[InstCombine] allow fmul fold with less than 'fast'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@326502 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 2959c09..1529911 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -676,7 +676,7 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
   //  latency of the instruction Y is amortized by the expression of X*X,
   //  and therefore Y is in a "less critical" position compared to what it
   //  was before the transformation.
-  if (I.isFast()) {
+  if (I.hasAllowReassoc()) {
     if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) &&
         Op1 != Y) {
       Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I);

diff --git a/test/Transforms/InstCombine/fmul.ll b/test/Transforms/InstCombine/fmul.ll
index 4c364f9..6c0642b 100644 (file)
--- a/test/Transforms/InstCombine/fmul.ll
+++ b/test/Transforms/InstCombine/fmul.ll
@@ -228,15 +228,17 @@ define float @fabs_x_fabs(float %x, float %y) {
 }
 
 ; (X*Y) * X => (X*X) * Y
+; The transform only requires 'reassoc', but test other FMF in 
+; the commuted variants to make sure FMF propagates as expected.
 
 define float @reassoc_common_operand1(float %x, float %y) {
 ; CHECK-LABEL: @reassoc_common_operand1(
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul fast float [[X:%.*]], [[X]]
-; CHECK-NEXT:    [[MUL2:%.*]] = fmul fast float [[TMP1]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc float [[X:%.*]], [[X]]
+; CHECK-NEXT:    [[MUL2:%.*]] = fmul reassoc float [[TMP1]], [[Y:%.*]]
 ; CHECK-NEXT:    ret float [[MUL2]]
 ;
   %mul1 = fmul float %x, %y
-  %mul2 = fmul fast float %mul1, %x
+  %mul2 = fmul reassoc float %mul1, %x
   ret float %mul2
 }
 
@@ -258,13 +260,13 @@ define float @reassoc_common_operand2(float %x, float %y) {
 define float @reassoc_common_operand3(float %x1, float %y) {
 ; CHECK-LABEL: @reassoc_common_operand3(
 ; CHECK-NEXT:    [[X:%.*]] = fdiv float [[X1:%.*]], 3.000000e+00
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul fast float [[X]], [[X]]
-; CHECK-NEXT:    [[MUL2:%.*]] = fmul fast float [[TMP1]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc nnan float [[X]], [[X]]
+; CHECK-NEXT:    [[MUL2:%.*]] = fmul reassoc nnan float [[TMP1]], [[Y:%.*]]
 ; CHECK-NEXT:    ret float [[MUL2]]
 ;
   %x = fdiv float %x1, 3.0 ; thwart complexity-based canonicalization
   %mul1 = fmul float %x, %y
-  %mul2 = fmul fast float %x, %mul1
+  %mul2 = fmul reassoc nnan float %x, %mul1
   ret float %mul2
 }
 
@@ -273,13 +275,13 @@ define float @reassoc_common_operand3(float %x1, float %y) {
 define float @reassoc_common_operand4(float %x1, float %y) {
 ; CHECK-LABEL: @reassoc_common_operand4(
 ; CHECK-NEXT:    [[X:%.*]] = fdiv float [[X1:%.*]], 3.000000e+00
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul fast float [[X]], [[X]]
-; CHECK-NEXT:    [[MUL2:%.*]] = fmul fast float [[TMP1]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fmul reassoc ninf float [[X]], [[X]]
+; CHECK-NEXT:    [[MUL2:%.*]] = fmul reassoc ninf float [[TMP1]], [[Y:%.*]]
 ; CHECK-NEXT:    ret float [[MUL2]]
 ;
   %x = fdiv float %x1, 3.0 ; thwart complexity-based canonicalization
   %mul1 = fmul float %y, %x
-  %mul2 = fmul fast float %x, %mul1
+  %mul2 = fmul reassoc ninf float %x, %mul1
   ret float %mul2
 }
 
@@ -288,12 +290,12 @@ define float @reassoc_common_operand4(float %x1, float %y) {
 define float @reassoc_common_operand_multi_use(float %x, float %y) {
 ; CHECK-LABEL: @reassoc_common_operand_multi_use(
 ; CHECK-NEXT:    [[MUL1:%.*]] = fmul float [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[MUL2:%.*]] = fmul fast float [[MUL1]], [[X]]
+; CHECK-NEXT:    [[MUL2:%.*]] = fmul reassoc float [[MUL1]], [[X]]
 ; CHECK-NEXT:    call void @use_f32(float [[MUL1]])
 ; CHECK-NEXT:    ret float [[MUL2]]
 ;
   %mul1 = fmul float %x, %y
-  %mul2 = fmul fast float %mul1, %x
+  %mul2 = fmul reassoc float %mul1, %x
   call void @use_f32(float %mul1)
   ret float %mul2
 }