RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-22 18:54:02 +01:00
Code Issues Projects Releases Wiki Activity

[AMDGPU] Fix regression in AMDGPULibCalls allowing native for doubles

Browse Source 
Under -cl-fast-relaxed-math we could use native_sqrt, but f64 was
allowed to produce HSAIL's nsqrt instruction. HSAIL is not here
and we stick with non-existing native_sqrt(double) as a result.

Add check for f64 to not return native functions and also remove
handling of f64 case for fold_sqrt.

Differential Revision: https://reviews.llvm.org/D37223

llvm-svn: 311900
This commit is contained in:
Stanislav Mekhanoshin 2017-08-28 18:00:08 +00:00
parent 5f48b3a89c
commit 22de6c878a
2 changed files with 14 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
lib/Target/AMDGPU/AMDGPULibCalls.cpp
View File

@ -1170,6 +1170,8 @@ bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
// Get a scalar native builtin signle argument FP function
Constant* AMDGPULibCalls::getNativeFunction(Module* M, const FuncInfo& FInfo) {
if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
return nullptr;
FuncInfo nf = FInfo;
nf.setPrefix(AMDGPULibFunc::NATIVE);
return getFunction(M, nf);
@ -1178,9 +1180,7 @@ Constant* AMDGPULibCalls::getNativeFunction(Module* M, const FuncInfo& FInfo) {
// fold sqrt -> native_sqrt (x)
bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
const FuncInfo &FInfo) {
if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
getArgType(FInfo) == AMDGPULibFunc::F64) &&
(getVecSize(FInfo) == 1) &&
if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
(FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
if (Constant *FPExpr = getNativeFunction(
CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {

11
test/CodeGen/AMDGPU/simplify-libcalls.ll
View File

@ -641,7 +641,18 @@ entry:
ret void
}
; GCN-LABEL: {{^}}define amdgpu_kernel void @test_dont_use_native_sqrt_fast_f64
; GCN: tail call fast double @_Z4sqrtd(double %tmp)
define amdgpu_kernel void @test_dont_use_native_sqrt_fast_f64(double addrspace(1)* nocapture %a) {
entry:
%tmp = load double, double addrspace(1)* %a, align 8
%call = tail call fast double @_Z4sqrtd(double %tmp)
store double %call, double addrspace(1)* %a, align 8
ret void
}
declare float @_Z4sqrtf(float)
declare double @_Z4sqrtd(double)
; GCN-LABEL: {{^}}define amdgpu_kernel void @test_use_native_rsqrt
; GCN-NATIVE: tail call fast float @_Z12native_rsqrtf(float %tmp)