[llvm] r326098 - [InstCombine] allow fdiv folds with less than fully 'fast' ops

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Mon Feb 26 08:02:45 2018
New Revision: 326098

URL: http://llvm.org/viewvc/llvm-project?rev=326098&view=rev
Log:
[InstCombine] allow fdiv folds with less than fully 'fast' ops

Note: gcc appears to allow this fold with -freciprocal-math alone, 
but clang/llvm require more than that with this patch. The wording
in the definitions seems fuzzy enough that it could go either way,
but we'll err on the conservative side of FMF interpretation.

This patch also changes the newly created fmul to have FMF propagated
by the last fdiv rather than intersecting the FMF of the fdivs. This
matches the behavior of other folds near here. The new fmul is only 
used to produce an intermediate op for the final fdiv result, so it
shouldn't be any stricter than that result. The previous behavior
could result in dropping FMF via other folds in instcombine or CSE.

Differential Revision: https://reviews.llvm.org/D43398

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
    llvm/trunk/test/Transforms/InstCombine/fdiv.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp?rev=326098&r1=326097&r2=326098&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp Mon Feb 26 08:02:45 2018
@@ -1347,28 +1347,18 @@ Instruction *InstCombiner::visitFDiv(Bin
       if (Instruction *R = FoldOpIntoSelect(I, SI))
         return R;
 
-  if (I.isFast()) {
+  if (I.hasAllowReassoc() && I.hasAllowReciprocal()) {
     Value *X, *Y;
     if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
         (!isa<Constant>(Y) || !isa<Constant>(Op1))) {
       // (X / Y) / Z => X / (Y * Z)
-      Value *YZ = Builder.CreateFMul(Y, Op1);
-      if (auto *YZInst = dyn_cast<Instruction>(YZ)) {
-        FastMathFlags FMFIntersect = I.getFastMathFlags();
-        FMFIntersect &= cast<Instruction>(Op0)->getFastMathFlags();
-        YZInst->setFastMathFlags(FMFIntersect);
-      }
+      Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I);
       return BinaryOperator::CreateFDivFMF(X, YZ, &I);
     }
     if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
         (!isa<Constant>(Y) || !isa<Constant>(Op0))) {
       // Z / (X / Y) => (Y * Z) / X
-      Value *YZ = Builder.CreateFMul(Y, Op0);
-      if (auto *YZInst = dyn_cast<Instruction>(YZ)) {
-        FastMathFlags FMFIntersect = I.getFastMathFlags();
-        FMFIntersect &= cast<Instruction>(Op1)->getFastMathFlags();
-        YZInst->setFastMathFlags(FMFIntersect);
-      }
+      Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I);
       return BinaryOperator::CreateFDivFMF(YZ, X, &I);
     }
   }

Modified: llvm/trunk/test/Transforms/InstCombine/fdiv.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/fdiv.ll?rev=326098&r1=326097&r2=326098&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/fdiv.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/fdiv.ll Mon Feb 26 08:02:45 2018
@@ -126,12 +126,12 @@ define <2 x float> @not_exact_inverse_ve
 
 define float @div_with_div_numerator(float %x, float %y, float %z) {
 ; CHECK-LABEL: @div_with_div_numerator(
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul ninf float [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[DIV2:%.*]] = fdiv fast float [[X:%.*]], [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc arcp float [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[DIV2:%.*]] = fdiv reassoc arcp float [[X:%.*]], [[TMP1]]
 ; CHECK-NEXT:    ret float [[DIV2]]
 ;
   %div1 = fdiv ninf float %x, %y
-  %div2 = fdiv fast float %div1, %z
+  %div2 = fdiv arcp reassoc float %div1, %z
   ret float %div2
 }
 
@@ -139,12 +139,12 @@ define float @div_with_div_numerator(flo
 
 define <2 x float> @div_with_div_denominator(<2 x float> %x, <2 x float> %y, <2 x float> %z) {
 ; CHECK-LABEL: @div_with_div_denominator(
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul nnan <2 x float> [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[DIV2:%.*]] = fdiv fast <2 x float> [[TMP1]], [[X:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[DIV2:%.*]] = fdiv reassoc arcp <2 x float> [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret <2 x float> [[DIV2]]
 ;
   %div1 = fdiv nnan <2 x float> %x, %y
-  %div2 = fdiv fast <2 x float> %z, %div1
+  %div2 = fdiv arcp reassoc <2 x float> %z, %div1
   ret <2 x float> %div2
 }
 
@@ -342,8 +342,8 @@ define <2 x float> @div_constant_dividen
 
 define <2 x float> @div_constant_dividend3(<2 x float> %x) {
 ; CHECK-LABEL: @div_constant_dividend3(
-; CHECK-NEXT:    [[T1:%.*]] = fdiv <2 x float> <float 3.000000e+00, float 7.000000e+00>, [[X:%.*]]
-; CHECK-NEXT:    [[T2:%.*]] = fdiv reassoc arcp <2 x float> <float 1.500000e+01, float -7.000000e+00>, [[T1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc arcp <2 x float> [[X:%.*]], <float 1.500000e+01, float -7.000000e+00>
+; CHECK-NEXT:    [[T2:%.*]] = fmul reassoc arcp <2 x float> [[TMP1]], <float 0x3FD5555560000000, float 0x3FC24924A0000000>
 ; CHECK-NEXT:    ret <2 x float> [[T2]]
 ;
   %t1 = fdiv <2 x float> <float 3.0e0, float 7.0e0>, %x