[InstCombine] Generalize fold of fcmp + copysign
#86387
Open
+133
−34
Conversation
This file contains 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
dtcxzyw
requested review from
arsenm,
goldsteinn,
andykaylor and
jcranmer-intel
|
@llvm/pr-subscribers-llvm-transforms Author: Yingwei Zheng (dtcxzyw) ChangesThis patch generalize the fold of It will improve the codegen of idiom Full diff: https://github.com/llvm/llvm-project/pull/86387.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index db302d7e526844..076452200fead5 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -8101,22 +8101,22 @@ Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
}
// Convert a sign-bit test of an FP value into a cast and integer compare.
- // TODO: Simplify if the copysign constant is 0.0 or NaN.
- // TODO: Handle non-zero compare constants.
- // TODO: Handle other predicates.
if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::copysign>(m_APFloat(C),
m_Value(X)))) &&
- match(Op1, m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
+ match(Op1, m_ImmConstant(RHSC))) {
Type *IntType = Builder.getIntNTy(X->getType()->getScalarSizeInBits());
if (auto *VecTy = dyn_cast<VectorType>(OpType))
IntType = VectorType::get(IntType, VecTy->getElementCount());
- // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
- if (Pred == FCmpInst::FCMP_OLT) {
- Value *IntX = Builder.CreateBitCast(X, IntType);
- return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
- ConstantInt::getNullValue(IntType));
- }
+ APFloat PosC = abs(*C);
+ if (Value *CmpPos = ConstantFoldCompareInstOperands(
+ Pred, ConstantFP::get(X->getType(), PosC), RHSC, DL, &TLI, &I))
+ if (Value *CmpNeg = ConstantFoldCompareInstOperands(
+ Pred, ConstantFP::get(X->getType(), -PosC), RHSC, DL, &TLI, &I)) {
+ Value *IntX = Builder.CreateBitCast(X, IntType);
+ Value *NotNeg = Builder.CreateIsNotNeg(IntX);
+ return SelectInst::Create(NotNeg, CmpPos, CmpNeg);
+ }
}
{
diff --git a/llvm/test/Transforms/InstCombine/fcmp.ll b/llvm/test/Transforms/InstCombine/fcmp.ll
index f2701d16d0f3d1..cff533ce5e9cc8 100644
--- a/llvm/test/Transforms/InstCombine/fcmp.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp.ll
@@ -596,8 +596,8 @@ define i1 @is_signbit_set(double %x) {
define i1 @is_signbit_set_1(double %x) {
; CHECK-LABEL: @is_signbit_set_1(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 1.000000e+00, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ult double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp slt i64 [[TMP1]], 0
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double 1.0, double %x)
@@ -607,8 +607,8 @@ define i1 @is_signbit_set_1(double %x) {
define i1 @is_signbit_set_2(double %x) {
; CHECK-LABEL: @is_signbit_set_2(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 1.000000e+00, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ole double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp slt i64 [[TMP1]], 0
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double 1.0, double %x)
@@ -618,8 +618,8 @@ define i1 @is_signbit_set_2(double %x) {
define i1 @is_signbit_set_3(double %x) {
; CHECK-LABEL: @is_signbit_set_3(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 1.000000e+00, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ule double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp slt i64 [[TMP1]], 0
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double 1.0, double %x)
@@ -640,12 +640,10 @@ define <2 x i1> @is_signbit_set_anyzero(<2 x double> %x) {
ret <2 x i1> %r
}
-; TODO: Handle different predicates.
-
define i1 @is_signbit_clear(double %x) {
; CHECK-LABEL: @is_signbit_clear(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ogt double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp sgt i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double -42.0, double %x)
@@ -655,8 +653,8 @@ define i1 @is_signbit_clear(double %x) {
define i1 @is_signbit_clear_1(double %x) {
; CHECK-LABEL: @is_signbit_clear_1(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ugt double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp sgt i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double -42.0, double %x)
@@ -666,8 +664,8 @@ define i1 @is_signbit_clear_1(double %x) {
define i1 @is_signbit_clear_2(double %x) {
; CHECK-LABEL: @is_signbit_clear_2(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp oge double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp sgt i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double -42.0, double %x)
@@ -677,8 +675,8 @@ define i1 @is_signbit_clear_2(double %x) {
define i1 @is_signbit_clear_3(double %x) {
; CHECK-LABEL: @is_signbit_clear_3(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp uge double [[S]], 0.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp sgt i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double -42.0, double %x)
@@ -701,12 +699,10 @@ define i1 @is_signbit_set_extra_use(double %x, ptr %p) {
ret i1 %r
}
-; TODO: Handle non-zero compare constant.
-
define i1 @is_signbit_clear_nonzero(double %x) {
; CHECK-LABEL: @is_signbit_clear_nonzero(
-; CHECK-NEXT: [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
-; CHECK-NEXT: [[R:%.*]] = fcmp ogt double [[S]], 1.000000e+00
+; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64
+; CHECK-NEXT: [[R:%.*]] = icmp sgt i64 [[TMP1]], -1
; CHECK-NEXT: ret i1 [[R]]
;
%s = call double @llvm.copysign.f64(double -42.0, double %x)
@@ -714,8 +710,6 @@ define i1 @is_signbit_clear_nonzero(double %x) {
ret i1 %r
}
-; TODO: Handle zero copysign constant.
-
define i1 @is_signbit_set_simplify_zero(double %x) {
; CHECK-LABEL: @is_signbit_set_simplify_zero(
; CHECK-NEXT: ret i1 false
@@ -725,8 +719,6 @@ define i1 @is_signbit_set_simplify_zero(double %x) {
ret i1 %r
}
-; TODO: Handle NaN copysign constant.
-
define i1 @is_signbit_set_simplify_nan(double %x) {
; CHECK-LABEL: @is_signbit_set_simplify_nan(
; CHECK-NEXT: ret i1 false
@@ -736,6 +728,54 @@ define i1 @is_signbit_set_simplify_nan(double %x) {
ret i1 %r
}
+define i1 @test_oeq(float %a) {
+; CHECK-LABEL: @test_oeq(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = bitcast float [[A:%.*]] to i32
+; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP0]], -1
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+entry:
+ %res = call float @llvm.copysign.f32(float 1.0, float %a)
+ %cmp = fcmp oeq float %res, 1.0
+ ret i1 %cmp
+}
+
+define i1 @test_one(float %a) {
+; CHECK-LABEL: @test_one(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = bitcast float [[A:%.*]] to i32
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP0]], 0
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+entry:
+ %res = call float @llvm.copysign.f32(float 1.0, float %a)
+ %cmp = fcmp one float %res, 1.0
+ ret i1 %cmp
+}
+
+define i1 @test_ogt_false(float %a) {
+; CHECK-LABEL: @test_ogt_false(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: ret i1 false
+;
+entry:
+ %res = call float @llvm.copysign.f32(float 1.0, float %a)
+ %cmp = fcmp ogt float %res, 2.0
+ ret i1 %cmp
+}
+
+define i1 @test_olt_true(float %a) {
+; CHECK-LABEL: @test_olt_true(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: ret i1 true
+;
+entry:
+ %res = call float @llvm.copysign.f32(float 1.0, float %a)
+ %cmp = fcmp olt float %res, 2.0
+ ret i1 %cmp
+}
+
define <2 x i1> @lossy_oeq(<2 x float> %x) {
; CHECK-LABEL: @lossy_oeq(
; CHECK-NEXT: ret <2 x i1> zeroinitializer
|
dtcxzyw
added a commit
to dtcxzyw/llvm-opt-benchmark
that referenced
this pull request
Mar 23, 2024
|
✅ With the latest revision this PR passed the Python code formatter. |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
andykaylor
reviewed
Mar 25, 2024
arsenm
reviewed
Mar 26, 2024
Comment on lines
+599
to
+600
| ; CHECK-NEXT: [[TMP1:%.*]] = bitcast double [[X:%.*]] to i64 | ||
| ; CHECK-NEXT: [[R:%.*]] = icmp slt i64 [[TMP1]], 0 |
There was a problem hiding this comment.
Given that we currently don't try to track FP class / sign bit through bitcast, this form seems worse? Plus, if we had fast math flags this would necessarily lose them
There was a problem hiding this comment.
Should we introduce a new intrinsic i1 @llvm.signbit?
See also https://discourse.llvm.org/t/do-we-need-intrinsics-for-floating-point-classification-functions/58852/8
There was a problem hiding this comment.
If it returns an i1, that wouldn't really be any better. It still wouldn't be a candidate for tracking. If we returned the signbit encoded in the original FP type, it may be more useful (i.e., it returns 0 or -0)
dtcxzyw
force-pushed
the
perf/fcmp-copysign
branch
from
25e1ce4
to
6085802
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.
This patch generalize the fold of
fcmp + copysign:It will improve the codegen of idiom
fcmp oeq (copysign C, X), +/-C.Godbolt: https://godbolt.org/z/a8zPheTqT