[AMDGPU] Fix regression in AMDGPULibCalls allowing native for doubles

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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311900 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPULibCalls.cpp b/lib/Target/AMDGPU/AMDGPULibCalls.cpp
index 57cb2a6a9214e50c4bd064986ccab23118fb5462..ff89ce7b96f9220139b6e7707567514c96cc9103 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPULibCalls.cpp
+++ b/lib/Target/AMDGPU/AMDGPULibCalls.cpp
@@ -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))) {

diff --git a/test/CodeGen/AMDGPU/simplify-libcalls.ll b/test/CodeGen/AMDGPU/simplify-libcalls.ll
index e487f3c417eae781c09bf6477ae7b1a2293b18ab..eab1fe4c70f79f54eb0dc9a79a88a4c81695ebda 100644 (file)
--- a/test/CodeGen/AMDGPU/simplify-libcalls.ll
+++ b/test/CodeGen/AMDGPU/simplify-libcalls.ll
@@ -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)