brintos / llvm-project-archived public Read only

1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN:  opt -S -mtriple=amdgcn-- -mcpu=gfx906 -passes='amdgpu-atomic-optimizer<strategy=iterative>,verify<domtree>' %s | FileCheck --check-prefixes=IR,IR-ITERATIVE %s3; RUN:  opt -S -mtriple=amdgcn-- -mcpu=gfx906 -passes='amdgpu-atomic-optimizer<strategy=dpp>,verify<domtree>' %s | FileCheck --check-prefixes=IR,IR-DPP %s4 5; Tests various combinations of uniform/divergent address and uniform/divergent value inputs of various types for atomic operations.6; Optimization remains same for Iterative and DPP strategies when value in uniform. These different scan/reduction7; strategies are valid for only divergent values. This optimization is valid for divergent addresses. Test also covers different scopes.8 9define amdgpu_ps void @global_atomic_fadd_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, float inreg %val) #0 {10; IR-LABEL: @global_atomic_fadd_uni_address_uni_value_agent_scope_unsafe(11; IR-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()12; IR-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]13; IR:       2:14; IR-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)15; IR-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i3216; IR-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 3217; IR-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i3218; IR-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)19; IR-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])20; IR-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]])21; IR-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i3222; IR-NEXT:    [[TMP11:%.*]] = uitofp i32 [[TMP10]] to float23; IR-NEXT:    [[TMP12:%.*]] = fmul float [[VAL:%.*]], [[TMP11]]24; IR-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 025; IR-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]26; IR:       14:27; IR-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] syncscope("agent") monotonic, align 428; IR-NEXT:    br label [[TMP16]]29; IR:       16:30; IR-NEXT:    br label [[TMP17]]31; IR:       17:32; IR-NEXT:    ret void33;34  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic, align 435  ret void36}37 38define amdgpu_ps void @global_atomic_fadd_uni_address_div_value_scope_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, float %val) #0 {39; IR-ITERATIVE-LABEL: @global_atomic_fadd_uni_address_div_value_scope_agent_scope_unsafe(40; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()41; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]42; IR-ITERATIVE:       2:43; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)44; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i3245; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 3246; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i3247; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)48; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])49; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)50; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]51; IR-ITERATIVE:       10:52; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] syncscope("agent") monotonic, align 453; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]54; IR-ITERATIVE:       12:55; IR-ITERATIVE-NEXT:    br label [[TMP13]]56; IR-ITERATIVE:       13:57; IR-ITERATIVE-NEXT:    ret void58; IR-ITERATIVE:       ComputeLoop:59; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]60; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]61; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true)62; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i3263; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]])64; IR-ITERATIVE-NEXT:    [[TMP17]] = fadd float [[ACCUMULATOR]], [[TMP16]]65; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]66; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -167; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]68; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 069; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]70; IR-ITERATIVE:       ComputeEnd:71; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 072; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]73;74; IR-DPP-LABEL: @global_atomic_fadd_uni_address_div_value_scope_agent_scope_unsafe(75; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()76; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]77; IR-DPP:       2:78; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)79; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i3280; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 3281; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i3282; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)83; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])84; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float -0.000000e+00)85; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP9]], i32 273, i32 15, i32 15, i1 false)86; IR-DPP-NEXT:    [[TMP11:%.*]] = fadd float [[TMP9]], [[TMP10]]87; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP11]], i32 274, i32 15, i32 15, i1 false)88; IR-DPP-NEXT:    [[TMP13:%.*]] = fadd float [[TMP11]], [[TMP12]]89; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP13]], i32 276, i32 15, i32 15, i1 false)90; IR-DPP-NEXT:    [[TMP15:%.*]] = fadd float [[TMP13]], [[TMP14]]91; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP15]], i32 280, i32 15, i32 15, i1 false)92; IR-DPP-NEXT:    [[TMP17:%.*]] = fadd float [[TMP15]], [[TMP16]]93; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP17]], i32 322, i32 10, i32 15, i1 false)94; IR-DPP-NEXT:    [[TMP19:%.*]] = fadd float [[TMP17]], [[TMP18]]95; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP19]], i32 323, i32 12, i32 15, i1 false)96; IR-DPP-NEXT:    [[TMP21:%.*]] = fadd float [[TMP19]], [[TMP20]]97; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63)98; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]])99; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0100; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]101; IR-DPP:       25:102; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] syncscope("agent") monotonic, align 4103; IR-DPP-NEXT:    br label [[TMP27]]104; IR-DPP:       27:105; IR-DPP-NEXT:    br label [[TMP28]]106; IR-DPP:       28:107; IR-DPP-NEXT:    ret void108;109  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic, align 4110  ret void111}112 113define amdgpu_ps void @global_atomic_fadd_uni_address_uni_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, float inreg %val) #1 {114; IR-ITERATIVE-LABEL: @global_atomic_fadd_uni_address_uni_value_one_as_scope_unsafe_strictfp(115; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7:[0-9]+]]116; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]117; IR-ITERATIVE:       2:118; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]119; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32120; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32121; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32122; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]123; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]124; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]125; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32126; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]127; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]128; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0129; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]130; IR-ITERATIVE:       14:131; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] syncscope("one-as") monotonic, align 4132; IR-ITERATIVE-NEXT:    br label [[TMP16]]133; IR-ITERATIVE:       16:134; IR-ITERATIVE-NEXT:    br label [[TMP17]]135; IR-ITERATIVE:       17:136; IR-ITERATIVE-NEXT:    ret void137;138; IR-DPP-LABEL: @global_atomic_fadd_uni_address_uni_value_one_as_scope_unsafe_strictfp(139; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8:[0-9]+]]140; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]141; IR-DPP:       2:142; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]143; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32144; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32145; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32146; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]147; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]148; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]149; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32150; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]151; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]152; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0153; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]154; IR-DPP:       14:155; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] syncscope("one-as") monotonic, align 4156; IR-DPP-NEXT:    br label [[TMP16]]157; IR-DPP:       16:158; IR-DPP-NEXT:    br label [[TMP17]]159; IR-DPP:       17:160; IR-DPP-NEXT:    ret void161;162  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic163  ret void164}165 166define amdgpu_ps void @global_atomic_fadd_uni_address_div_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, float %val) #1 {167; IR-ITERATIVE-LABEL: @global_atomic_fadd_uni_address_div_value_one_as_scope_unsafe_strictfp(168; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]169; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]170; IR-ITERATIVE:       2:171; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]172; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32173; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32174; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32175; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]176; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]177; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]178; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]179; IR-ITERATIVE:       10:180; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] syncscope("one-as") monotonic, align 4181; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]182; IR-ITERATIVE:       12:183; IR-ITERATIVE-NEXT:    br label [[TMP13]]184; IR-ITERATIVE:       13:185; IR-ITERATIVE-NEXT:    ret void186; IR-ITERATIVE:       ComputeLoop:187; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]188; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]189; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]190; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32191; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]192; IR-ITERATIVE-NEXT:    [[TMP17]] = call float @llvm.experimental.constrained.fadd.f32(float [[ACCUMULATOR]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]193; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]194; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1195; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]196; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0197; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]198; IR-ITERATIVE:       ComputeEnd:199; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0200; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]201;202; IR-DPP-LABEL: @global_atomic_fadd_uni_address_div_value_one_as_scope_unsafe_strictfp(203; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]204; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]205; IR-DPP:       2:206; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]207; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32208; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32209; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32210; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]211; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]212; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float -0.000000e+00) #[[ATTR8]]213; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]214; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP9]], float [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]215; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]216; IR-DPP-NEXT:    [[TMP13:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP11]], float [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]217; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]218; IR-DPP-NEXT:    [[TMP15:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP13]], float [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]219; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]220; IR-DPP-NEXT:    [[TMP17:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP15]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]221; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]222; IR-DPP-NEXT:    [[TMP19:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP17]], float [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]223; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]224; IR-DPP-NEXT:    [[TMP21:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP19]], float [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]225; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63) #[[ATTR8]]226; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]]) #[[ATTR8]]227; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0228; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]229; IR-DPP:       25:230; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] syncscope("one-as") monotonic, align 4231; IR-DPP-NEXT:    br label [[TMP27]]232; IR-DPP:       27:233; IR-DPP-NEXT:    br label [[TMP28]]234; IR-DPP:       28:235; IR-DPP-NEXT:    ret void236;237  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic238  ret void239}240 241define amdgpu_ps void @global_atomic_fsub_uni_address_uni_value_agent_scope_strictfp(ptr addrspace(1) inreg %ptr, float inreg %val) #2 {242; IR-ITERATIVE-LABEL: @global_atomic_fsub_uni_address_uni_value_agent_scope_strictfp(243; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]244; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]245; IR-ITERATIVE:       2:246; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]247; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32248; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32249; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32250; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]251; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]252; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]253; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32254; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]255; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]256; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0257; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]258; IR-ITERATIVE:       14:259; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] syncscope("agent") monotonic, align 4260; IR-ITERATIVE-NEXT:    br label [[TMP16]]261; IR-ITERATIVE:       16:262; IR-ITERATIVE-NEXT:    br label [[TMP17]]263; IR-ITERATIVE:       17:264; IR-ITERATIVE-NEXT:    ret void265;266; IR-DPP-LABEL: @global_atomic_fsub_uni_address_uni_value_agent_scope_strictfp(267; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]268; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]269; IR-DPP:       2:270; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]271; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32272; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32273; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32274; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]275; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]276; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]277; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32278; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]279; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]280; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0281; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]282; IR-DPP:       14:283; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] syncscope("agent") monotonic, align 4284; IR-DPP-NEXT:    br label [[TMP16]]285; IR-DPP:       16:286; IR-DPP-NEXT:    br label [[TMP17]]287; IR-DPP:       17:288; IR-DPP-NEXT:    ret void289;290  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic291  ret void292}293 294define amdgpu_ps void @global_atomic_fsub_uni_address_div_value_agent_scope_strictfp(ptr addrspace(1) inreg %ptr, float %val) #2 {295; IR-ITERATIVE-LABEL: @global_atomic_fsub_uni_address_div_value_agent_scope_strictfp(296; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]297; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]298; IR-ITERATIVE:       2:299; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]300; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32301; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32302; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32303; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]304; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]305; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]306; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]307; IR-ITERATIVE:       10:308; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] syncscope("agent") monotonic, align 4309; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]310; IR-ITERATIVE:       12:311; IR-ITERATIVE-NEXT:    br label [[TMP13]]312; IR-ITERATIVE:       13:313; IR-ITERATIVE-NEXT:    ret void314; IR-ITERATIVE:       ComputeLoop:315; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]316; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]317; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]318; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32319; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]320; IR-ITERATIVE-NEXT:    [[TMP17]] = call float @llvm.experimental.constrained.fadd.f32(float [[ACCUMULATOR]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]321; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]322; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1323; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]324; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0325; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]326; IR-ITERATIVE:       ComputeEnd:327; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0328; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]329;330; IR-DPP-LABEL: @global_atomic_fsub_uni_address_div_value_agent_scope_strictfp(331; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]332; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]333; IR-DPP:       2:334; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]335; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32336; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32337; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32338; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]339; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]340; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float -0.000000e+00) #[[ATTR8]]341; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]342; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP9]], float [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]343; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]344; IR-DPP-NEXT:    [[TMP13:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP11]], float [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]345; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]346; IR-DPP-NEXT:    [[TMP15:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP13]], float [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]347; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]348; IR-DPP-NEXT:    [[TMP17:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP15]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]349; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]350; IR-DPP-NEXT:    [[TMP19:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP17]], float [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]351; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]352; IR-DPP-NEXT:    [[TMP21:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP19]], float [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]353; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63) #[[ATTR8]]354; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]]) #[[ATTR8]]355; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0356; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]357; IR-DPP:       25:358; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] syncscope("agent") monotonic, align 4359; IR-DPP-NEXT:    br label [[TMP27]]360; IR-DPP:       27:361; IR-DPP-NEXT:    br label [[TMP28]]362; IR-DPP:       28:363; IR-DPP-NEXT:    ret void364;365  %result = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic366  ret void367}368 369define amdgpu_ps void @global_atomic_fmin_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, float inreg %val) #0 {370; IR-LABEL: @global_atomic_fmin_uni_address_uni_value_agent_scope_unsafe(371; IR-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()372; IR-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]373; IR:       2:374; IR-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)375; IR-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32376; IR-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32377; IR-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32378; IR-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)379; IR-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])380; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 0381; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]382; IR:       10:383; IR-NEXT:    [[TMP11:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4384; IR-NEXT:    br label [[TMP12]]385; IR:       12:386; IR-NEXT:    br label [[TMP13]]387; IR:       13:388; IR-NEXT:    ret void389;390  %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic391  ret void392}393 394define amdgpu_ps void @global_atomic_fmin_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, float %val) #0 {395; IR-ITERATIVE-LABEL: @global_atomic_fmin_uni_address_div_value_agent_scope_unsafe(396; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()397; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]398; IR-ITERATIVE:       2:399; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)400; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32401; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32402; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32403; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)404; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])405; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)406; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]407; IR-ITERATIVE:       10:408; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] syncscope("agent") monotonic, align 4409; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]410; IR-ITERATIVE:       12:411; IR-ITERATIVE-NEXT:    br label [[TMP13]]412; IR-ITERATIVE:       13:413; IR-ITERATIVE-NEXT:    ret void414; IR-ITERATIVE:       ComputeLoop:415; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ 0x7FF8000000000000, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]416; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]417; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true)418; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32419; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]])420; IR-ITERATIVE-NEXT:    [[TMP17]] = call float @llvm.minnum.f32(float [[ACCUMULATOR]], float [[TMP16]])421; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]422; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1423; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]424; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0425; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]426; IR-ITERATIVE:       ComputeEnd:427; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0428; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]429;430; IR-DPP-LABEL: @global_atomic_fmin_uni_address_div_value_agent_scope_unsafe(431; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()432; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]433; IR-DPP:       2:434; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)435; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32436; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32437; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32438; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)439; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])440; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float 0x7FF8000000000000)441; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP9]], i32 273, i32 15, i32 15, i1 false)442; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.minnum.f32(float [[TMP9]], float [[TMP10]])443; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP11]], i32 274, i32 15, i32 15, i1 false)444; IR-DPP-NEXT:    [[TMP13:%.*]] = call float @llvm.minnum.f32(float [[TMP11]], float [[TMP12]])445; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP13]], i32 276, i32 15, i32 15, i1 false)446; IR-DPP-NEXT:    [[TMP15:%.*]] = call float @llvm.minnum.f32(float [[TMP13]], float [[TMP14]])447; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP15]], i32 280, i32 15, i32 15, i1 false)448; IR-DPP-NEXT:    [[TMP17:%.*]] = call float @llvm.minnum.f32(float [[TMP15]], float [[TMP16]])449; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP17]], i32 322, i32 10, i32 15, i1 false)450; IR-DPP-NEXT:    [[TMP19:%.*]] = call float @llvm.minnum.f32(float [[TMP17]], float [[TMP18]])451; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP19]], i32 323, i32 12, i32 15, i1 false)452; IR-DPP-NEXT:    [[TMP21:%.*]] = call float @llvm.minnum.f32(float [[TMP19]], float [[TMP20]])453; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63)454; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]])455; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0456; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]457; IR-DPP:       25:458; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] syncscope("agent") monotonic, align 4459; IR-DPP-NEXT:    br label [[TMP27]]460; IR-DPP:       27:461; IR-DPP-NEXT:    br label [[TMP28]]462; IR-DPP:       28:463; IR-DPP-NEXT:    ret void464;465  %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic466  ret void467}468 469define amdgpu_ps void @global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, float inreg %val) #1{470; IR-ITERATIVE-LABEL: @global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe_strictfp(471; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]472; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]473; IR-ITERATIVE:       2:474; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]475; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32476; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32477; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32478; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]479; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]480; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 0481; IR-ITERATIVE-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]482; IR-ITERATIVE:       10:483; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4484; IR-ITERATIVE-NEXT:    br label [[TMP12]]485; IR-ITERATIVE:       12:486; IR-ITERATIVE-NEXT:    br label [[TMP13]]487; IR-ITERATIVE:       13:488; IR-ITERATIVE-NEXT:    ret void489;490; IR-DPP-LABEL: @global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe_strictfp(491; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]492; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]493; IR-DPP:       2:494; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]495; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32496; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32497; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32498; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]499; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]500; IR-DPP-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 0501; IR-DPP-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]502; IR-DPP:       10:503; IR-DPP-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4504; IR-DPP-NEXT:    br label [[TMP12]]505; IR-DPP:       12:506; IR-DPP-NEXT:    br label [[TMP13]]507; IR-DPP:       13:508; IR-DPP-NEXT:    ret void509;510  %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic511  ret void512}513 514define amdgpu_ps void @global_atomic_fmax_uni_address_div_value_agent_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, float %val) #1{515; IR-ITERATIVE-LABEL: @global_atomic_fmax_uni_address_div_value_agent_scope_unsafe_strictfp(516; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]517; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]518; IR-ITERATIVE:       2:519; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]520; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32521; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32522; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32523; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]524; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]525; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]526; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]527; IR-ITERATIVE:       10:528; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] syncscope("agent") monotonic, align 4529; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]530; IR-ITERATIVE:       12:531; IR-ITERATIVE-NEXT:    br label [[TMP13]]532; IR-ITERATIVE:       13:533; IR-ITERATIVE-NEXT:    ret void534; IR-ITERATIVE:       ComputeLoop:535; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ 0x7FF8000000000000, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]536; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]537; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]538; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32539; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]540; IR-ITERATIVE-NEXT:    [[TMP17]] = call float @llvm.experimental.constrained.maxnum.f32(float [[ACCUMULATOR]], float [[TMP16]], metadata !"fpexcept.strict") #[[ATTR7]]541; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]542; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1543; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]544; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0545; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]546; IR-ITERATIVE:       ComputeEnd:547; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0548; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]549;550; IR-DPP-LABEL: @global_atomic_fmax_uni_address_div_value_agent_scope_unsafe_strictfp(551; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]552; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]553; IR-DPP:       2:554; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]555; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32556; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32557; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32558; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]559; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]560; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float 0x7FF8000000000000) #[[ATTR8]]561; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]562; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP9]], float [[TMP10]], metadata !"fpexcept.strict") #[[ATTR8]]563; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]564; IR-DPP-NEXT:    [[TMP13:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP11]], float [[TMP12]], metadata !"fpexcept.strict") #[[ATTR8]]565; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]566; IR-DPP-NEXT:    [[TMP15:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP13]], float [[TMP14]], metadata !"fpexcept.strict") #[[ATTR8]]567; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]568; IR-DPP-NEXT:    [[TMP17:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP15]], float [[TMP16]], metadata !"fpexcept.strict") #[[ATTR8]]569; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]570; IR-DPP-NEXT:    [[TMP19:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP17]], float [[TMP18]], metadata !"fpexcept.strict") #[[ATTR8]]571; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float 0x7FF8000000000000, float [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]572; IR-DPP-NEXT:    [[TMP21:%.*]] = call float @llvm.experimental.constrained.maxnum.f32(float [[TMP19]], float [[TMP20]], metadata !"fpexcept.strict") #[[ATTR8]]573; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63) #[[ATTR8]]574; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]]) #[[ATTR8]]575; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0576; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]577; IR-DPP:       25:578; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] syncscope("agent") monotonic, align 4579; IR-DPP-NEXT:    br label [[TMP27]]580; IR-DPP:       27:581; IR-DPP-NEXT:    br label [[TMP28]]582; IR-DPP:       28:583; IR-DPP-NEXT:    ret void584;585  %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic586  ret void587}588 589define amdgpu_ps void @global_atomic_fadd_uni_address_uni_value_system_scope_strictfp(ptr addrspace(1) inreg %ptr, float inreg %val) #2 {590; IR-ITERATIVE-LABEL: @global_atomic_fadd_uni_address_uni_value_system_scope_strictfp(591; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]592; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]593; IR-ITERATIVE:       2:594; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]595; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32596; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32597; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32598; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]599; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]600; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]601; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32602; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]603; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]604; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0605; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]606; IR-ITERATIVE:       14:607; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] monotonic, align 4608; IR-ITERATIVE-NEXT:    br label [[TMP16]]609; IR-ITERATIVE:       16:610; IR-ITERATIVE-NEXT:    br label [[TMP17]]611; IR-ITERATIVE:       17:612; IR-ITERATIVE-NEXT:    ret void613;614; IR-DPP-LABEL: @global_atomic_fadd_uni_address_uni_value_system_scope_strictfp(615; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]616; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]617; IR-DPP:       2:618; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]619; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32620; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32621; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32622; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]623; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]624; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]625; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32626; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.uitofp.f32.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]627; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.experimental.constrained.fmul.f32(float [[VAL:%.*]], float [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]628; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0629; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]630; IR-DPP:       14:631; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP12]] monotonic, align 4632; IR-DPP-NEXT:    br label [[TMP16]]633; IR-DPP:       16:634; IR-DPP-NEXT:    br label [[TMP17]]635; IR-DPP:       17:636; IR-DPP-NEXT:    ret void637;638  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val monotonic, align 4639  ret void640}641 642define amdgpu_ps void @global_atomic_fadd_uni_address_div_value_system_scope_strictfp(ptr addrspace(1) inreg %ptr, float %val) #2 {643; IR-ITERATIVE-LABEL: @global_atomic_fadd_uni_address_div_value_system_scope_strictfp(644; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]645; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]646; IR-ITERATIVE:       2:647; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]648; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32649; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32650; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32651; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]652; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]653; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]654; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]655; IR-ITERATIVE:       10:656; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP17:%.*]] monotonic, align 4657; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]658; IR-ITERATIVE:       12:659; IR-ITERATIVE-NEXT:    br label [[TMP13]]660; IR-ITERATIVE:       13:661; IR-ITERATIVE-NEXT:    ret void662; IR-ITERATIVE:       ComputeLoop:663; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi float [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]664; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]665; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]666; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32667; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]668; IR-ITERATIVE-NEXT:    [[TMP17]] = call float @llvm.experimental.constrained.fadd.f32(float [[ACCUMULATOR]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]669; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]670; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1671; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]672; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0673; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]674; IR-ITERATIVE:       ComputeEnd:675; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0676; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]677;678; IR-DPP-LABEL: @global_atomic_fadd_uni_address_div_value_system_scope_strictfp(679; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]680; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]681; IR-DPP:       2:682; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]683; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32684; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32685; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32686; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]687; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]688; IR-DPP-NEXT:    [[TMP9:%.*]] = call float @llvm.amdgcn.set.inactive.f32(float [[VAL:%.*]], float -0.000000e+00) #[[ATTR8]]689; IR-DPP-NEXT:    [[TMP10:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]690; IR-DPP-NEXT:    [[TMP11:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP9]], float [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]691; IR-DPP-NEXT:    [[TMP12:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]692; IR-DPP-NEXT:    [[TMP13:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP11]], float [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]693; IR-DPP-NEXT:    [[TMP14:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]694; IR-DPP-NEXT:    [[TMP15:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP13]], float [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]695; IR-DPP-NEXT:    [[TMP16:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]696; IR-DPP-NEXT:    [[TMP17:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP15]], float [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]697; IR-DPP-NEXT:    [[TMP18:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]698; IR-DPP-NEXT:    [[TMP19:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP17]], float [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]699; IR-DPP-NEXT:    [[TMP20:%.*]] = call float @llvm.amdgcn.update.dpp.f32(float -0.000000e+00, float [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]700; IR-DPP-NEXT:    [[TMP21:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[TMP19]], float [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]701; IR-DPP-NEXT:    [[TMP22:%.*]] = call float @llvm.amdgcn.readlane.f32(float [[TMP21]], i32 63) #[[ATTR8]]702; IR-DPP-NEXT:    [[TMP23:%.*]] = call float @llvm.amdgcn.strict.wwm.f32(float [[TMP22]]) #[[ATTR8]]703; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0704; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]705; IR-DPP:       25:706; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[TMP23]] monotonic, align 4707; IR-DPP-NEXT:    br label [[TMP27]]708; IR-DPP:       27:709; IR-DPP-NEXT:    br label [[TMP28]]710; IR-DPP:       28:711; IR-DPP-NEXT:    ret void712;713  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val monotonic, align 4714  ret void715}716 717define amdgpu_ps void @global_atomic_fadd_div_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr, float inreg %val) #0 {718; IR-LABEL: @global_atomic_fadd_div_address_uni_value_agent_scope_unsafe(719; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4720; IR-NEXT:    ret void721;722  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic, align 4723  ret void724}725 726define amdgpu_ps void @global_atomic_fadd_div_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr, float %val) #0 {727; IR-LABEL: @global_atomic_fadd_div_address_div_value_agent_scope_unsafe(728; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4729; IR-NEXT:    ret void730;731  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic, align 4732  ret void733}734 735define amdgpu_ps void @global_atomic_fadd_div_address_uni_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) %ptr, float inreg %val) #1 {736; IR-LABEL: @global_atomic_fadd_div_address_uni_value_one_as_scope_unsafe_strictfp(737; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("one-as") monotonic, align 4738; IR-NEXT:    ret void739;740  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic741  ret void742}743 744define amdgpu_ps void @global_atomic_fadd_div_address_div_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) %ptr, float %val) #1 {745; IR-LABEL: @global_atomic_fadd_div_address_div_value_one_as_scope_unsafe_strictfp(746; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("one-as") monotonic, align 4747; IR-NEXT:    ret void748;749  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic750  ret void751}752 753define amdgpu_ps void @global_atomic_fsub_div_address_uni_value_agent_scope_strictfp(ptr addrspace(1) %ptr, float inreg %val) #2 {754; IR-LABEL: @global_atomic_fsub_div_address_uni_value_agent_scope_strictfp(755; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4756; IR-NEXT:    ret void757;758  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic759  ret void760}761 762define amdgpu_ps void @global_atomic_fsub_div_address_div_value_agent_scope_strictfp(ptr addrspace(1) %ptr, float %val) #2 {763; IR-LABEL: @global_atomic_fsub_div_address_div_value_agent_scope_strictfp(764; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4765; IR-NEXT:    ret void766;767  %result = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic768  ret void769}770 771define amdgpu_ps void @global_atomic_fmin_div_address_uni_value_agent_scope(ptr addrspace(1) %ptr, float inreg %val) #0 {772; IR-LABEL: @global_atomic_fmin_div_address_uni_value_agent_scope(773; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4774; IR-NEXT:    ret void775;776  %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic777  ret void778}779 780define amdgpu_ps void @global_atomic_fmin_div_address_div_value_agent_scope(ptr addrspace(1) %ptr, float %val) #0 {781; IR-LABEL: @global_atomic_fmin_div_address_div_value_agent_scope(782; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4783; IR-NEXT:    ret void784;785  %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic786  ret void787}788 789define amdgpu_ps void @global_atomic_fmax_div_address_uni_value_agent_scope_unsafe_strictfp(ptr addrspace(1) %ptr, float inreg %val) #1{790; IR-LABEL: @global_atomic_fmax_div_address_uni_value_agent_scope_unsafe_strictfp(791; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4792; IR-NEXT:    ret void793;794  %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic795  ret void796}797 798define amdgpu_ps void @global_atomic_fmax_div_address_div_value_agent_scope_unsafe_strictfp(ptr addrspace(1) %ptr, float %val) #1{799; IR-LABEL: @global_atomic_fmax_div_address_div_value_agent_scope_unsafe_strictfp(800; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] syncscope("agent") monotonic, align 4801; IR-NEXT:    ret void802;803  %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") monotonic804  ret void805}806 807define amdgpu_ps void @global_atomic_fadd_div_address_uni_value_system_scope_strictfp(ptr addrspace(1) %ptr, float inreg %val) #2 {808; IR-LABEL: @global_atomic_fadd_div_address_uni_value_system_scope_strictfp(809; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] monotonic, align 4810; IR-NEXT:    ret void811;812  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val monotonic, align 4813  ret void814}815 816define amdgpu_ps void @global_atomic_fadd_div_address_div_value_system_scope_strictfp(ptr addrspace(1) %ptr, float %val) #2 {817; IR-LABEL: @global_atomic_fadd_div_address_div_value_system_scope_strictfp(818; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]] monotonic, align 4819; IR-NEXT:    ret void820;821  %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val monotonic, align 4822  ret void823}824 825define amdgpu_ps void @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, double inreg %val) #0 {826; IR-LABEL: @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(827; IR-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()828; IR-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]829; IR:       2:830; IR-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)831; IR-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32832; IR-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32833; IR-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32834; IR-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)835; IR-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])836; IR-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]])837; IR-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32838; IR-NEXT:    [[TMP11:%.*]] = uitofp i32 [[TMP10]] to double839; IR-NEXT:    [[TMP12:%.*]] = fmul double [[VAL:%.*]], [[TMP11]]840; IR-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0841; IR-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]842; IR:       14:843; IR-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("agent") monotonic, align 4844; IR-NEXT:    br label [[TMP16]]845; IR:       16:846; IR-NEXT:    br label [[TMP17]]847; IR:       17:848; IR-NEXT:    ret void849;850  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic, align 4851  ret void852}853 854define amdgpu_ps void @global_atomic_fadd_double_uni_address_div_value_scope_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, double %val) #0 {855; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_div_value_scope_agent_scope_unsafe(856; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()857; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]858; IR-ITERATIVE:       2:859; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)860; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32861; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32862; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32863; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)864; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])865; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)866; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]867; IR-ITERATIVE:       10:868; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] syncscope("agent") monotonic, align 4869; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]870; IR-ITERATIVE:       12:871; IR-ITERATIVE-NEXT:    br label [[TMP13]]872; IR-ITERATIVE:       13:873; IR-ITERATIVE-NEXT:    ret void874; IR-ITERATIVE:       ComputeLoop:875; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]876; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]877; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true)878; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i32879; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]])880; IR-ITERATIVE-NEXT:    [[TMP17]] = fadd double [[ACCUMULATOR]], [[TMP16]]881; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]882; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -1883; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]884; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 0885; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]886; IR-ITERATIVE:       ComputeEnd:887; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 0888; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]889;890; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_div_value_scope_agent_scope_unsafe(891; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()892; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]893; IR-DPP:       2:894; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)895; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32896; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32897; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32898; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)899; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])900; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double -0.000000e+00)901; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP9]], i32 273, i32 15, i32 15, i1 false)902; IR-DPP-NEXT:    [[TMP11:%.*]] = fadd double [[TMP9]], [[TMP10]]903; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP11]], i32 274, i32 15, i32 15, i1 false)904; IR-DPP-NEXT:    [[TMP13:%.*]] = fadd double [[TMP11]], [[TMP12]]905; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP13]], i32 276, i32 15, i32 15, i1 false)906; IR-DPP-NEXT:    [[TMP15:%.*]] = fadd double [[TMP13]], [[TMP14]]907; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP15]], i32 280, i32 15, i32 15, i1 false)908; IR-DPP-NEXT:    [[TMP17:%.*]] = fadd double [[TMP15]], [[TMP16]]909; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP17]], i32 322, i32 10, i32 15, i1 false)910; IR-DPP-NEXT:    [[TMP19:%.*]] = fadd double [[TMP17]], [[TMP18]]911; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP19]], i32 323, i32 12, i32 15, i1 false)912; IR-DPP-NEXT:    [[TMP21:%.*]] = fadd double [[TMP19]], [[TMP20]]913; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63)914; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]])915; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 0916; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]917; IR-DPP:       25:918; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] syncscope("agent") monotonic, align 4919; IR-DPP-NEXT:    br label [[TMP27]]920; IR-DPP:       27:921; IR-DPP-NEXT:    br label [[TMP28]]922; IR-DPP:       28:923; IR-DPP-NEXT:    ret void924;925  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic, align 4926  ret void927}928 929define amdgpu_ps void @global_atomic_fadd_double_uni_address_uni_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, double inreg %val) #1 {930; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_uni_value_one_as_scope_unsafe_strictfp(931; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]932; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]933; IR-ITERATIVE:       2:934; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]935; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32936; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32937; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32938; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]939; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]940; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]941; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32942; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]943; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]944; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0945; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]946; IR-ITERATIVE:       14:947; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("one-as") monotonic, align 8948; IR-ITERATIVE-NEXT:    br label [[TMP16]]949; IR-ITERATIVE:       16:950; IR-ITERATIVE-NEXT:    br label [[TMP17]]951; IR-ITERATIVE:       17:952; IR-ITERATIVE-NEXT:    ret void953;954; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_uni_value_one_as_scope_unsafe_strictfp(955; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]956; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]957; IR-DPP:       2:958; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]959; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32960; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32961; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32962; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]963; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]964; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]965; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32966; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]967; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]968; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0969; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]970; IR-DPP:       14:971; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("one-as") monotonic, align 8972; IR-DPP-NEXT:    br label [[TMP16]]973; IR-DPP:       16:974; IR-DPP-NEXT:    br label [[TMP17]]975; IR-DPP:       17:976; IR-DPP-NEXT:    ret void977;978  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic979  ret void980}981 982define amdgpu_ps void @global_atomic_fadd_double_uni_address_div_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, double %val) #1 {983; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_div_value_one_as_scope_unsafe_strictfp(984; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]985; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]986; IR-ITERATIVE:       2:987; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]988; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32989; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32990; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32991; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]992; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]993; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]994; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]995; IR-ITERATIVE:       10:996; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] syncscope("one-as") monotonic, align 8997; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]998; IR-ITERATIVE:       12:999; IR-ITERATIVE-NEXT:    br label [[TMP13]]1000; IR-ITERATIVE:       13:1001; IR-ITERATIVE-NEXT:    ret void1002; IR-ITERATIVE:       ComputeLoop:1003; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]1004; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]1005; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]1006; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i321007; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]1008; IR-ITERATIVE-NEXT:    [[TMP17]] = call double @llvm.experimental.constrained.fadd.f64(double [[ACCUMULATOR]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1009; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]1010; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -11011; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]1012; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 01013; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]1014; IR-ITERATIVE:       ComputeEnd:1015; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 01016; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]1017;1018; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_div_value_one_as_scope_unsafe_strictfp(1019; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1020; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]1021; IR-DPP:       2:1022; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1023; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321024; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321025; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321026; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1027; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1028; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double -0.000000e+00) #[[ATTR8]]1029; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]1030; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP9]], double [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1031; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]1032; IR-DPP-NEXT:    [[TMP13:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP11]], double [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1033; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]1034; IR-DPP-NEXT:    [[TMP15:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP13]], double [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1035; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]1036; IR-DPP-NEXT:    [[TMP17:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP15]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1037; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]1038; IR-DPP-NEXT:    [[TMP19:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP17]], double [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1039; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]1040; IR-DPP-NEXT:    [[TMP21:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP19]], double [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1041; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63) #[[ATTR8]]1042; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]]) #[[ATTR8]]1043; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 01044; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]1045; IR-DPP:       25:1046; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] syncscope("one-as") monotonic, align 81047; IR-DPP-NEXT:    br label [[TMP27]]1048; IR-DPP:       27:1049; IR-DPP-NEXT:    br label [[TMP28]]1050; IR-DPP:       28:1051; IR-DPP-NEXT:    ret void1052;1053  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic1054  ret void1055}1056 1057define amdgpu_ps void @global_atomic_fsub_double_uni_address_uni_value_agent_scope_strictfp(ptr addrspace(1) inreg %ptr, double inreg %val) #2 {1058; IR-ITERATIVE-LABEL: @global_atomic_fsub_double_uni_address_uni_value_agent_scope_strictfp(1059; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1060; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]1061; IR-ITERATIVE:       2:1062; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1063; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321064; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321065; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321066; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1067; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1068; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]1069; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i321070; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1071; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1072; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 01073; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]1074; IR-ITERATIVE:       14:1075; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("agent") monotonic, align 81076; IR-ITERATIVE-NEXT:    br label [[TMP16]]1077; IR-ITERATIVE:       16:1078; IR-ITERATIVE-NEXT:    br label [[TMP17]]1079; IR-ITERATIVE:       17:1080; IR-ITERATIVE-NEXT:    ret void1081;1082; IR-DPP-LABEL: @global_atomic_fsub_double_uni_address_uni_value_agent_scope_strictfp(1083; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1084; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]1085; IR-DPP:       2:1086; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1087; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321088; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321089; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321090; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1091; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1092; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]1093; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i321094; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1095; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1096; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 01097; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]1098; IR-DPP:       14:1099; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("agent") monotonic, align 81100; IR-DPP-NEXT:    br label [[TMP16]]1101; IR-DPP:       16:1102; IR-DPP-NEXT:    br label [[TMP17]]1103; IR-DPP:       17:1104; IR-DPP-NEXT:    ret void1105;1106  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1107  ret void1108}1109 1110define amdgpu_ps void @global_atomic_fsub_double_uni_address_div_value_agent_scope_strictfp(ptr addrspace(1) inreg %ptr, double %val) #2 {1111; IR-ITERATIVE-LABEL: @global_atomic_fsub_double_uni_address_div_value_agent_scope_strictfp(1112; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1113; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1114; IR-ITERATIVE:       2:1115; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1116; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321117; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321118; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321119; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1120; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1121; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1122; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]1123; IR-ITERATIVE:       10:1124; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] syncscope("agent") monotonic, align 81125; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]1126; IR-ITERATIVE:       12:1127; IR-ITERATIVE-NEXT:    br label [[TMP13]]1128; IR-ITERATIVE:       13:1129; IR-ITERATIVE-NEXT:    ret void1130; IR-ITERATIVE:       ComputeLoop:1131; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]1132; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]1133; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]1134; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i321135; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]1136; IR-ITERATIVE-NEXT:    [[TMP17]] = call double @llvm.experimental.constrained.fadd.f64(double [[ACCUMULATOR]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1137; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]1138; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -11139; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]1140; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 01141; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]1142; IR-ITERATIVE:       ComputeEnd:1143; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 01144; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]1145;1146; IR-DPP-LABEL: @global_atomic_fsub_double_uni_address_div_value_agent_scope_strictfp(1147; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1148; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]1149; IR-DPP:       2:1150; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1151; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321152; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321153; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321154; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1155; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1156; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double -0.000000e+00) #[[ATTR8]]1157; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]1158; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP9]], double [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1159; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]1160; IR-DPP-NEXT:    [[TMP13:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP11]], double [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1161; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]1162; IR-DPP-NEXT:    [[TMP15:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP13]], double [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1163; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]1164; IR-DPP-NEXT:    [[TMP17:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP15]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1165; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]1166; IR-DPP-NEXT:    [[TMP19:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP17]], double [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1167; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]1168; IR-DPP-NEXT:    [[TMP21:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP19]], double [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1169; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63) #[[ATTR8]]1170; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]]) #[[ATTR8]]1171; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 01172; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]1173; IR-DPP:       25:1174; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] syncscope("agent") monotonic, align 81175; IR-DPP-NEXT:    br label [[TMP27]]1176; IR-DPP:       27:1177; IR-DPP-NEXT:    br label [[TMP28]]1178; IR-DPP:       28:1179; IR-DPP-NEXT:    ret void1180;1181  %result = atomicrmw fsub ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1182  ret void1183}1184 1185define amdgpu_ps void @global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, double inreg %val) #0 {1186; IR-LABEL: @global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe(1187; IR-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()1188; IR-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1189; IR:       2:1190; IR-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)1191; IR-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321192; IR-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321193; IR-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321194; IR-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)1195; IR-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])1196; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 01197; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]1198; IR:       10:1199; IR-NEXT:    [[TMP11:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81200; IR-NEXT:    br label [[TMP12]]1201; IR:       12:1202; IR-NEXT:    br label [[TMP13]]1203; IR:       13:1204; IR-NEXT:    ret void1205;1206  %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1207  ret void1208}1209 1210define amdgpu_ps void @global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, double %val) #0 {1211; IR-ITERATIVE-LABEL: @global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe(1212; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()1213; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1214; IR-ITERATIVE:       2:1215; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)1216; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321217; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321218; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321219; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)1220; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])1221; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)1222; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]1223; IR-ITERATIVE:       10:1224; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] syncscope("agent") monotonic, align 81225; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]1226; IR-ITERATIVE:       12:1227; IR-ITERATIVE-NEXT:    br label [[TMP13]]1228; IR-ITERATIVE:       13:1229; IR-ITERATIVE-NEXT:    ret void1230; IR-ITERATIVE:       ComputeLoop:1231; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ 0x7FF8000000000000, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]1232; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]1233; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true)1234; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i321235; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]])1236; IR-ITERATIVE-NEXT:    [[TMP17]] = call double @llvm.minnum.f64(double [[ACCUMULATOR]], double [[TMP16]])1237; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]1238; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -11239; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]1240; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 01241; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]1242; IR-ITERATIVE:       ComputeEnd:1243; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 01244; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]1245;1246; IR-DPP-LABEL: @global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe(1247; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()1248; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]1249; IR-DPP:       2:1250; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)1251; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321252; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321253; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321254; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)1255; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])1256; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double 0x7FF8000000000000)1257; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP9]], i32 273, i32 15, i32 15, i1 false)1258; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.minnum.f64(double [[TMP9]], double [[TMP10]])1259; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP11]], i32 274, i32 15, i32 15, i1 false)1260; IR-DPP-NEXT:    [[TMP13:%.*]] = call double @llvm.minnum.f64(double [[TMP11]], double [[TMP12]])1261; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP13]], i32 276, i32 15, i32 15, i1 false)1262; IR-DPP-NEXT:    [[TMP15:%.*]] = call double @llvm.minnum.f64(double [[TMP13]], double [[TMP14]])1263; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP15]], i32 280, i32 15, i32 15, i1 false)1264; IR-DPP-NEXT:    [[TMP17:%.*]] = call double @llvm.minnum.f64(double [[TMP15]], double [[TMP16]])1265; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP17]], i32 322, i32 10, i32 15, i1 false)1266; IR-DPP-NEXT:    [[TMP19:%.*]] = call double @llvm.minnum.f64(double [[TMP17]], double [[TMP18]])1267; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP19]], i32 323, i32 12, i32 15, i1 false)1268; IR-DPP-NEXT:    [[TMP21:%.*]] = call double @llvm.minnum.f64(double [[TMP19]], double [[TMP20]])1269; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63)1270; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]])1271; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 01272; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]1273; IR-DPP:       25:1274; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] syncscope("agent") monotonic, align 81275; IR-DPP-NEXT:    br label [[TMP27]]1276; IR-DPP:       27:1277; IR-DPP-NEXT:    br label [[TMP28]]1278; IR-DPP:       28:1279; IR-DPP-NEXT:    ret void1280;1281  %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1282  ret void1283}1284 1285define amdgpu_ps void @global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, double inreg %val) #1{1286; IR-ITERATIVE-LABEL: @global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe_strictfp(1287; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1288; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1289; IR-ITERATIVE:       2:1290; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1291; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321292; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321293; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321294; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1295; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1296; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 01297; IR-ITERATIVE-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]1298; IR-ITERATIVE:       10:1299; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81300; IR-ITERATIVE-NEXT:    br label [[TMP12]]1301; IR-ITERATIVE:       12:1302; IR-ITERATIVE-NEXT:    br label [[TMP13]]1303; IR-ITERATIVE:       13:1304; IR-ITERATIVE-NEXT:    ret void1305;1306; IR-DPP-LABEL: @global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe_strictfp(1307; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1308; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1309; IR-DPP:       2:1310; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1311; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321312; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321313; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321314; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1315; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1316; IR-DPP-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP8]], 01317; IR-DPP-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]1318; IR-DPP:       10:1319; IR-DPP-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81320; IR-DPP-NEXT:    br label [[TMP12]]1321; IR-DPP:       12:1322; IR-DPP-NEXT:    br label [[TMP13]]1323; IR-DPP:       13:1324; IR-DPP-NEXT:    ret void1325;1326  %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1327  ret void1328}1329 1330define amdgpu_ps void @global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe_strictfp(ptr addrspace(1) inreg %ptr, double %val) #1{1331; IR-ITERATIVE-LABEL: @global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe_strictfp(1332; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1333; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1334; IR-ITERATIVE:       2:1335; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1336; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321337; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321338; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321339; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1340; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1341; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1342; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]1343; IR-ITERATIVE:       10:1344; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] syncscope("agent") monotonic, align 81345; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]1346; IR-ITERATIVE:       12:1347; IR-ITERATIVE-NEXT:    br label [[TMP13]]1348; IR-ITERATIVE:       13:1349; IR-ITERATIVE-NEXT:    ret void1350; IR-ITERATIVE:       ComputeLoop:1351; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ 0x7FF8000000000000, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]1352; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]1353; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]1354; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i321355; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]1356; IR-ITERATIVE-NEXT:    [[TMP17]] = call double @llvm.experimental.constrained.maxnum.f64(double [[ACCUMULATOR]], double [[TMP16]], metadata !"fpexcept.strict") #[[ATTR7]]1357; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]1358; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -11359; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]1360; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 01361; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]1362; IR-ITERATIVE:       ComputeEnd:1363; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 01364; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]1365;1366; IR-DPP-LABEL: @global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe_strictfp(1367; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1368; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]1369; IR-DPP:       2:1370; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1371; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321372; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321373; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321374; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1375; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1376; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double 0x7FF8000000000000) #[[ATTR8]]1377; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]1378; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP9]], double [[TMP10]], metadata !"fpexcept.strict") #[[ATTR8]]1379; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]1380; IR-DPP-NEXT:    [[TMP13:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP11]], double [[TMP12]], metadata !"fpexcept.strict") #[[ATTR8]]1381; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]1382; IR-DPP-NEXT:    [[TMP15:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP13]], double [[TMP14]], metadata !"fpexcept.strict") #[[ATTR8]]1383; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]1384; IR-DPP-NEXT:    [[TMP17:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP15]], double [[TMP16]], metadata !"fpexcept.strict") #[[ATTR8]]1385; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]1386; IR-DPP-NEXT:    [[TMP19:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP17]], double [[TMP18]], metadata !"fpexcept.strict") #[[ATTR8]]1387; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double 0x7FF8000000000000, double [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]1388; IR-DPP-NEXT:    [[TMP21:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[TMP19]], double [[TMP20]], metadata !"fpexcept.strict") #[[ATTR8]]1389; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63) #[[ATTR8]]1390; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]]) #[[ATTR8]]1391; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 01392; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]1393; IR-DPP:       25:1394; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] syncscope("agent") monotonic, align 81395; IR-DPP-NEXT:    br label [[TMP27]]1396; IR-DPP:       27:1397; IR-DPP-NEXT:    br label [[TMP28]]1398; IR-DPP:       28:1399; IR-DPP-NEXT:    ret void1400;1401  %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1402  ret void1403}1404 1405define amdgpu_ps void @global_atomic_fadd_double_uni_address_uni_value_system_scope_strictfp(ptr addrspace(1) inreg %ptr, double inreg %val) #2 {1406; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_uni_value_system_scope_strictfp(1407; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1408; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]1409; IR-ITERATIVE:       2:1410; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1411; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321412; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321413; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321414; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1415; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1416; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR7]]1417; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i321418; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1419; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1420; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 01421; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]1422; IR-ITERATIVE:       14:1423; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] monotonic, align 41424; IR-ITERATIVE-NEXT:    br label [[TMP16]]1425; IR-ITERATIVE:       16:1426; IR-ITERATIVE-NEXT:    br label [[TMP17]]1427; IR-ITERATIVE:       17:1428; IR-ITERATIVE-NEXT:    ret void1429;1430; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_uni_value_system_scope_strictfp(1431; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1432; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]1433; IR-DPP:       2:1434; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1435; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321436; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321437; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321438; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1439; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1440; IR-DPP-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]]) #[[ATTR8]]1441; IR-DPP-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i321442; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.uitofp.f64.i32(i32 [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1443; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.experimental.constrained.fmul.f64(double [[VAL:%.*]], double [[TMP11]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1444; IR-DPP-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 01445; IR-DPP-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]1446; IR-DPP:       14:1447; IR-DPP-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] monotonic, align 41448; IR-DPP-NEXT:    br label [[TMP16]]1449; IR-DPP:       16:1450; IR-DPP-NEXT:    br label [[TMP17]]1451; IR-DPP:       17:1452; IR-DPP-NEXT:    ret void1453;1454  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val monotonic, align 41455  ret void1456}1457 1458define amdgpu_ps void @global_atomic_fadd_double_uni_address_div_value_system_scope_strictfp(ptr addrspace(1) inreg %ptr, double %val) #2 {1459; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_div_value_system_scope_strictfp(1460; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR7]]1461; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP13:%.*]]1462; IR-ITERATIVE:       2:1463; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1464; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321465; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321466; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321467; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR7]]1468; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR7]]1469; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR7]]1470; IR-ITERATIVE-NEXT:    br label [[COMPUTELOOP:%.*]]1471; IR-ITERATIVE:       10:1472; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP17:%.*]] monotonic, align 41473; IR-ITERATIVE-NEXT:    br label [[TMP12:%.*]]1474; IR-ITERATIVE:       12:1475; IR-ITERATIVE-NEXT:    br label [[TMP13]]1476; IR-ITERATIVE:       13:1477; IR-ITERATIVE-NEXT:    ret void1478; IR-ITERATIVE:       ComputeLoop:1479; IR-ITERATIVE-NEXT:    [[ACCUMULATOR:%.*]] = phi double [ -0.000000e+00, [[TMP2]] ], [ [[TMP17]], [[COMPUTELOOP]] ]1480; IR-ITERATIVE-NEXT:    [[ACTIVEBITS:%.*]] = phi i64 [ [[TMP9]], [[TMP2]] ], [ [[TMP20:%.*]], [[COMPUTELOOP]] ]1481; IR-ITERATIVE-NEXT:    [[TMP14:%.*]] = call i64 @llvm.cttz.i64(i64 [[ACTIVEBITS]], i1 true) #[[ATTR7]]1482; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = trunc i64 [[TMP14]] to i321483; IR-ITERATIVE-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[VAL:%.*]], i32 [[TMP15]]) #[[ATTR7]]1484; IR-ITERATIVE-NEXT:    [[TMP17]] = call double @llvm.experimental.constrained.fadd.f64(double [[ACCUMULATOR]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR7]]1485; IR-ITERATIVE-NEXT:    [[TMP18:%.*]] = shl i64 1, [[TMP14]]1486; IR-ITERATIVE-NEXT:    [[TMP19:%.*]] = xor i64 [[TMP18]], -11487; IR-ITERATIVE-NEXT:    [[TMP20]] = and i64 [[ACTIVEBITS]], [[TMP19]]1488; IR-ITERATIVE-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[TMP20]], 01489; IR-ITERATIVE-NEXT:    br i1 [[TMP21]], label [[COMPUTEEND:%.*]], label [[COMPUTELOOP]]1490; IR-ITERATIVE:       ComputeEnd:1491; IR-ITERATIVE-NEXT:    [[TMP22:%.*]] = icmp eq i32 [[TMP8]], 01492; IR-ITERATIVE-NEXT:    br i1 [[TMP22]], label [[TMP10:%.*]], label [[TMP12]]1493;1494; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_div_value_system_scope_strictfp(1495; IR-DPP-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live() #[[ATTR8]]1496; IR-DPP-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP28:%.*]]1497; IR-DPP:       2:1498; IR-DPP-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true) #[[ATTR8]]1499; IR-DPP-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i321500; IR-DPP-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 321501; IR-DPP-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i321502; IR-DPP-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0) #[[ATTR8]]1503; IR-DPP-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]]) #[[ATTR8]]1504; IR-DPP-NEXT:    [[TMP9:%.*]] = call double @llvm.amdgcn.set.inactive.f64(double [[VAL:%.*]], double -0.000000e+00) #[[ATTR8]]1505; IR-DPP-NEXT:    [[TMP10:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP9]], i32 273, i32 15, i32 15, i1 false) #[[ATTR8]]1506; IR-DPP-NEXT:    [[TMP11:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP9]], double [[TMP10]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1507; IR-DPP-NEXT:    [[TMP12:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP11]], i32 274, i32 15, i32 15, i1 false) #[[ATTR8]]1508; IR-DPP-NEXT:    [[TMP13:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP11]], double [[TMP12]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1509; IR-DPP-NEXT:    [[TMP14:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP13]], i32 276, i32 15, i32 15, i1 false) #[[ATTR8]]1510; IR-DPP-NEXT:    [[TMP15:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP13]], double [[TMP14]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1511; IR-DPP-NEXT:    [[TMP16:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP15]], i32 280, i32 15, i32 15, i1 false) #[[ATTR8]]1512; IR-DPP-NEXT:    [[TMP17:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP15]], double [[TMP16]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1513; IR-DPP-NEXT:    [[TMP18:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP17]], i32 322, i32 10, i32 15, i1 false) #[[ATTR8]]1514; IR-DPP-NEXT:    [[TMP19:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP17]], double [[TMP18]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1515; IR-DPP-NEXT:    [[TMP20:%.*]] = call double @llvm.amdgcn.update.dpp.f64(double -0.000000e+00, double [[TMP19]], i32 323, i32 12, i32 15, i1 false) #[[ATTR8]]1516; IR-DPP-NEXT:    [[TMP21:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[TMP19]], double [[TMP20]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR8]]1517; IR-DPP-NEXT:    [[TMP22:%.*]] = call double @llvm.amdgcn.readlane.f64(double [[TMP21]], i32 63) #[[ATTR8]]1518; IR-DPP-NEXT:    [[TMP23:%.*]] = call double @llvm.amdgcn.strict.wwm.f64(double [[TMP22]]) #[[ATTR8]]1519; IR-DPP-NEXT:    [[TMP24:%.*]] = icmp eq i32 [[TMP8]], 01520; IR-DPP-NEXT:    br i1 [[TMP24]], label [[TMP25:%.*]], label [[TMP27:%.*]]1521; IR-DPP:       25:1522; IR-DPP-NEXT:    [[TMP26:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP23]] monotonic, align 41523; IR-DPP-NEXT:    br label [[TMP27]]1524; IR-DPP:       27:1525; IR-DPP-NEXT:    br label [[TMP28]]1526; IR-DPP:       28:1527; IR-DPP-NEXT:    ret void1528;1529  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val monotonic, align 41530  ret void1531}1532 1533define amdgpu_ps void @global_atomic_fadd_double_div_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr, double inreg %val) #0 {1534; IR-LABEL: @global_atomic_fadd_double_div_address_uni_value_agent_scope_unsafe(1535; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 41536; IR-NEXT:    ret void1537;1538  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic, align 41539  ret void1540}1541 1542define amdgpu_ps void @global_atomic_fadd_double_div_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr, double %val) #0 {1543; IR-LABEL: @global_atomic_fadd_double_div_address_div_value_agent_scope_unsafe(1544; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 41545; IR-NEXT:    ret void1546;1547  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic, align 41548  ret void1549}1550 1551define amdgpu_ps void @global_atomic_fadd_double_div_address_uni_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) %ptr, double inreg %val) #1 {1552; IR-LABEL: @global_atomic_fadd_double_div_address_uni_value_one_as_scope_unsafe_strictfp(1553; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("one-as") monotonic, align 81554; IR-NEXT:    ret void1555;1556  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic1557  ret void1558}1559 1560define amdgpu_ps void @global_atomic_fadd_double_div_address_div_value_one_as_scope_unsafe_strictfp(ptr addrspace(1) %ptr, double %val) #1 {1561; IR-LABEL: @global_atomic_fadd_double_div_address_div_value_one_as_scope_unsafe_strictfp(1562; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("one-as") monotonic, align 81563; IR-NEXT:    ret void1564;1565  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic1566  ret void1567}1568 1569define amdgpu_ps void @global_atomic_fsub_double_div_address_uni_value_agent_scope_strictfp(ptr addrspace(1) %ptr, double inreg %val) #2 {1570; IR-LABEL: @global_atomic_fsub_double_div_address_uni_value_agent_scope_strictfp(1571; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81572; IR-NEXT:    ret void1573;1574  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1575  ret void1576}1577 1578define amdgpu_ps void @global_atomic_fsub_double_div_address_div_value_agent_scope_strictfp(ptr addrspace(1) %ptr, double %val) #2 {1579; IR-LABEL: @global_atomic_fsub_double_div_address_div_value_agent_scope_strictfp(1580; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81581; IR-NEXT:    ret void1582;1583  %result = atomicrmw fsub ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1584  ret void1585}1586 1587define amdgpu_ps void @global_atomic_fmin_double_div_address_uni_value_agent_scope(ptr addrspace(1) %ptr, double inreg %val) #0 {1588; IR-LABEL: @global_atomic_fmin_double_div_address_uni_value_agent_scope(1589; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81590; IR-NEXT:    ret void1591;1592  %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1593  ret void1594}1595 1596define amdgpu_ps void @global_atomic_fmin_double_div_address_div_value_agent_scope(ptr addrspace(1) %ptr, double %val) #0 {1597; IR-LABEL: @global_atomic_fmin_double_div_address_div_value_agent_scope(1598; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81599; IR-NEXT:    ret void1600;1601  %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1602  ret void1603}1604 1605define amdgpu_ps void @global_atomic_fmax_double_div_address_uni_value_agent_scope_unsafe_strictfp(ptr addrspace(1) %ptr, double inreg %val) #1{1606; IR-LABEL: @global_atomic_fmax_double_div_address_uni_value_agent_scope_unsafe_strictfp(1607; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81608; IR-NEXT:    ret void1609;1610  %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1611  ret void1612}1613 1614define amdgpu_ps void @global_atomic_fmax_double_div_address_div_value_agent_scope_unsafe_strictfp(ptr addrspace(1) %ptr, double %val) #1{1615; IR-LABEL: @global_atomic_fmax_double_div_address_div_value_agent_scope_unsafe_strictfp(1616; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] syncscope("agent") monotonic, align 81617; IR-NEXT:    ret void1618;1619  %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") monotonic1620  ret void1621}1622 1623define amdgpu_ps void @global_atomic_fadd_double_div_address_uni_value_system_scope_strictfp(ptr addrspace(1) %ptr, double inreg %val) #2 {1624; IR-LABEL: @global_atomic_fadd_double_div_address_uni_value_system_scope_strictfp(1625; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] monotonic, align 41626; IR-NEXT:    ret void1627;1628  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val monotonic, align 41629  ret void1630}1631 1632define amdgpu_ps void @global_atomic_fadd_double_div_address_div_value_system_scope_strictfp(ptr addrspace(1) %ptr, double %val) #2 {1633; IR-LABEL: @global_atomic_fadd_double_div_address_div_value_system_scope_strictfp(1634; IR-NEXT:    [[RESULT:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VAL:%.*]] monotonic, align 41635; IR-NEXT:    ret void1636;1637  %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val monotonic, align 41638  ret void1639}1640 1641attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" }1642attributes #1 = { strictfp "denormal-fp-math-f32"="preserve-sign,preserve-sign" }1643attributes #2 = { strictfp }1644