[AMDGPU] Fix regression in AMDGPULibCalls allowing native for doubles

[rampitec]

Summary

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.

[vpykhtin]

LGTM.

llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp

if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
     }
   }
 
   return false;
 }
 
 // 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);
 }
 
 // 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))) {
       Value *opr0 = CI->getArgOperand(0);
       DEBUG(errs() << "AMDIC: " << *CI << " ---> "
                    << "sqrt(" << *opr0 << ")\n");
       Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
       replaceCall(nval);

llvm/trunk/test/CodeGen/AMDGPU/simplify-libcalls.ll

 define amdgpu_kernel void @test_use_native_sqrt(float addrspace(1)* nocapture %a) {
 entry:
   %tmp = load float, float addrspace(1)* %a, align 4
   %call = tail call fast float @_Z4sqrtf(float %tmp)
   store float %call, float addrspace(1)* %a, align 4
   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)
 define amdgpu_kernel void @test_use_native_rsqrt(float addrspace(1)* nocapture %a) {
 entry:
   %tmp = load float, float addrspace(1)* %a, align 4
   %call = tail call fast float @_Z5rsqrtf(float %tmp)
   store float %call, float addrspace(1)* %a, align 4