81 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 42; RUN: llvm-as < %s | llvm-dis | FileCheck %s3 4; amdgpu-unsafe-fp-atomics attribute should be removed and replaced5; with metadata attached to any atomicrmw with floating-point6; operations.7 8; Maybe the attribute should be dropped from declarations, but it9; didn't do anything on one and clang never added it.10declare void @unsafe_fp_atomics_true_decl() "amdgpu-unsafe-fp-atomics"="true"11declare void @unsafe_fp_atomics_false_decl() "amdgpu-unsafe-fp-atomics"="false"12 13; Delete the attribute and replace with the most aggressive metadata possible14define void @unsafe_fp_atomics_true(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="true" {15; CHECK-LABEL: define void @unsafe_fp_atomics_true(16; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) {17; CHECK-NEXT: [[RMW_FADD:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0:![0-9]+]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]18; CHECK-NEXT: [[RMW_FSUB:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]19; CHECK-NEXT: [[RMW_FMIN:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]20; CHECK-NEXT: [[RMW_FMAX:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]21; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 422; CHECK-NEXT: [[RMW_FADD_VECTOR:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]23; CHECK-NEXT: [[RMW_FSUB_VECTOR:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]24; CHECK-NEXT: [[RMW_FMIN_VECTOR:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]25; CHECK-NEXT: [[RMW_FMAX_VECTOR:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4, !amdgpu.no.fine.grained.host.memory [[META0]], !amdgpu.no.remote.memory.access [[META0]], !amdgpu.ignore.denormal.mode [[META0]]26; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 427; CHECK-NEXT: ret void28;29 %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst30 %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst31 %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst32 %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst33 %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst34 %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst35 %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst36 %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst37 %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst38 39 ; xchg doesn't need any metadata40 %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst41 ret void42}43 44; Should just delete the effectless attribute if it exists45define void @unsafe_fp_atomics_false(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="false" {46; CHECK-LABEL: define void @unsafe_fp_atomics_false(47; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) {48; CHECK-NEXT: [[RMW_FADD:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 449; CHECK-NEXT: [[RMW_FSUB:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 450; CHECK-NEXT: [[RMW_FMIN:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 451; CHECK-NEXT: [[RMW_FMAX:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 452; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 453; CHECK-NEXT: [[RMW_FADD_VECTOR:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 454; CHECK-NEXT: [[RMW_FSUB_VECTOR:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 455; CHECK-NEXT: [[RMW_FMIN_VECTOR:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 456; CHECK-NEXT: [[RMW_FMAX_VECTOR:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 457; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 458; CHECK-NEXT: ret void59;60 %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst61 %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst62 %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst63 %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst64 %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst65 %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst66 %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst67 %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst68 %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst69 70 ; xchg doesn't need any metadata71 %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst72 ret void73}74 75;.76; CHECK: attributes #[[ATTR0:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="true" }77; CHECK: attributes #[[ATTR1:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="false" }78;.79; CHECK: [[META0]] = !{}80;.81