Do not translate rint into nearbyint, but truncate it like nearbyint

[joerg]

Summary

A common way to implement nearbyint is by fiddling with the floating point environment and calling rint. This is used at least by the BSD libm and musl. As such, canonicalizing the latter to the former will create infinite loops for libm and generally pessimize performance, at least when the generic C versions are used. This patch preserves the rint in the libcall translation and also handle the domain truncation logic, so that rint with float argument will be reduced to rintf etc.

[scanon]

LGTM.

[arsenm]

LGTM

[efriedma]

If rint() is faster than nearbyint(), should we canonicalize nearbyint()->rint()?

[scanon]

In D31533#715623, @efriedma wrote:

If rint() is faster than nearbyint(), should we canonicalize nearbyint()->rint()?

nearbyint( ) is not allowed to raise inexact. rint( ) is. Replacing nearbyint( ) with rint( ) requires that we support the FENV_ACCESS pragma (and only do the transformation when FENV_ACCESS is off).

Also, this is relatively low-value, because on newer platforms there's no difference; e.g. SSE4.1 and armv8 have single instruction implementations for both.

[joerg]

r299247.

lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::fabs: {
     }
 
     LLVM_FALLTHROUGH;
   }
   case Intrinsic::ceil:
   case Intrinsic::floor:
   case Intrinsic::round:
   case Intrinsic::nearbyint:
+  case Intrinsic::rint:
   case Intrinsic::trunc: {
     Value *ExtSrc;
     if (match(II->getArgOperand(0), m_FPExt(m_Value(ExtSrc))) &&
         II->getArgOperand(0)->hasOneUse()) {
       // fabs (fpext x) -> fpext (fabs x)
       Value *F = Intrinsic::getDeclaration(II->getModule(), II->getIntrinsicID(),
                                            { ExtSrc->getType() });
       CallInst *NewFabs = Builder->CreateCall(F, ExtSrc);

lib/Transforms/Utils/SimplifyLibCalls.cpp

case LibFunc_fputc:
       return optimizeErrorReporting(CI, Builder, 1);
     case LibFunc_ceil:
       return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
     case LibFunc_floor:
       return replaceUnaryCall(CI, Builder, Intrinsic::floor);
     case LibFunc_round:
       return replaceUnaryCall(CI, Builder, Intrinsic::round);
     case LibFunc_nearbyint:
-    case LibFunc_rint:
       return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
+    case LibFunc_rint:
+      return replaceUnaryCall(CI, Builder, Intrinsic::rint);
     case LibFunc_trunc:
       return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
     case LibFunc_acos:
     case LibFunc_acosh:
     case LibFunc_asin:
     case LibFunc_asinh:
     case LibFunc_atan:
     case LibFunc_atanh:

test/Transforms/InstCombine/float-shrink-compare.ll

 define i32 @test5(float %x, float %y) nounwind uwtable {
   %x.ext = fpext float %x to double
   %rint = call double @rint(double %x.ext) nounwind
   %y.ext = fpext float %y to double
   %cmp = fcmp oeq double %rint, %y.ext
   %cmp.ext = zext i1 %cmp to i32
   ret i32 %cmp.ext
 ; CHECK-LABEL: @test5(
-; CHECK-NEXT: %rint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: %rint = call float @llvm.rint.f32(float %x)
 ; CHECK-NEXT: fcmp oeq float %rint, %y
 }
 
 define i32 @test6(float %x, float %y) nounwind uwtable {
   %x.ext = fpext float %x to double
   %round = call double @round(double %x.ext) nounwind readnone
   %y.ext = fpext float %y to double
   %cmp = fcmp oeq double %round, %y.ext

 define i32 @test12(float %x, float %y) nounwind uwtable {
   %x.ext = fpext float %x to double
   %y.ext = fpext float %y to double
   %rint = call double @rint(double %x.ext) nounwind
   %cmp = fcmp oeq double %y.ext, %rint
   %cmp.ext = zext i1 %cmp to i32
   ret i32 %cmp.ext
 ; CHECK-LABEL: @test12(
-; CHECK-NEXT: %rint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: %rint = call float @llvm.rint.f32(float %x)
 ; CHECK-NEXT: fcmp oeq float %rint, %y
 }
 
 define i32 @test13(float %x, float %y) nounwind uwtable {
   %x.ext = fpext float %x to double
   %y.ext = fpext float %y to double
   %round = call double @round(double %x.ext) nounwind readnone
   %cmp = fcmp oeq double %y.ext, %round