[llvm] Update foldFMulReassoc to respect absent fast-math flags (PR #88589)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Apr 12 15:54:38 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-transforms
Author: Andy Kaylor (andykaylor)
<details>
<summary>Changes</summary>
This change updates a few of the transformations in foldFMulReassoc to
respect absent fast-math flags in a few cases where fmul and fdiv
instructions were being folded but the code was not checking for
fast-math flags on the fdiv instruction and was transferring flags to
the folded instruction that were not present on the original fdiv
instruction.
This fixes https://github.com/llvm/llvm-project/issues/82857
---
Full diff: https://github.com/llvm/llvm-project/pull/88589.diff
2 Files Affected:
- (modified) llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp (+25-8)
- (modified) llvm/test/Transforms/InstCombine/fmul.ll (+47-6)
``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 4dc1319f1c437f..96485c40484007 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -636,26 +636,43 @@ Instruction *InstCombinerImpl::foldFMulReassoc(BinaryOperator &I) {
// expression.
if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP()) {
Constant *C1;
- if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) {
+ if (match(Op0,
+ m_AllowReassoc(m_OneUse(m_FDiv(m_Constant(C1), m_Value(X)))))) {
// (C1 / X) * C --> (C * C1) / X
Constant *CC1 =
ConstantFoldBinaryOpOperands(Instruction::FMul, C, C1, DL);
- if (CC1 && CC1->isNormalFP())
- return BinaryOperator::CreateFDivFMF(CC1, X, &I);
+ if (CC1 && CC1->isNormalFP()) {
+ // Preserve only fast-math flags that were set on both of the original
+ // instructions
+ auto *NewDiv = BinaryOperator::CreateFDivFMF(CC1, X, &I);
+ NewDiv->andIRFlags(Op0);
+ return NewDiv;
+ }
}
- if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
+ if (match(Op0, m_AllowReassoc(m_FDiv(m_Value(X), m_Constant(C1))))) {
+ // FIXME: This seems like it should also be checking for arcp
// (X / C1) * C --> X * (C / C1)
Constant *CDivC1 =
ConstantFoldBinaryOpOperands(Instruction::FDiv, C, C1, DL);
- if (CDivC1 && CDivC1->isNormalFP())
- return BinaryOperator::CreateFMulFMF(X, CDivC1, &I);
+ if (CDivC1 && CDivC1->isNormalFP()) {
+ // Preserve only fast-math flags that were set on both of the original
+ // instructions
+ auto *NewMul = BinaryOperator::CreateFMulFMF(X, CDivC1, &I);
+ NewMul->andIRFlags(Op0);
+ return NewMul;
+ }
// If the constant was a denormal, try reassociating differently.
// (X / C1) * C --> X / (C1 / C)
Constant *C1DivC =
ConstantFoldBinaryOpOperands(Instruction::FDiv, C1, C, DL);
- if (C1DivC && Op0->hasOneUse() && C1DivC->isNormalFP())
- return BinaryOperator::CreateFDivFMF(X, C1DivC, &I);
+ if (C1DivC && Op0->hasOneUse() && C1DivC->isNormalFP()) {
+ // Preserve only fast-math flags that were set on both of the original
+ // instructions
+ auto *NewDiv = BinaryOperator::CreateFDivFMF(X, C1DivC, &I);
+ NewDiv->andIRFlags(Op0);
+ return NewDiv;
+ }
}
// We do not need to match 'fadd C, X' and 'fsub X, C' because they are
diff --git a/llvm/test/Transforms/InstCombine/fmul.ll b/llvm/test/Transforms/InstCombine/fmul.ll
index f6435f0032891e..b541f455939c7e 100644
--- a/llvm/test/Transforms/InstCombine/fmul.ll
+++ b/llvm/test/Transforms/InstCombine/fmul.ll
@@ -652,12 +652,49 @@ define float @fdiv_constant_numerator_fmul(float %x) {
; CHECK-LABEL: @fdiv_constant_numerator_fmul(
; CHECK-NEXT: [[T3:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
; CHECK-NEXT: ret float [[T3]]
+;
+ %t1 = fdiv reassoc float 2.0e+3, %x
+ %t3 = fmul reassoc float %t1, 6.0e+3
+ ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with mixed fast-math flags
+
+define float @fdiv_constant_numerator_fmul_mixed(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_mixed(
+; CHECK-NEXT: [[T3:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT: ret float [[T3]]
+;
+ %t1 = fdiv reassoc float 2.0e+3, %x
+ %t3 = fmul fast float %t1, 6.0e+3
+ ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with full fast-math flags
+
+define float @fdiv_constant_numerator_fmul_fast(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_fast(
+; CHECK-NEXT: [[T3:%.*]] = fdiv fast float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT: ret float [[T3]]
+;
+ %t1 = fdiv fast float 2.0e+3, %x
+ %t3 = fmul fast float %t1, 6.0e+3
+ ret float %t3
+}
+
+; C1/X * C2 => (C1*C2) / X with no fast-math flags on the fdiv
+
+define float @fdiv_constant_numerator_fmul_precdiv(float %x) {
+; CHECK-LABEL: @fdiv_constant_numerator_fmul_precdiv(
+; CHECK-NEXT: [[T4:%.*]] = fdiv reassoc float 1.200000e+07, [[X:%.*]]
+; CHECK-NEXT: ret float [[T4]]
;
%t1 = fdiv float 2.0e+3, %x
%t3 = fmul reassoc float %t1, 6.0e+3
ret float %t3
}
+
; C1/X * C2 => (C1*C2) / X is disabled if C1/X has multiple uses
@fmul2_external = external global float
@@ -679,7 +716,8 @@ define float @fdiv_constant_numerator_fmul_extra_use(float %x) {
define float @fdiv_constant_denominator_fmul(float %x) {
; CHECK-LABEL: @fdiv_constant_denominator_fmul(
-; CHECK-NEXT: [[T3:%.*]] = fmul reassoc float [[X:%.*]], 3.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc float [[X:%.*]], 6.000000e+03
+; CHECK-NEXT: [[T3:%.*]] = fdiv reassoc float [[TMP1]], 2.000000e+03
; CHECK-NEXT: ret float [[T3]]
;
%t1 = fdiv float %x, 2.0e+3
@@ -692,7 +730,7 @@ define <4 x float> @fdiv_constant_denominator_fmul_vec(<4 x float> %x) {
; CHECK-NEXT: [[T3:%.*]] = fmul reassoc <4 x float> [[X:%.*]], <float 3.000000e+00, float 2.000000e+00, float 1.000000e+00, float 1.000000e+00>
; CHECK-NEXT: ret <4 x float> [[T3]]
;
- %t1 = fdiv <4 x float> %x, <float 2.0e+3, float 3.0e+3, float 2.0e+3, float 1.0e+3>
+ %t1 = fdiv reassoc <4 x float> %x, <float 2.0e+3, float 3.0e+3, float 2.0e+3, float 1.0e+3>
%t3 = fmul reassoc <4 x float> %t1, <float 6.0e+3, float 6.0e+3, float 2.0e+3, float 1.0e+3>
ret <4 x float> %t3
}
@@ -705,7 +743,7 @@ define <4 x float> @fdiv_constant_denominator_fmul_vec_constexpr(<4 x float> %x)
; CHECK-NEXT: ret <4 x float> [[T3]]
;
%constExprMul = bitcast i128 trunc (i160 bitcast (<5 x float> <float 6.0e+3, float 6.0e+3, float 2.0e+3, float 1.0e+3, float undef> to i160) to i128) to <4 x float>
- %t1 = fdiv <4 x float> %x, <float 2.0e+3, float 3.0e+3, float 2.0e+3, float 1.0e+3>
+ %t1 = fdiv reassoc <4 x float> %x, <float 2.0e+3, float 3.0e+3, float 2.0e+3, float 1.0e+3>
%t3 = fmul reassoc <4 x float> %t1, %constExprMul
ret <4 x float> %t3
}
@@ -745,7 +783,8 @@ define float @fdiv_constant_denominator_fmul_denorm(float %x) {
define float @fdiv_constant_denominator_fmul_denorm_try_harder(float %x) {
; CHECK-LABEL: @fdiv_constant_denominator_fmul_denorm_try_harder(
-; CHECK-NEXT: [[T3:%.*]] = fdiv reassoc float [[X:%.*]], 0x47E8000000000000
+; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc float [[X:%.*]], 0x3810000000000000
+; CHECK-NEXT: [[T3:%.*]] = fdiv reassoc float [[TMP1]], 3.000000e+00
; CHECK-NEXT: ret float [[T3]]
;
%t1 = fdiv float %x, 3.0
@@ -868,7 +907,8 @@ define float @fmul_fadd_distribute_extra_use(float %x) {
define double @fmul_fadd_fdiv_distribute2(double %x) {
; CHECK-LABEL: @fmul_fadd_fdiv_distribute2(
-; CHECK-NEXT: [[TMP1:%.*]] = fdiv reassoc double [[X:%.*]], 0x7FE8000000000000
+; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc double [[X:%.*]], 0x10000000000000
+; CHECK-NEXT: [[TMP1:%.*]] = fdiv reassoc double [[TMP2]], 3.000000e+00
; CHECK-NEXT: [[T3:%.*]] = fadd reassoc double [[TMP1]], 0x34000000000000
; CHECK-NEXT: ret double [[T3]]
;
@@ -883,7 +923,8 @@ define double @fmul_fadd_fdiv_distribute2(double %x) {
define double @fmul_fadd_fdiv_distribute3(double %x) {
; CHECK-LABEL: @fmul_fadd_fdiv_distribute3(
-; CHECK-NEXT: [[TMP1:%.*]] = fdiv reassoc double [[X:%.*]], 0x7FE8000000000000
+; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc double [[X:%.*]], 0x10000000000000
+; CHECK-NEXT: [[TMP1:%.*]] = fdiv reassoc double [[TMP2]], 3.000000e+00
; CHECK-NEXT: [[T3:%.*]] = fadd reassoc double [[TMP1]], 0x34000000000000
; CHECK-NEXT: ret double [[T3]]
;
``````````
</details>
https://github.com/llvm/llvm-project/pull/88589
More information about the llvm-commits
mailing list