| ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 4 |
| ; RUN: llvm-as < %s | llvm-dis | FileCheck %s |
| |
| ; amdgpu-unsafe-fp-atomics attribute should be removed and replaced |
| ; with metadata attached to any atomicrmw with floating-point |
| ; operations. |
| |
| ; Maybe the attribute should be dropped from declarations, but it |
| ; didn't do anything on one and clang never added it. |
| declare void @unsafe_fp_atomics_true_decl() "amdgpu-unsafe-fp-atomics"="true" |
| declare void @unsafe_fp_atomics_false_decl() "amdgpu-unsafe-fp-atomics"="false" |
| |
| ; Delete the attribute and replace with the most aggressive metadata possible |
| define void @unsafe_fp_atomics_true(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="true" { |
| ; CHECK-LABEL: define void @unsafe_fp_atomics_true( |
| ; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) { |
| ; 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]] |
| ; 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]] |
| ; 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]] |
| ; 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]] |
| ; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 4 |
| ; 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]] |
| ; 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]] |
| ; 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]] |
| ; 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]] |
| ; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: ret void |
| ; |
| %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst |
| %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| |
| ; xchg doesn't need any metadata |
| %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| ret void |
| } |
| |
| ; Should just delete the effectless attribute if it exists |
| define void @unsafe_fp_atomics_false(ptr addrspace(1) %ptr, float %val, i32 %ival, <2 x half> %vval) "amdgpu-unsafe-fp-atomics"="false" { |
| ; CHECK-LABEL: define void @unsafe_fp_atomics_false( |
| ; CHECK-SAME: ptr addrspace(1) [[PTR:%.*]], float [[VAL:%.*]], i32 [[IVAL:%.*]], <2 x half> [[VVAL:%.*]]) { |
| ; CHECK-NEXT: [[RMW_FADD:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FSUB:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FMIN:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FMAX:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[PTR]], i32 [[IVAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FADD_VECTOR:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FSUB_VECTOR:%.*]] = atomicrmw fsub ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FMIN_VECTOR:%.*]] = atomicrmw fmin ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_FMAX_VECTOR:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR]], <2 x half> [[VVAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: [[RMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[PTR]], float [[VAL]] syncscope("one-as") seq_cst, align 4 |
| ; CHECK-NEXT: ret void |
| ; |
| %rmw.fadd = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fsub = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fmin = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.fmax = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| %rmw.xor = atomicrmw xor ptr addrspace(1) %ptr, i32 %ival syncscope("one-as") seq_cst |
| %rmw.fadd.vector = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fsub.vector = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fmin.vector = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| %rmw.fmax.vector = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %vval syncscope("one-as") seq_cst |
| |
| ; xchg doesn't need any metadata |
| %rmw.xchg = atomicrmw xchg ptr addrspace(1) %ptr, float %val syncscope("one-as") seq_cst |
| ret void |
| } |
| |
| ;. |
| ; CHECK: attributes #[[ATTR0:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="true" } |
| ; CHECK: attributes #[[ATTR1:[0-9]+]] = { "amdgpu-unsafe-fp-atomics"="false" } |
| ;. |
| ; CHECK: [[META0]] = !{} |
| ;. |