Disable call to fma for soft-float

[long5hot]

Summary

PowerPC backend generate calls to libc function calls
for soft-float, regardless of the -nostdlib /-ffreestanding flag.
fma is not a function provided by compiler-rt builtins and
thus should not be generated here.
PR : #55230

Below is patch given by @nemanjai

[long5hot]

This comment was removed by long5hot.

[long5hot]

Apologies for updating diff too many times, used arcanist first time. Won't happend again!

[shchenz]

compiler-rt builtins library should not just undefining fma for PPC, right (see https://gcc.gnu.org/onlinedocs/gccint/Soft-float-library-routines.html)? If so, maybe we should at least first try not generating fma for soft-float at the first place where fmul + fadd is fused, like tryEmitFMulAdd in clang front end?

[long5hot]

In D156344#4583559, @shchenz wrote:

compiler-rt builtins library should not just undefining fma for PPC, right (see https://gcc.gnu.org/onlinedocs/gccint/Soft-float-library-routines.html)? If so, maybe we should at least first try not generating fma for soft-float at the first place where fmul + fadd is fused, like tryEmitFMulAdd in clang front end?

I actually raised D154605 for that one, but it's only for X86.
I was relying on target-features containing "+fma", which is wrong. The review was raised just to get feedback from community working on various targets.
Plan is to disable fma-intrinsic from IR for all targets which doesn't support fma.

[shchenz]

In D156344#4583767, @long5hot wrote:

In D156344#4583559, @shchenz wrote:

compiler-rt builtins library should not just undefining fma for PPC, right (see https://gcc.gnu.org/onlinedocs/gccint/Soft-float-library-routines.html)? If so, maybe we should at least first try not generating fma for soft-float at the first place where fmul + fadd is fused, like tryEmitFMulAdd in clang front end?

I actually raised D154605 for that one, but it's only for X86.
I was relying on target-features containing "+fma", which is wrong. The review was raised just to get feedback from community working on various targets.
Plan is to disable fma-intrinsic from IR for all targets which doesn't support fma.

Thanks. I think soft-float looks like a candidate for the hasFMA function in D154605.

[long5hot]

@nemanjai , @jhibbits should i land on this commit, because as @shchenz said, similiar to D158632 review we are having plan to do same for PPC.

[long5hot]

ping!

[long5hot]

landing on this PR, from the conclusion of #55230 and D158632 conversation..
"If the llvm.fmuladd intrinsic reaches the backend, the backend is responsible for deciding how to represent it."

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
                                                    EVT VT) const {
   return isFMAFasterThanFMulAndFAdd(
       MF.getFunction(), VT.getTypeForEVT(MF.getFunction().getContext()));
 }
 
 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F,
                                                    Type *Ty) const {
-  if (Subtarget.hasSPE())
+  if (Subtarget.hasSPE() || Subtarget.useSoftFloat())
     return false;
   switch (Ty->getScalarType()->getTypeID()) {
   case Type::FloatTyID:
   case Type::DoubleTyID:
     return true;
   case Type::FP128TyID:
     return Subtarget.hasP9Vector();
   default:

llvm/test/CodeGen/PowerPC/ppcsoftops.ll

entry:
   %0 = load double, ptr %a, align 8
   %1 = load double, ptr %b, align 8
   %div = fdiv double %0, %1
   ret double %div
 
   ; CHECK-LABEL:      __divdf3
 }
 
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local zeroext i32 @func(double noundef %0, double noundef %1) #0 {
+  %3 = alloca double, align 8
+  %4 = alloca double, align 8
+  store double %0, ptr %3, align 8
+  store double %1, ptr %4, align 8
+  %5 = load double, ptr %3, align 8
+  %6 = load double, ptr %4, align 8
+  %7 = fneg double %6
+  %8 = call double @llvm.fmuladd.f64(double %7, double 0x41F0000000000000, double %5)
+  %9 = fptoui double %8 to i32
+  ret i32 %9
+
+  ; CHECK-LABEL:      __muldf3
+  ; CHECK-LABEL:      __adddf3
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare double @llvm.fmuladd.f64(double, double, double) #1
+
 attributes #0 = {"use-soft-float"="true" }
+attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }