IR: Stop requiring nsz to reassociate fmul #171726
Merged
+10
−18
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Author
arsenm
added
floating-point
Graphite App
arsenm
requested review from
andykaylor,
dtcxzyw,
jayfoad and
jcranmer-intel
Member
|
@llvm/pr-subscribers-llvm-ir @llvm/pr-subscribers-llvm-transforms Author: Matt Arsenault (arsenm) Changesnsz can only change the behavior of the sign bit. Fixes #64967 Full diff: https://github.com/llvm/llvm-project/pull/171726.diff 4 Files Affected:
diff --git a/llvm/lib/IR/Instruction.cpp b/llvm/lib/IR/Instruction.cpp
index 33ca46ca1c2c6..b95c1466871bc 100644
--- a/llvm/lib/IR/Instruction.cpp
+++ b/llvm/lib/IR/Instruction.cpp
@@ -1271,6 +1271,7 @@ bool Instruction::isAssociative() const {
switch (Opcode) {
case FMul:
+ return cast<FPMathOperator>(this)->hasAllowReassoc();
case FAdd:
return cast<FPMathOperator>(this)->hasAllowReassoc() &&
cast<FPMathOperator>(this)->hasNoSignedZeros();
diff --git a/llvm/test/Transforms/InstCombine/2006-10-26-VectorReassoc.ll b/llvm/test/Transforms/InstCombine/2006-10-26-VectorReassoc.ll
index fb860a5e7bdf3..6509797e0d3dc 100644
--- a/llvm/test/Transforms/InstCombine/2006-10-26-VectorReassoc.ll
+++ b/llvm/test/Transforms/InstCombine/2006-10-26-VectorReassoc.ll
@@ -35,12 +35,10 @@ define <4 x float> @test_fmul_reassoc_nsz(<4 x float> %V) {
}
; (V * C1) * C2 => V * (C1 * C2)
-; TODO: This doesn't require 'nsz'. It should fold to V * { 1.0, 4.0e+05, -9.0, 16.0 }
define <4 x float> @test_fmul_reassoc(<4 x float> %V) {
; CHECK-LABEL: @test_fmul_reassoc(
-; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc <4 x float> [[V:%.*]], <float 1.000000e+00, float 2.000000e+00, float 3.000000e+00, float 4.000000e+00>
-; CHECK-NEXT: [[TMP2:%.*]] = fmul reassoc <4 x float> [[TMP1]], <float 1.000000e+00, float 2.000000e+05, float -3.000000e+00, float 4.000000e+00>
-; CHECK-NEXT: ret <4 x float> [[TMP2]]
+; CHECK: [[TMP1:%.*]] = fmul reassoc <4 x float> %V, <float 1.000000e+00, float 4.000000e+05, float -9.000000e+00, float 1.600000e+01>
+; CHECK-NEXT: ret <4 x float> [[TMP1]]
%Y = fmul reassoc <4 x float> %V, < float 1.000000e+00, float 2.000000e+00, float 3.000000e+00, float 4.000000e+00 >
%Z = fmul reassoc <4 x float> %Y, < float 1.000000e+00, float 2.000000e+05, float -3.000000e+00, float 4.000000e+00 >
ret <4 x float> %Z
diff --git a/llvm/test/Transforms/InstCombine/fdiv.ll b/llvm/test/Transforms/InstCombine/fdiv.ll
index 54b0bf8c50ac7..3465781e3af9d 100644
--- a/llvm/test/Transforms/InstCombine/fdiv.ll
+++ b/llvm/test/Transforms/InstCombine/fdiv.ll
@@ -525,8 +525,7 @@ define <2 x float> @div_constant_dividend2_reassoc_only(<2 x float> %x) {
define <2 x float> @div_constant_dividend3(<2 x float> %x) {
; CHECK-LABEL: @div_constant_dividend3(
-; 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: [[T2:%.*]] = fmul reassoc arcp <2 x float> [[X:%.*]], <float 5.000000e+00, float -1.000000e+00>
; CHECK-NEXT: ret <2 x float> [[T2]]
;
%t1 = fdiv <2 x float> <float 3.0e0, float 7.0e0>, %x
diff --git a/llvm/test/Transforms/InstCombine/issue64967-reassoc-fmul.ll b/llvm/test/Transforms/InstCombine/issue64967-reassoc-fmul.ll
index 16f9cf2dd64c5..5d064234bf609 100644
--- a/llvm/test/Transforms/InstCombine/issue64967-reassoc-fmul.ll
+++ b/llvm/test/Transforms/InstCombine/issue64967-reassoc-fmul.ll
@@ -25,8 +25,7 @@ define float @fmul(float %x) {
define float @fmul_reassoc(float %x) {
; CHECK-LABEL: define float @fmul_reassoc(
; CHECK-SAME: float [[X:%.*]]) {
-; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc float [[X]], 2.000000e+00
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[FMUL0]], 4.000000e+00
+; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[X]], 8.000000e+00
; CHECK-NEXT: ret float [[FMUL1]]
;
%fmul0 = fmul reassoc float %x, 2.0
@@ -37,8 +36,7 @@ define float @fmul_reassoc(float %x) {
define <2 x float> @fmul_reassoc_v2(<2 x float> %x) {
; CHECK-LABEL: define <2 x float> @fmul_reassoc_v2(
; CHECK-SAME: <2 x float> [[X:%.*]]) {
-; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc <2 x float> [[X]], splat (float 2.000000e+00)
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc <2 x float> [[FMUL0]], splat (float 4.000000e+00)
+; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc <2 x float> [[X]], splat (float 8.000000e+00)
; CHECK-NEXT: ret <2 x float> [[FMUL1]]
;
%fmul0 = fmul reassoc <2 x float> %x, splat (float 2.0)
@@ -54,8 +52,7 @@ define <2 x float> @fmul_reassoc_v2(<2 x float> %x) {
define float @fmul_reassoc_negative_0(float %x) {
; CHECK-LABEL: define float @fmul_reassoc_negative_0(
; CHECK-SAME: float [[X:%.*]]) {
-; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc float [[X]], 2.000000e+00
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[FMUL0]], -4.000000e+00
+; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[X]], -8.000000e+00
; CHECK-NEXT: ret float [[FMUL1]]
;
%fmul0 = fmul reassoc float %x, 2.0
@@ -71,8 +68,7 @@ define float @fmul_reassoc_negative_0(float %x) {
define float @fmul_reassoc_negative_1(float %x) {
; CHECK-LABEL: define float @fmul_reassoc_negative_1(
; CHECK-SAME: float [[X:%.*]]) {
-; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc float [[X]], -2.000000e+00
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[FMUL0]], 4.000000e+00
+; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[X]], -8.000000e+00
; CHECK-NEXT: ret float [[FMUL1]]
;
%fmul0 = fmul reassoc float %x, -2.0
@@ -95,8 +91,7 @@ define float @fmul_reassoc_nsz(float %x) {
define float @fmul_reassoc_posk_neg0(float %x) {
; CHECK-LABEL: define float @fmul_reassoc_posk_neg0(
; CHECK-SAME: float [[X:%.*]]) {
-; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc float [[X]], 4.000000e+00
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[FMUL0]], -0.000000e+00
+; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[X]], -0.000000e+00
; CHECK-NEXT: ret float [[FMUL1]]
;
%fmul0 = fmul reassoc float %x, 4.0
@@ -108,8 +103,7 @@ define float @fmul_reassoc_neg0_posk(float %x) {
; CHECK-LABEL: define float @fmul_reassoc_neg0_posk(
; CHECK-SAME: float [[X:%.*]]) {
; CHECK-NEXT: [[FMUL0:%.*]] = fmul reassoc float [[X]], -0.000000e+00
-; CHECK-NEXT: [[FMUL1:%.*]] = fmul reassoc float [[FMUL0]], 4.000000e+00
-; CHECK-NEXT: ret float [[FMUL1]]
+; CHECK-NEXT: ret float [[FMUL0]]
;
%fmul0 = fmul reassoc float %x, -0.0
%fmul1 = fmul reassoc float %fmul0, 4.0
|
This was referenced Dec 10, 2025
andykaylor
approved these changes
Dec 10, 2025
Contributor
andykaylor
left a comment
andykaylor
left a comment
There was a problem hiding this comment.
This change looks good, but I have a concern about the prior state.
| ; CHECK-LABEL: @div_constant_dividend3( | ||
| ; 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: [[T2:%.*]] = fmul reassoc arcp <2 x float> [[X:%.*]], <float 5.000000e+00, float -1.000000e+00> |
Contributor
There was a problem hiding this comment.
This one doesn't look right, but not because of the change in this PR. Shouldn't we have required 1reassocandrcpon bothfdivinstructions before converting them into twofmulinstructions withressocandarcp` on both?
Contributor
Author
There was a problem hiding this comment.
Yes. Many places still treat reassoc as viral
Base automatically changed from
users/arsenm/issue64697/instcombine-add-baseline-test-reassoc-fmul
to
main
December 11, 2025 10:37
nsz can only change the behavior of the sign bit. The sign bit for fmul can be implemented as xor, which is associative. DAGCombiner already reassociates the multiply by 2 constants without nsz. Fixes #64967
arsenm
force-pushed
the
users/arsenm/issue64697/allow-reassoc-fmul-no-nsz
branch
from
a7f8f32 to
966cb03
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
nsz can only change the behavior of the sign bit.
The sign bit for fmul can be implemented as xor,
which is associative. DAGCombiner already reassociates
the multiply by 2 constants without nsz.
Fixes #64967