AMDGPU: Replace sqrt OpenCL libcalls with llvm.sqrt (#74197)

The library implementation is just a wrapper around a call to the
intrinsic, but loses metadata. Swap out the call site to the intrinsic
so that the lowering can see the !fpmath metadata and fast math flags.

Since d56e0d0, clang started placing
!fpmath on OpenCL library sqrt calls. Also don't bother emitting
native_sqrt anymore, it's just another wrapper around llvm.sqrt.

llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp

@@ -87,9 +87,6 @@ class AMDGPULibCalls {
                               Constant *copr0, Constant *copr1);
   bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo);
 
-  // sqrt
-  bool fold_sqrt(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
-
   /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value
   /// of cos, sincos call).
   std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg,
@@ -672,8 +669,6 @@ bool AMDGPULibCalls::fold(CallInst *CI) {
 
     // Specialized optimizations for each function call.
     //
-    // TODO: Handle other simple intrinsic wrappers. Sqrt.
-    //
     // TODO: Handle native functions
     switch (FInfo.getId()) {
     case AMDGPULibFunc::EI_EXP:
@@ -794,7 +789,9 @@ bool AMDGPULibCalls::fold(CallInst *CI) {
     case AMDGPULibFunc::EI_ROOTN:
       return fold_rootn(FPOp, B, FInfo);
     case AMDGPULibFunc::EI_SQRT:
-      return fold_sqrt(FPOp, B, FInfo);
+      // TODO: Allow with strictfp + constrained intrinsic
+      return tryReplaceLibcallWithSimpleIntrinsic(
+          B, CI, Intrinsic::sqrt, true, true, /*AllowStrictFP=*/false);
     case AMDGPULibFunc::EI_COS:
     case AMDGPULibFunc::EI_SIN:
       return fold_sincos(FPOp, B, FInfo);
@@ -1273,29 +1270,6 @@ bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic(
   return true;
 }
 
-// fold sqrt -> native_sqrt (x)
-bool AMDGPULibCalls::fold_sqrt(FPMathOperator *FPOp, IRBuilder<> &B,
-                               const FuncInfo &FInfo) {
-  if (!isUnsafeMath(FPOp))
-    return false;
-
-  if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
-      (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
-    Module *M = B.GetInsertBlock()->getModule();
-
-    if (FunctionCallee FPExpr = getNativeFunction(
-            M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
-      Value *opr0 = FPOp->getOperand(0);
-      LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
-                        << "sqrt(" << *opr0 << ")\n");
-      Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
-      replaceCall(FPOp, nval);
-      return true;
-    }
-  }
-  return false;
-}
-
 std::tuple<Value *, Value *, Value *>
 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B,
                              FunctionCallee Fsincos) {