2954 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 42; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX803 %s3; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX906 %s4; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX908 %s5; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX90A %s6; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX942 %s7; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX10 %s8; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX11 %s9; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -passes=atomic-expand %s | FileCheck -check-prefixes=COMMON,GFX12 %s10 11;---------------------------------------------------------------------12; atomicrmw xchg13;---------------------------------------------------------------------14 15; xchg is supported over PCIe, so no expansion is necessary16define double @test_atomicrmw_xchg_f64_global_system(ptr addrspace(1) %ptr, double %value) {17; COMMON-LABEL: define double @test_atomicrmw_xchg_f64_global_system(18; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0:[0-9]+]] {19; COMMON-NEXT: [[RES:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 820; COMMON-NEXT: ret double [[RES]]21;22 %res = atomicrmw xchg ptr addrspace(1) %ptr, double %value seq_cst23 ret double %res24}25 26; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.27define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {28; COMMON-LABEL: define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_fine_grained_memory(29; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {30; COMMON-NEXT: [[RES:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0:![0-9]+]]31; COMMON-NEXT: ret double [[RES]]32;33 %res = atomicrmw xchg ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !034 ret double %res35}36 37; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.38define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {39; COMMON-LABEL: define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_remote_memory(40; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {41; COMMON-NEXT: [[RES:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]42; COMMON-NEXT: ret double [[RES]]43;44 %res = atomicrmw xchg ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.remote.memory !045 ret double %res46}47 48; xchg is supported over PCIe, so no expansion is necessary. Metadata should be ignored.49define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {50; COMMON-LABEL: define double @test_atomicrmw_xchg_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(51; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {52; COMMON-NEXT: [[RES:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]53; COMMON-NEXT: ret double [[RES]]54;55 %res = atomicrmw xchg ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !056 ret double %res57}58 59;---------------------------------------------------------------------60; atomicrmw fadd61;---------------------------------------------------------------------62 63define double @test_atomicrmw_fadd_f64_global_system(ptr addrspace(1) %ptr, double %value) {64; COMMON-LABEL: define double @test_atomicrmw_fadd_f64_global_system(65; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {66; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 867; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]68; COMMON: atomicrmw.start:69; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]70; COMMON-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]71; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i6472; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i6473; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 874; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 175; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 076; COMMON-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double77; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]78; COMMON: atomicrmw.end:79; COMMON-NEXT: ret double [[TMP5]]80;81 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst82 ret double %res83}84 85define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {86; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(87; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {88; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 889; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]90; GFX803: atomicrmw.start:91; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]92; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]93; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i6494; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i6495; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]96; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 197; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 098; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double99; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]100; GFX803: atomicrmw.end:101; GFX803-NEXT: ret double [[TMP5]]102;103; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(104; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {105; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8106; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]107; GFX906: atomicrmw.start:108; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]109; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]110; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64111; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64112; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]113; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1114; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0115; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double116; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]117; GFX906: atomicrmw.end:118; GFX906-NEXT: ret double [[TMP5]]119;120; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(121; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {122; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8123; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]124; GFX908: atomicrmw.start:125; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]126; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]127; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64128; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64129; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]130; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1131; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0132; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double133; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]134; GFX908: atomicrmw.end:135; GFX908-NEXT: ret double [[TMP5]]136;137; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(138; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {139; GFX90A-NEXT: [[TMP5:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]140; GFX90A-NEXT: ret double [[TMP5]]141;142; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(143; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {144; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]145; GFX942-NEXT: ret double [[RES]]146;147; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(148; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {149; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8150; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]151; GFX10: atomicrmw.start:152; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]153; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]154; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64155; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64156; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]157; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1158; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0159; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double160; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]161; GFX10: atomicrmw.end:162; GFX10-NEXT: ret double [[TMP5]]163;164; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(165; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {166; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8167; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]168; GFX11: atomicrmw.start:169; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]170; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]171; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64172; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64173; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]174; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1175; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0176; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double177; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]178; GFX11: atomicrmw.end:179; GFX11-NEXT: ret double [[TMP5]]180;181; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory(182; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {183; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8184; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]185; GFX12: atomicrmw.start:186; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]187; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]188; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64189; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64190; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]191; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1192; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0193; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double194; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]195; GFX12: atomicrmw.end:196; GFX12-NEXT: ret double [[TMP5]]197;198 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0199 ret double %res200}201 202define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {203; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(204; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {205; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8206; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]207; GFX803: atomicrmw.start:208; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]209; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]210; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64211; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64212; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]213; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1214; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0215; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double216; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]217; GFX803: atomicrmw.end:218; GFX803-NEXT: ret double [[TMP5]]219;220; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(221; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {222; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8223; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]224; GFX906: atomicrmw.start:225; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]226; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]227; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64228; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64229; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]230; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1231; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0232; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double233; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]234; GFX906: atomicrmw.end:235; GFX906-NEXT: ret double [[TMP5]]236;237; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(238; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {239; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8240; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]241; GFX908: atomicrmw.start:242; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]243; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]244; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64245; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64246; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]247; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1248; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0249; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double250; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]251; GFX908: atomicrmw.end:252; GFX908-NEXT: ret double [[TMP5]]253;254; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(255; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {256; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8257; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]258; GFX90A: atomicrmw.start:259; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]260; GFX90A-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]261; GFX90A-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64262; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64263; GFX90A-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]264; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1265; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0266; GFX90A-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double267; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]268; GFX90A: atomicrmw.end:269; GFX90A-NEXT: ret double [[TMP5]]270;271; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(272; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {273; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]274; GFX942-NEXT: ret double [[RES]]275;276; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(277; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {278; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8279; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]280; GFX10: atomicrmw.start:281; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]282; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]283; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64284; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64285; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]286; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1287; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0288; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double289; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]290; GFX10: atomicrmw.end:291; GFX10-NEXT: ret double [[TMP5]]292;293; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(294; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {295; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8296; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]297; GFX11: atomicrmw.start:298; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]299; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]300; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64301; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64302; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]303; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1304; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0305; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double306; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]307; GFX11: atomicrmw.end:308; GFX11-NEXT: ret double [[TMP5]]309;310; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_remote_memory(311; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {312; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8313; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]314; GFX12: atomicrmw.start:315; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]316; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]317; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64318; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64319; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]320; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1321; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0322; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double323; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]324; GFX12: atomicrmw.end:325; GFX12-NEXT: ret double [[TMP5]]326;327 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.remote.memory !0328 ret double %res329}330 331define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {332; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(333; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {334; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8335; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]336; GFX803: atomicrmw.start:337; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]338; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]339; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64340; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64341; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]342; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1343; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0344; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double345; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]346; GFX803: atomicrmw.end:347; GFX803-NEXT: ret double [[TMP5]]348;349; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(350; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {351; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8352; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]353; GFX906: atomicrmw.start:354; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]355; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]356; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64357; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64358; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]359; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1360; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0361; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double362; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]363; GFX906: atomicrmw.end:364; GFX906-NEXT: ret double [[TMP5]]365;366; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(367; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {368; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8369; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]370; GFX908: atomicrmw.start:371; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]372; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]373; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64374; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64375; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]376; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1377; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0378; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double379; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]380; GFX908: atomicrmw.end:381; GFX908-NEXT: ret double [[TMP5]]382;383; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(384; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {385; GFX90A-NEXT: [[TMP5:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]386; GFX90A-NEXT: ret double [[TMP5]]387;388; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(389; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {390; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]391; GFX942-NEXT: ret double [[RES]]392;393; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(394; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {395; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8396; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]397; GFX10: atomicrmw.start:398; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]399; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]400; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64401; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64402; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]403; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1404; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0405; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double406; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]407; GFX10: atomicrmw.end:408; GFX10-NEXT: ret double [[TMP5]]409;410; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(411; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {412; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8413; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]414; GFX11: atomicrmw.start:415; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]416; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]417; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64418; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64419; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]420; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1421; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0422; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double423; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]424; GFX11: atomicrmw.end:425; GFX11-NEXT: ret double [[TMP5]]426;427; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(428; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {429; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8430; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]431; GFX12: atomicrmw.start:432; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]433; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]434; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64435; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64436; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]437; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1438; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0439; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double440; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]441; GFX12: atomicrmw.end:442; GFX12-NEXT: ret double [[TMP5]]443;444 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0445 ret double %res446}447 448define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(ptr addrspace(1) %ptr, double %value) #0 {449; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(450; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {451; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8452; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]453; GFX803: atomicrmw.start:454; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]455; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]456; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64457; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64458; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]459; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1460; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0461; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double462; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]463; GFX803: atomicrmw.end:464; GFX803-NEXT: ret double [[TMP5]]465;466; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(467; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {468; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8469; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]470; GFX906: atomicrmw.start:471; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]472; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]473; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64474; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64475; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]476; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1477; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0478; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double479; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]480; GFX906: atomicrmw.end:481; GFX906-NEXT: ret double [[TMP5]]482;483; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(484; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {485; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8486; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]487; GFX908: atomicrmw.start:488; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]489; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]490; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64491; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64492; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]493; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1494; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0495; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double496; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]497; GFX908: atomicrmw.end:498; GFX908-NEXT: ret double [[TMP5]]499;500; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(501; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {502; GFX90A-NEXT: [[TMP5:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]503; GFX90A-NEXT: ret double [[TMP5]]504;505; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(506; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {507; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]508; GFX942-NEXT: ret double [[RES]]509;510; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(511; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {512; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8513; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]514; GFX10: atomicrmw.start:515; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]516; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]517; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64518; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64519; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]520; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1521; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0522; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double523; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]524; GFX10: atomicrmw.end:525; GFX10-NEXT: ret double [[TMP5]]526;527; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(528; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {529; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8530; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]531; GFX11: atomicrmw.start:532; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]533; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]534; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64535; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64536; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]537; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1538; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0539; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double540; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]541; GFX11: atomicrmw.end:542; GFX11-NEXT: ret double [[TMP5]]543;544; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_daz(545; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1:[0-9]+]] {546; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8547; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]548; GFX12: atomicrmw.start:549; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]550; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]551; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64552; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64553; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]554; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1555; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0556; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double557; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]558; GFX12: atomicrmw.end:559; GFX12-NEXT: ret double [[TMP5]]560;561 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0562 ret double %res563}564 565define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(ptr addrspace(1) %ptr, double %value) #1 {566; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(567; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {568; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8569; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]570; GFX803: atomicrmw.start:571; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]572; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]573; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64574; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64575; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]576; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1577; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0578; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double579; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]580; GFX803: atomicrmw.end:581; GFX803-NEXT: ret double [[TMP5]]582;583; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(584; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {585; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8586; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]587; GFX906: atomicrmw.start:588; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]589; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]590; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64591; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64592; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]593; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1594; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0595; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double596; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]597; GFX906: atomicrmw.end:598; GFX906-NEXT: ret double [[TMP5]]599;600; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(601; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {602; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8603; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]604; GFX908: atomicrmw.start:605; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]606; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]607; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64608; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64609; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]610; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1611; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0612; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double613; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]614; GFX908: atomicrmw.end:615; GFX908-NEXT: ret double [[TMP5]]616;617; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(618; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {619; GFX90A-NEXT: [[TMP5:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]620; GFX90A-NEXT: ret double [[TMP5]]621;622; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(623; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {624; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]625; GFX942-NEXT: ret double [[RES]]626;627; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(628; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {629; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8630; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]631; GFX10: atomicrmw.start:632; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]633; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]634; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64635; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64636; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]637; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1638; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0639; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double640; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]641; GFX10: atomicrmw.end:642; GFX10-NEXT: ret double [[TMP5]]643;644; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(645; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {646; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8647; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]648; GFX11: atomicrmw.start:649; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]650; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]651; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64652; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64653; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]654; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1655; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0656; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double657; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]658; GFX11: atomicrmw.end:659; GFX11-NEXT: ret double [[TMP5]]660;661; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory___denormal_fp_mode_f64_dynamic(662; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2:[0-9]+]] {663; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8664; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]665; GFX12: atomicrmw.start:666; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]667; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]668; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64669; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64670; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]671; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1672; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0673; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double674; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]675; GFX12: atomicrmw.end:676; GFX12-NEXT: ret double [[TMP5]]677;678 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0679 ret double %res680}681 682define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, double %value) {683; COMMON-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode(684; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {685; COMMON-NEXT: [[TMP3:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8686; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]687; COMMON: atomicrmw.start:688; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]689; COMMON-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]690; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64691; COMMON-NEXT: [[TMP5:%.*]] = bitcast double [[LOADED]] to i64692; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP5]], i64 [[TMP2]] seq_cst seq_cst, align 8693; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1694; COMMON-NEXT: [[NEWLOADED1:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0695; COMMON-NEXT: [[NEWLOADED]] = bitcast i64 [[NEWLOADED1]] to double696; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]697; COMMON: atomicrmw.end:698; COMMON-NEXT: ret double [[NEWLOADED]]699;700 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.ignore.denormal.mode !0701 ret double %res702}703 704define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {705; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(706; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {707; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8708; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]709; GFX803: atomicrmw.start:710; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]711; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]712; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64713; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64714; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]715; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1716; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0717; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double718; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]719; GFX803: atomicrmw.end:720; GFX803-NEXT: ret double [[TMP5]]721;722; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(723; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {724; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8725; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]726; GFX906: atomicrmw.start:727; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]728; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]729; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64730; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64731; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]732; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1733; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0734; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double735; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]736; GFX906: atomicrmw.end:737; GFX906-NEXT: ret double [[TMP5]]738;739; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(740; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {741; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8742; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]743; GFX908: atomicrmw.start:744; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]745; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]746; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64747; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64748; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]749; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1750; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0751; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double752; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]753; GFX908: atomicrmw.end:754; GFX908-NEXT: ret double [[TMP5]]755;756; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(757; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {758; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]759; GFX90A-NEXT: ret double [[RES]]760;761; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(762; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {763; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]764; GFX942-NEXT: ret double [[RES]]765;766; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(767; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {768; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8769; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]770; GFX10: atomicrmw.start:771; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]772; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]773; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64774; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64775; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]776; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1777; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0778; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double779; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]780; GFX10: atomicrmw.end:781; GFX10-NEXT: ret double [[TMP5]]782;783; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(784; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {785; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8786; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]787; GFX11: atomicrmw.start:788; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]789; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]790; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64791; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64792; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]793; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1794; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0795; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double796; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]797; GFX11: atomicrmw.end:798; GFX11-NEXT: ret double [[TMP5]]799;800; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(801; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {802; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8803; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]804; GFX12: atomicrmw.start:805; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]806; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]807; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64808; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64809; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]810; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1811; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0812; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double813; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]814; GFX12: atomicrmw.end:815; GFX12-NEXT: ret double [[TMP5]]816;817 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !0818 ret double %res819}820 821define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {822; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(823; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {824; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8825; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]826; GFX803: atomicrmw.start:827; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]828; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]829; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64830; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64831; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]832; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1833; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0834; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double835; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]836; GFX803: atomicrmw.end:837; GFX803-NEXT: ret double [[TMP5]]838;839; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(840; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {841; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8842; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]843; GFX906: atomicrmw.start:844; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]845; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]846; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64847; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64848; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]849; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1850; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0851; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double852; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]853; GFX906: atomicrmw.end:854; GFX906-NEXT: ret double [[TMP5]]855;856; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(857; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {858; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8859; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]860; GFX908: atomicrmw.start:861; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]862; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]863; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64864; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64865; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]866; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1867; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0868; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double869; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]870; GFX908: atomicrmw.end:871; GFX908-NEXT: ret double [[TMP5]]872;873; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(874; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {875; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8876; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]877; GFX90A: atomicrmw.start:878; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]879; GFX90A-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]880; GFX90A-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64881; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64882; GFX90A-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]883; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1884; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0885; GFX90A-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double886; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]887; GFX90A: atomicrmw.end:888; GFX90A-NEXT: ret double [[TMP5]]889;890; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(891; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {892; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]893; GFX942-NEXT: ret double [[RES]]894;895; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(896; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {897; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8898; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]899; GFX10: atomicrmw.start:900; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]901; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]902; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64903; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64904; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]905; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1906; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0907; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double908; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]909; GFX10: atomicrmw.end:910; GFX10-NEXT: ret double [[TMP5]]911;912; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(913; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {914; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8915; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]916; GFX11: atomicrmw.start:917; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]918; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]919; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64920; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64921; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]922; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1923; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0924; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double925; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]926; GFX11: atomicrmw.end:927; GFX11-NEXT: ret double [[TMP5]]928;929; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(930; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {931; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8932; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]933; GFX12: atomicrmw.start:934; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]935; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]936; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64937; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64938; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]939; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1940; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0941; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double942; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]943; GFX12: atomicrmw.end:944; GFX12-NEXT: ret double [[TMP5]]945;946 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !0947 ret double %res948}949 950define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {951; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(952; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {953; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8954; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]955; GFX803: atomicrmw.start:956; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]957; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]958; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64959; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64960; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]961; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1962; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0963; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double964; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]965; GFX803: atomicrmw.end:966; GFX803-NEXT: ret double [[TMP5]]967;968; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(969; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {970; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8971; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]972; GFX906: atomicrmw.start:973; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]974; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]975; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64976; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64977; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]978; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1979; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0980; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double981; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]982; GFX906: atomicrmw.end:983; GFX906-NEXT: ret double [[TMP5]]984;985; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(986; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {987; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 8988; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]989; GFX908: atomicrmw.start:990; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]991; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]992; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i64993; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i64994; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]995; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1996; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0997; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double998; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]999; GFX908: atomicrmw.end:1000; GFX908-NEXT: ret double [[TMP5]]1001;1002; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1003; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1004; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1005; GFX90A-NEXT: ret double [[RES]]1006;1007; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1008; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1009; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1010; GFX942-NEXT: ret double [[RES]]1011;1012; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1013; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1014; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81015; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]1016; GFX10: atomicrmw.start:1017; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1018; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1019; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641020; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641021; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1022; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11023; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01024; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1025; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1026; GFX10: atomicrmw.end:1027; GFX10-NEXT: ret double [[TMP5]]1028;1029; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1030; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1031; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81032; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1033; GFX11: atomicrmw.start:1034; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1035; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1036; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641037; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641038; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1039; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11040; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01041; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1042; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1043; GFX11: atomicrmw.end:1044; GFX11-NEXT: ret double [[TMP5]]1045;1046; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1047; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1048; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81049; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1050; GFX12: atomicrmw.start:1051; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1052; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1053; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641054; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641055; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1056; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11057; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01058; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1059; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1060; GFX12: atomicrmw.end:1061; GFX12-NEXT: ret double [[TMP5]]1062;1063 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !01064 ret double %res1065}1066 1067define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(ptr addrspace(1) %ptr, double %value) #0 {1068; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1069; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1070; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81071; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1072; GFX803: atomicrmw.start:1073; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1074; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1075; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641076; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641077; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1078; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11079; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01080; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1081; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1082; GFX803: atomicrmw.end:1083; GFX803-NEXT: ret double [[TMP5]]1084;1085; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1086; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1087; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81088; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1089; GFX906: atomicrmw.start:1090; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1091; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1092; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641093; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641094; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1095; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11096; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01097; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1098; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1099; GFX906: atomicrmw.end:1100; GFX906-NEXT: ret double [[TMP5]]1101;1102; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1103; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1104; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81105; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1106; GFX908: atomicrmw.start:1107; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1108; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1109; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641110; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641111; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1112; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11113; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01114; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1115; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1116; GFX908: atomicrmw.end:1117; GFX908-NEXT: ret double [[TMP5]]1118;1119; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1120; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1121; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1122; GFX90A-NEXT: ret double [[RES]]1123;1124; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1125; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1126; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1127; GFX942-NEXT: ret double [[RES]]1128;1129; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1130; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1131; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81132; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]1133; GFX10: atomicrmw.start:1134; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1135; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1136; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641137; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641138; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1139; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11140; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01141; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1142; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1143; GFX10: atomicrmw.end:1144; GFX10-NEXT: ret double [[TMP5]]1145;1146; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1147; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1148; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81149; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1150; GFX11: atomicrmw.start:1151; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1152; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1153; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641154; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641155; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1156; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11157; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01158; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1159; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1160; GFX11: atomicrmw.end:1161; GFX11-NEXT: ret double [[TMP5]]1162;1163; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_daz(1164; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR1]] {1165; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81166; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1167; GFX12: atomicrmw.start:1168; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1169; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1170; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641171; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641172; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1173; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11174; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01175; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1176; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1177; GFX12: atomicrmw.end:1178; GFX12-NEXT: ret double [[TMP5]]1179;1180 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !01181 ret double %res1182}1183 1184define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(ptr addrspace(1) %ptr, double %value) #1 {1185; GFX803-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1186; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1187; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81188; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1189; GFX803: atomicrmw.start:1190; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1191; GFX803-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1192; GFX803-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641193; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641194; GFX803-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1195; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11196; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01197; GFX803-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1198; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1199; GFX803: atomicrmw.end:1200; GFX803-NEXT: ret double [[TMP5]]1201;1202; GFX906-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1203; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1204; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81205; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1206; GFX906: atomicrmw.start:1207; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1208; GFX906-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1209; GFX906-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641210; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641211; GFX906-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1212; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11213; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01214; GFX906-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1215; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1216; GFX906: atomicrmw.end:1217; GFX906-NEXT: ret double [[TMP5]]1218;1219; GFX908-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1220; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1221; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81222; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1223; GFX908: atomicrmw.start:1224; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1225; GFX908-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1226; GFX908-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641227; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641228; GFX908-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1229; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11230; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01231; GFX908-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1232; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1233; GFX908: atomicrmw.end:1234; GFX908-NEXT: ret double [[TMP5]]1235;1236; GFX90A-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1237; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1238; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1239; GFX90A-NEXT: ret double [[RES]]1240;1241; GFX942-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1242; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1243; GFX942-NEXT: [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1244; GFX942-NEXT: ret double [[RES]]1245;1246; GFX10-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1247; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1248; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81249; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]1250; GFX10: atomicrmw.start:1251; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1252; GFX10-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1253; GFX10-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641254; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641255; GFX10-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1256; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11257; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01258; GFX10-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1259; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1260; GFX10: atomicrmw.end:1261; GFX10-NEXT: ret double [[TMP5]]1262;1263; GFX11-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1264; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1265; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81266; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1267; GFX11: atomicrmw.start:1268; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1269; GFX11-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1270; GFX11-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641271; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641272; GFX11-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1273; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11274; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01275; GFX11-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1276; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1277; GFX11: atomicrmw.end:1278; GFX11-NEXT: ret double [[TMP5]]1279;1280; GFX12-LABEL: define double @test_atomicrmw_fadd_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory__denormal_mode_dynamic(1281; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR2]] {1282; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81283; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1284; GFX12: atomicrmw.start:1285; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1286; GFX12-NEXT: [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE]]1287; GFX12-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641288; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641289; GFX12-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1290; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11291; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01292; GFX12-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1293; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1294; GFX12: atomicrmw.end:1295; GFX12-NEXT: ret double [[TMP5]]1296;1297 %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !01298 ret double %res1299}1300 1301;---------------------------------------------------------------------1302; atomicrmw fsub1303;---------------------------------------------------------------------1304 1305define double @test_atomicrmw_fsub_f64_global_system(ptr addrspace(1) %ptr, double %value) {1306; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system(1307; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1308; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81309; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1310; COMMON: atomicrmw.start:1311; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1312; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1313; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641314; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641315; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 81316; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11317; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01318; COMMON-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1319; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1320; COMMON: atomicrmw.end:1321; COMMON-NEXT: ret double [[RES]]1322;1323 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst1324 ret double %res1325}1326 1327define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {1328; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_fine_grained_memory(1329; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1330; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81331; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1332; COMMON: atomicrmw.start:1333; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1334; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1335; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641336; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641337; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1338; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11339; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01340; COMMON-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1341; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1342; COMMON: atomicrmw.end:1343; COMMON-NEXT: ret double [[RES]]1344;1345 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !01346 ret double %res1347}1348 1349define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1350; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_remote_memory(1351; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1352; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81353; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1354; COMMON: atomicrmw.start:1355; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1356; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1357; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641358; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641359; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1360; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11361; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01362; COMMON-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1363; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1364; COMMON: atomicrmw.end:1365; COMMON-NEXT: ret double [[RES]]1366;1367 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.remote.memory !01368 ret double %res1369}1370 1371define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1372; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1373; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1374; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81375; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1376; COMMON: atomicrmw.start:1377; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1378; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1379; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641380; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641381; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1382; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11383; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01384; COMMON-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1385; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1386; COMMON: atomicrmw.end:1387; COMMON-NEXT: ret double [[RES]]1388;1389 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !01390 ret double %res1391}1392 1393define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, double %value) {1394; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode(1395; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1396; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81397; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1398; COMMON: atomicrmw.start:1399; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1400; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1401; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641402; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641403; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 81404; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11405; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01406; COMMON-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1407; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1408; COMMON: atomicrmw.end:1409; COMMON-NEXT: ret double [[TMP5]]1410;1411 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.ignore.denormal.mode !01412 ret double %res1413}1414 1415define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {1416; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1417; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1418; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81419; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1420; COMMON: atomicrmw.start:1421; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1422; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1423; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641424; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641425; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1426; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11427; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01428; COMMON-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1429; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1430; COMMON: atomicrmw.end:1431; COMMON-NEXT: ret double [[TMP5]]1432;1433 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !01434 ret double %res1435}1436 1437define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1438; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(1439; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1440; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81441; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1442; COMMON: atomicrmw.start:1443; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1444; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1445; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641446; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641447; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1448; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11449; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01450; COMMON-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1451; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1452; COMMON: atomicrmw.end:1453; COMMON-NEXT: ret double [[TMP5]]1454;1455 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !01456 ret double %res1457}1458 1459define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1460; COMMON-LABEL: define double @test_atomicrmw_fsub_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1461; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1462; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81463; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1464; COMMON: atomicrmw.start:1465; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]1466; COMMON-NEXT: [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE]]1467; COMMON-NEXT: [[TMP2:%.*]] = bitcast double [[NEW]] to i641468; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[LOADED]] to i641469; COMMON-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1470; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 11471; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 01472; COMMON-NEXT: [[TMP5]] = bitcast i64 [[NEWLOADED]] to double1473; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1474; COMMON: atomicrmw.end:1475; COMMON-NEXT: ret double [[TMP5]]1476;1477 %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !01478 ret double %res1479}1480 1481;---------------------------------------------------------------------1482; atomicrmw fmax1483;---------------------------------------------------------------------1484 1485define double @test_atomicrmw_fmax_f64_global_system(ptr addrspace(1) %ptr, double %value) {1486; COMMON-LABEL: define double @test_atomicrmw_fmax_f64_global_system(1487; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1488; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81489; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1490; COMMON: atomicrmw.start:1491; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1492; COMMON-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1493; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641494; COMMON-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641495; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 81496; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11497; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01498; COMMON-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1499; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1500; COMMON: atomicrmw.end:1501; COMMON-NEXT: ret double [[TMP6]]1502;1503 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst1504 ret double %res1505}1506 1507define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {1508; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1509; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1510; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81511; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1512; GFX803: atomicrmw.start:1513; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1514; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1515; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641516; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641517; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1518; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11519; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01520; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1521; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1522; GFX803: atomicrmw.end:1523; GFX803-NEXT: ret double [[TMP6]]1524;1525; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1526; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1527; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81528; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1529; GFX906: atomicrmw.start:1530; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1531; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1532; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641533; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641534; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1535; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11536; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01537; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1538; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1539; GFX906: atomicrmw.end:1540; GFX906-NEXT: ret double [[TMP6]]1541;1542; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1543; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1544; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81545; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1546; GFX908: atomicrmw.start:1547; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1548; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1549; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641550; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641551; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1552; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11553; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01554; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1555; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1556; GFX908: atomicrmw.end:1557; GFX908-NEXT: ret double [[TMP6]]1558;1559; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1560; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1561; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1562; GFX90A-NEXT: ret double [[RES]]1563;1564; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1565; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1566; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1567; GFX942-NEXT: ret double [[RES]]1568;1569; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1570; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1571; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1572; GFX10-NEXT: ret double [[RES]]1573;1574; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1575; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1576; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81577; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1578; GFX11: atomicrmw.start:1579; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1580; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1581; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641582; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641583; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1584; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11585; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01586; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1587; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1588; GFX11: atomicrmw.end:1589; GFX11-NEXT: ret double [[TMP6]]1590;1591; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory(1592; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1593; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81594; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1595; GFX12: atomicrmw.start:1596; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1597; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1598; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641599; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641600; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1601; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11602; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01603; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1604; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1605; GFX12: atomicrmw.end:1606; GFX12-NEXT: ret double [[TMP6]]1607;1608 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !01609 ret double %res1610}1611 1612define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1613; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1614; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1615; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81616; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1617; GFX803: atomicrmw.start:1618; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1619; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1620; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641621; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641622; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1623; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11624; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01625; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1626; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1627; GFX803: atomicrmw.end:1628; GFX803-NEXT: ret double [[TMP6]]1629;1630; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1631; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1632; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81633; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1634; GFX906: atomicrmw.start:1635; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1636; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1637; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641638; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641639; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1640; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11641; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01642; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1643; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1644; GFX906: atomicrmw.end:1645; GFX906-NEXT: ret double [[TMP6]]1646;1647; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1648; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1649; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81650; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1651; GFX908: atomicrmw.start:1652; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1653; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1654; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641655; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641656; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1657; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11658; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01659; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1660; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1661; GFX908: atomicrmw.end:1662; GFX908-NEXT: ret double [[TMP6]]1663;1664; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1665; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1666; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81667; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]1668; GFX90A: atomicrmw.start:1669; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1670; GFX90A-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1671; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641672; GFX90A-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641673; GFX90A-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1674; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11675; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01676; GFX90A-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1677; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1678; GFX90A: atomicrmw.end:1679; GFX90A-NEXT: ret double [[RES]]1680;1681; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1682; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1683; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1684; GFX942-NEXT: ret double [[RES]]1685;1686; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1687; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1688; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81689; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]1690; GFX10: atomicrmw.start:1691; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]1692; GFX10-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1693; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641694; GFX10-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641695; GFX10-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1696; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11697; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01698; GFX10-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double1699; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1700; GFX10: atomicrmw.end:1701; GFX10-NEXT: ret double [[RES]]1702;1703; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1704; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1705; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81706; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1707; GFX11: atomicrmw.start:1708; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1709; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1710; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641711; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641712; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1713; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11714; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01715; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1716; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1717; GFX11: atomicrmw.end:1718; GFX11-NEXT: ret double [[TMP6]]1719;1720; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_remote_memory(1721; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1722; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81723; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1724; GFX12: atomicrmw.start:1725; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1726; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1727; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641728; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641729; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1730; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11731; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01732; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1733; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1734; GFX12: atomicrmw.end:1735; GFX12-NEXT: ret double [[TMP6]]1736;1737 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.remote.memory !01738 ret double %res1739}1740 1741define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1742; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1743; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1744; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81745; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1746; GFX803: atomicrmw.start:1747; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1748; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1749; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641750; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641751; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1752; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11753; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01754; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1755; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1756; GFX803: atomicrmw.end:1757; GFX803-NEXT: ret double [[TMP6]]1758;1759; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1760; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1761; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81762; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1763; GFX906: atomicrmw.start:1764; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1765; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1766; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641767; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641768; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1769; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11770; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01771; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1772; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1773; GFX906: atomicrmw.end:1774; GFX906-NEXT: ret double [[TMP6]]1775;1776; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1777; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1778; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81779; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1780; GFX908: atomicrmw.start:1781; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1782; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1783; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641784; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641785; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1786; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11787; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01788; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1789; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1790; GFX908: atomicrmw.end:1791; GFX908-NEXT: ret double [[TMP6]]1792;1793; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1794; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1795; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1796; GFX90A-NEXT: ret double [[RES]]1797;1798; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1799; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1800; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1801; GFX942-NEXT: ret double [[RES]]1802;1803; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1804; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1805; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1806; GFX10-NEXT: ret double [[RES]]1807;1808; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1809; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1810; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81811; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1812; GFX11: atomicrmw.start:1813; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1814; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1815; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641816; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641817; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1818; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11819; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01820; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1821; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1822; GFX11: atomicrmw.end:1823; GFX11-NEXT: ret double [[TMP6]]1824;1825; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(1826; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1827; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81828; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1829; GFX12: atomicrmw.start:1830; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1831; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1832; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641833; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641834; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]1835; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11836; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01837; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1838; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1839; GFX12: atomicrmw.end:1840; GFX12-NEXT: ret double [[TMP6]]1841;1842 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !01843 ret double %res1844}1845 1846define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, double %value) {1847; COMMON-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode(1848; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1849; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81850; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]1851; COMMON: atomicrmw.start:1852; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1853; COMMON-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1854; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641855; COMMON-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641856; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 81857; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11858; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01859; COMMON-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1860; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1861; COMMON: atomicrmw.end:1862; COMMON-NEXT: ret double [[TMP6]]1863;1864 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.ignore.denormal.mode !01865 ret double %res1866}1867 1868define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {1869; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1870; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1871; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81872; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1873; GFX803: atomicrmw.start:1874; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1875; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1876; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641877; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641878; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1879; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11880; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01881; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1882; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1883; GFX803: atomicrmw.end:1884; GFX803-NEXT: ret double [[TMP6]]1885;1886; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1887; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1888; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81889; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1890; GFX906: atomicrmw.start:1891; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1892; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1893; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641894; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641895; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1896; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11897; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01898; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1899; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1900; GFX906: atomicrmw.end:1901; GFX906-NEXT: ret double [[TMP6]]1902;1903; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1904; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1905; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81906; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]1907; GFX908: atomicrmw.start:1908; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1909; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1910; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641911; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641912; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1913; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11914; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01915; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1916; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1917; GFX908: atomicrmw.end:1918; GFX908-NEXT: ret double [[TMP6]]1919;1920; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1921; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1922; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1923; GFX90A-NEXT: ret double [[RES]]1924;1925; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1926; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1927; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1928; GFX942-NEXT: ret double [[RES]]1929;1930; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1931; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1932; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]1933; GFX10-NEXT: ret double [[RES]]1934;1935; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1936; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1937; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81938; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]1939; GFX11: atomicrmw.start:1940; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1941; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1942; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641943; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641944; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1945; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11946; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01947; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1948; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1949; GFX11: atomicrmw.end:1950; GFX11-NEXT: ret double [[TMP6]]1951;1952; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(1953; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1954; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81955; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]1956; GFX12: atomicrmw.start:1957; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1958; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1959; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641960; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641961; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]1962; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11963; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01964; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1965; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1966; GFX12: atomicrmw.end:1967; GFX12-NEXT: ret double [[TMP6]]1968;1969 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !01970 ret double %res1971}1972 1973define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {1974; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(1975; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1976; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81977; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]1978; GFX803: atomicrmw.start:1979; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1980; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1981; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641982; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i641983; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]1984; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 11985; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 01986; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double1987; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]1988; GFX803: atomicrmw.end:1989; GFX803-NEXT: ret double [[TMP6]]1990;1991; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(1992; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {1993; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 81994; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]1995; GFX906: atomicrmw.start:1996; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]1997; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])1998; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i641999; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642000; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2001; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12002; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02003; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2004; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2005; GFX906: atomicrmw.end:2006; GFX906-NEXT: ret double [[TMP6]]2007;2008; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2009; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2010; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82011; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2012; GFX908: atomicrmw.start:2013; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2014; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2015; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642016; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642017; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2018; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12019; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02020; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2021; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2022; GFX908: atomicrmw.end:2023; GFX908-NEXT: ret double [[TMP6]]2024;2025; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2026; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2027; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82028; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]2029; GFX90A: atomicrmw.start:2030; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2031; GFX90A-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2032; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642033; GFX90A-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642034; GFX90A-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2035; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12036; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02037; GFX90A-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2038; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2039; GFX90A: atomicrmw.end:2040; GFX90A-NEXT: ret double [[TMP6]]2041;2042; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2043; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2044; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2045; GFX942-NEXT: ret double [[RES]]2046;2047; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2048; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2049; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82050; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]2051; GFX10: atomicrmw.start:2052; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2053; GFX10-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2054; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642055; GFX10-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642056; GFX10-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2057; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12058; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02059; GFX10-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2060; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2061; GFX10: atomicrmw.end:2062; GFX10-NEXT: ret double [[TMP6]]2063;2064; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2065; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2066; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82067; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2068; GFX11: atomicrmw.start:2069; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2070; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2071; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642072; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642073; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2074; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12075; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02076; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2077; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2078; GFX11: atomicrmw.end:2079; GFX11-NEXT: ret double [[TMP6]]2080;2081; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2082; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2083; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82084; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2085; GFX12: atomicrmw.start:2086; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2087; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2088; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642089; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642090; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2091; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12092; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02093; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2094; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2095; GFX12: atomicrmw.end:2096; GFX12-NEXT: ret double [[TMP6]]2097;2098 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !02099 ret double %res2100}2101 2102define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {2103; GFX803-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2104; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2105; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82106; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2107; GFX803: atomicrmw.start:2108; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2109; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2110; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642111; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642112; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2113; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12114; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02115; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2116; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2117; GFX803: atomicrmw.end:2118; GFX803-NEXT: ret double [[TMP6]]2119;2120; GFX906-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2121; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2122; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82123; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2124; GFX906: atomicrmw.start:2125; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2126; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2127; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642128; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642129; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2130; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12131; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02132; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2133; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2134; GFX906: atomicrmw.end:2135; GFX906-NEXT: ret double [[TMP6]]2136;2137; GFX908-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2138; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2139; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82140; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2141; GFX908: atomicrmw.start:2142; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2143; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2144; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642145; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642146; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2147; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12148; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02149; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2150; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2151; GFX908: atomicrmw.end:2152; GFX908-NEXT: ret double [[TMP6]]2153;2154; GFX90A-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2155; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2156; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2157; GFX90A-NEXT: ret double [[RES]]2158;2159; GFX942-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2160; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2161; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2162; GFX942-NEXT: ret double [[RES]]2163;2164; GFX10-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2165; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2166; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2167; GFX10-NEXT: ret double [[RES]]2168;2169; GFX11-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2170; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2171; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82172; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2173; GFX11: atomicrmw.start:2174; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2175; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2176; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642177; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642178; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2179; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12180; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02181; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2182; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2183; GFX11: atomicrmw.end:2184; GFX11-NEXT: ret double [[TMP6]]2185;2186; GFX12-LABEL: define double @test_atomicrmw_fmax_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2187; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2188; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82189; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2190; GFX12: atomicrmw.start:2191; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2192; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE]])2193; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642194; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642195; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2196; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12197; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02198; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2199; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2200; GFX12: atomicrmw.end:2201; GFX12-NEXT: ret double [[TMP6]]2202;2203 %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !02204 ret double %res2205}2206 2207;---------------------------------------------------------------------2208; atomicrmw fmin2209;---------------------------------------------------------------------2210 2211define double @test_atomicrmw_fmin_f64_global_system(ptr addrspace(1) %ptr, double %value) {2212; COMMON-LABEL: define double @test_atomicrmw_fmin_f64_global_system(2213; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2214; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82215; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]2216; COMMON: atomicrmw.start:2217; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2218; COMMON-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2219; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642220; COMMON-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642221; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 82222; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12223; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02224; COMMON-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2225; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2226; COMMON: atomicrmw.end:2227; COMMON-NEXT: ret double [[TMP6]]2228;2229 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst2230 ret double %res2231}2232 2233define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {2234; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2235; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2236; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82237; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2238; GFX803: atomicrmw.start:2239; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2240; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2241; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642242; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642243; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2244; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12245; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02246; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2247; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2248; GFX803: atomicrmw.end:2249; GFX803-NEXT: ret double [[TMP6]]2250;2251; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2252; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2253; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82254; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2255; GFX906: atomicrmw.start:2256; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2257; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2258; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642259; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642260; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2261; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12262; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02263; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2264; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2265; GFX906: atomicrmw.end:2266; GFX906-NEXT: ret double [[TMP6]]2267;2268; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2269; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2270; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82271; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2272; GFX908: atomicrmw.start:2273; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2274; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2275; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642276; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642277; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2278; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12279; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02280; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2281; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2282; GFX908: atomicrmw.end:2283; GFX908-NEXT: ret double [[TMP6]]2284;2285; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2286; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2287; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2288; GFX90A-NEXT: ret double [[RES]]2289;2290; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2291; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2292; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2293; GFX942-NEXT: ret double [[RES]]2294;2295; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2296; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2297; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2298; GFX10-NEXT: ret double [[RES]]2299;2300; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2301; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2302; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82303; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2304; GFX11: atomicrmw.start:2305; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2306; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2307; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642308; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642309; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2310; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12311; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02312; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2313; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2314; GFX11: atomicrmw.end:2315; GFX11-NEXT: ret double [[TMP6]]2316;2317; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory(2318; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2319; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82320; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2321; GFX12: atomicrmw.start:2322; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2323; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2324; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642325; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642326; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2327; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12328; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02329; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2330; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2331; GFX12: atomicrmw.end:2332; GFX12-NEXT: ret double [[TMP6]]2333;2334 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !02335 ret double %res2336}2337 2338define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {2339; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2340; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2341; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82342; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2343; GFX803: atomicrmw.start:2344; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2345; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2346; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642347; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642348; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2349; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12350; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02351; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2352; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2353; GFX803: atomicrmw.end:2354; GFX803-NEXT: ret double [[TMP6]]2355;2356; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2357; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2358; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82359; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2360; GFX906: atomicrmw.start:2361; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2362; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2363; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642364; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642365; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2366; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12367; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02368; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2369; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2370; GFX906: atomicrmw.end:2371; GFX906-NEXT: ret double [[TMP6]]2372;2373; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2374; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2375; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82376; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2377; GFX908: atomicrmw.start:2378; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2379; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2380; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642381; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642382; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2383; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12384; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02385; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2386; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2387; GFX908: atomicrmw.end:2388; GFX908-NEXT: ret double [[TMP6]]2389;2390; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2391; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2392; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82393; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]2394; GFX90A: atomicrmw.start:2395; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]2396; GFX90A-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2397; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642398; GFX90A-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642399; GFX90A-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2400; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12401; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02402; GFX90A-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double2403; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2404; GFX90A: atomicrmw.end:2405; GFX90A-NEXT: ret double [[RES]]2406;2407; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2408; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2409; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2410; GFX942-NEXT: ret double [[RES]]2411;2412; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2413; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2414; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82415; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]2416; GFX10: atomicrmw.start:2417; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[RES:%.*]], [[ATOMICRMW_START]] ]2418; GFX10-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2419; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642420; GFX10-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642421; GFX10-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2422; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12423; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02424; GFX10-NEXT: [[RES]] = bitcast i64 [[NEWLOADED]] to double2425; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2426; GFX10: atomicrmw.end:2427; GFX10-NEXT: ret double [[RES]]2428;2429; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2430; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2431; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82432; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2433; GFX11: atomicrmw.start:2434; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2435; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2436; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642437; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642438; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2439; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12440; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02441; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2442; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2443; GFX11: atomicrmw.end:2444; GFX11-NEXT: ret double [[TMP6]]2445;2446; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_remote_memory(2447; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2448; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82449; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2450; GFX12: atomicrmw.start:2451; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2452; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2453; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642454; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642455; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2456; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12457; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02458; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2459; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2460; GFX12: atomicrmw.end:2461; GFX12-NEXT: ret double [[TMP6]]2462;2463 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.remote.memory !02464 ret double %res2465}2466 2467define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {2468; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2469; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2470; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82471; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2472; GFX803: atomicrmw.start:2473; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2474; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2475; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642476; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642477; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2478; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12479; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02480; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2481; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2482; GFX803: atomicrmw.end:2483; GFX803-NEXT: ret double [[TMP6]]2484;2485; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2486; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2487; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82488; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2489; GFX906: atomicrmw.start:2490; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2491; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2492; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642493; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642494; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2495; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12496; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02497; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2498; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2499; GFX906: atomicrmw.end:2500; GFX906-NEXT: ret double [[TMP6]]2501;2502; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2503; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2504; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82505; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2506; GFX908: atomicrmw.start:2507; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2508; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2509; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642510; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642511; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2512; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12513; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02514; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2515; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2516; GFX908: atomicrmw.end:2517; GFX908-NEXT: ret double [[TMP6]]2518;2519; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2520; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2521; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2522; GFX90A-NEXT: ret double [[RES]]2523;2524; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2525; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2526; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2527; GFX942-NEXT: ret double [[RES]]2528;2529; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2530; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2531; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2532; GFX10-NEXT: ret double [[RES]]2533;2534; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2535; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2536; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82537; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2538; GFX11: atomicrmw.start:2539; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2540; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2541; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642542; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642543; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2544; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12545; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02546; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2547; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2548; GFX11: atomicrmw.end:2549; GFX11-NEXT: ret double [[TMP6]]2550;2551; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2552; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2553; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82554; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2555; GFX12: atomicrmw.start:2556; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2557; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2558; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642559; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642560; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2561; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12562; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02563; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2564; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2565; GFX12: atomicrmw.end:2566; GFX12-NEXT: ret double [[TMP6]]2567;2568 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !02569 ret double %res2570}2571 2572define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode(ptr addrspace(1) %ptr, double %value) {2573; COMMON-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode(2574; COMMON-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2575; COMMON-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82576; COMMON-NEXT: br label [[ATOMICRMW_START:%.*]]2577; COMMON: atomicrmw.start:2578; COMMON-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2579; COMMON-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2580; COMMON-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642581; COMMON-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642582; COMMON-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 82583; COMMON-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12584; COMMON-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02585; COMMON-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2586; COMMON-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2587; COMMON: atomicrmw.end:2588; COMMON-NEXT: ret double [[TMP6]]2589;2590 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.ignore.denormal.mode !02591 ret double %res2592}2593 2594define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, double %value) {2595; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2596; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2597; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82598; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2599; GFX803: atomicrmw.start:2600; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2601; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2602; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642603; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642604; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2605; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12606; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02607; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2608; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2609; GFX803: atomicrmw.end:2610; GFX803-NEXT: ret double [[TMP6]]2611;2612; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2613; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2614; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82615; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2616; GFX906: atomicrmw.start:2617; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2618; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2619; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642620; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642621; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2622; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12623; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02624; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2625; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2626; GFX906: atomicrmw.end:2627; GFX906-NEXT: ret double [[TMP6]]2628;2629; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2630; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2631; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82632; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2633; GFX908: atomicrmw.start:2634; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2635; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2636; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642637; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642638; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2639; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12640; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02641; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2642; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2643; GFX908: atomicrmw.end:2644; GFX908-NEXT: ret double [[TMP6]]2645;2646; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2647; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2648; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2649; GFX90A-NEXT: ret double [[RES]]2650;2651; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2652; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2653; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2654; GFX942-NEXT: ret double [[RES]]2655;2656; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2657; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2658; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2659; GFX10-NEXT: ret double [[RES]]2660;2661; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2662; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2663; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82664; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2665; GFX11: atomicrmw.start:2666; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2667; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2668; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642669; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642670; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2671; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12672; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02673; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2674; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2675; GFX11: atomicrmw.end:2676; GFX11-NEXT: ret double [[TMP6]]2677;2678; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory(2679; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2680; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82681; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2682; GFX12: atomicrmw.start:2683; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2684; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2685; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642686; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642687; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]]2688; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12689; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02690; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2691; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2692; GFX12: atomicrmw.end:2693; GFX12-NEXT: ret double [[TMP6]]2694;2695 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.ignore.denormal.mode !02696 ret double %res2697}2698 2699define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {2700; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2701; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2702; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82703; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2704; GFX803: atomicrmw.start:2705; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2706; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2707; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642708; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642709; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2710; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12711; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02712; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2713; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2714; GFX803: atomicrmw.end:2715; GFX803-NEXT: ret double [[TMP6]]2716;2717; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2718; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2719; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82720; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2721; GFX906: atomicrmw.start:2722; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2723; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2724; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642725; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642726; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2727; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12728; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02729; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2730; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2731; GFX906: atomicrmw.end:2732; GFX906-NEXT: ret double [[TMP6]]2733;2734; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2735; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2736; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82737; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2738; GFX908: atomicrmw.start:2739; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2740; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2741; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642742; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642743; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2744; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12745; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02746; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2747; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2748; GFX908: atomicrmw.end:2749; GFX908-NEXT: ret double [[TMP6]]2750;2751; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2752; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2753; GFX90A-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82754; GFX90A-NEXT: br label [[ATOMICRMW_START:%.*]]2755; GFX90A: atomicrmw.start:2756; GFX90A-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2757; GFX90A-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2758; GFX90A-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642759; GFX90A-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642760; GFX90A-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2761; GFX90A-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12762; GFX90A-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02763; GFX90A-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2764; GFX90A-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2765; GFX90A: atomicrmw.end:2766; GFX90A-NEXT: ret double [[TMP6]]2767;2768; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2769; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2770; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2771; GFX942-NEXT: ret double [[RES]]2772;2773; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2774; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2775; GFX10-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82776; GFX10-NEXT: br label [[ATOMICRMW_START:%.*]]2777; GFX10: atomicrmw.start:2778; GFX10-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2779; GFX10-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2780; GFX10-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642781; GFX10-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642782; GFX10-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2783; GFX10-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12784; GFX10-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02785; GFX10-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2786; GFX10-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2787; GFX10: atomicrmw.end:2788; GFX10-NEXT: ret double [[TMP6]]2789;2790; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2791; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2792; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82793; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2794; GFX11: atomicrmw.start:2795; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2796; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2797; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642798; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642799; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2800; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12801; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02802; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2803; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2804; GFX11: atomicrmw.end:2805; GFX11-NEXT: ret double [[TMP6]]2806;2807; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_remote_memory(2808; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2809; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82810; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2811; GFX12: atomicrmw.start:2812; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2813; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2814; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642815; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642816; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.remote.memory [[META0]]2817; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12818; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02819; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2820; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2821; GFX12: atomicrmw.end:2822; GFX12-NEXT: ret double [[TMP6]]2823;2824 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !02825 ret double %res2826}2827 2828define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, double %value) {2829; GFX803-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2830; GFX803-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2831; GFX803-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82832; GFX803-NEXT: br label [[ATOMICRMW_START:%.*]]2833; GFX803: atomicrmw.start:2834; GFX803-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2835; GFX803-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2836; GFX803-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642837; GFX803-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642838; GFX803-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2839; GFX803-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12840; GFX803-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02841; GFX803-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2842; GFX803-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2843; GFX803: atomicrmw.end:2844; GFX803-NEXT: ret double [[TMP6]]2845;2846; GFX906-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2847; GFX906-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2848; GFX906-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82849; GFX906-NEXT: br label [[ATOMICRMW_START:%.*]]2850; GFX906: atomicrmw.start:2851; GFX906-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2852; GFX906-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2853; GFX906-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642854; GFX906-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642855; GFX906-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2856; GFX906-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12857; GFX906-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02858; GFX906-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2859; GFX906-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2860; GFX906: atomicrmw.end:2861; GFX906-NEXT: ret double [[TMP6]]2862;2863; GFX908-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2864; GFX908-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2865; GFX908-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82866; GFX908-NEXT: br label [[ATOMICRMW_START:%.*]]2867; GFX908: atomicrmw.start:2868; GFX908-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2869; GFX908-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2870; GFX908-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642871; GFX908-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642872; GFX908-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2873; GFX908-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12874; GFX908-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02875; GFX908-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2876; GFX908-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2877; GFX908: atomicrmw.end:2878; GFX908-NEXT: ret double [[TMP6]]2879;2880; GFX90A-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2881; GFX90A-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2882; GFX90A-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2883; GFX90A-NEXT: ret double [[RES]]2884;2885; GFX942-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2886; GFX942-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2887; GFX942-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2888; GFX942-NEXT: ret double [[RES]]2889;2890; GFX10-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2891; GFX10-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2892; GFX10-NEXT: [[RES:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], double [[VALUE]] seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]], !amdgpu.ignore.denormal.mode [[META0]]2893; GFX10-NEXT: ret double [[RES]]2894;2895; GFX11-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2896; GFX11-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2897; GFX11-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82898; GFX11-NEXT: br label [[ATOMICRMW_START:%.*]]2899; GFX11: atomicrmw.start:2900; GFX11-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2901; GFX11-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2902; GFX11-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642903; GFX11-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642904; GFX11-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2905; GFX11-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12906; GFX11-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02907; GFX11-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2908; GFX11-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2909; GFX11: atomicrmw.end:2910; GFX11-NEXT: ret double [[TMP6]]2911;2912; GFX12-LABEL: define double @test_atomicrmw_fmin_f64_global_system__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(2913; GFX12-SAME: ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]]) #[[ATTR0]] {2914; GFX12-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR]], align 82915; GFX12-NEXT: br label [[ATOMICRMW_START:%.*]]2916; GFX12: atomicrmw.start:2917; GFX12-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]2918; GFX12-NEXT: [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE]])2919; GFX12-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i642920; GFX12-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i642921; GFX12-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8, !amdgpu.no.fine.grained.memory [[META0]], !amdgpu.no.remote.memory [[META0]]2922; GFX12-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 12923; GFX12-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 02924; GFX12-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double2925; GFX12-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]2926; GFX12: atomicrmw.end:2927; GFX12-NEXT: ret double [[TMP6]]2928;2929 %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst, align 8, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0, !amdgpu.ignore.denormal.mode !02930 ret double %res2931}2932 2933attributes #0 = { "denormal-fp-mode"="preserve-sign,preserve-sign" }2934attributes #1 = { "denormal-fp-mode"="dynamic,dynamic" }2935 2936!0 = !{}2937;.2938; GFX803: [[META0]] = !{}2939;.2940; GFX906: [[META0]] = !{}2941;.2942; GFX908: [[META0]] = !{}2943;.2944; GFX90A: [[META0]] = !{}2945;.2946; GFX942: [[META0]] = !{}2947;.2948; GFX10: [[META0]] = !{}2949;.2950; GFX11: [[META0]] = !{}2951;.2952; GFX12: [[META0]] = !{}2953;.2954