[llvm] 7bee944 - [InstCombine] form copysign from select of FP constants (PR44153)
Sanjay Patel via llvm-commits
llvm-commits at lists.llvm.org
Mon Jan 20 07:53:41 PST 2020
Author: Sanjay Patel
Date: 2020-01-20T10:51:14-05:00
New Revision: 7bee94410ce265833695128dfad7fbd7a8eef6ba
URL: https://github.com/llvm/llvm-project/commit/7bee94410ce265833695128dfad7fbd7a8eef6ba
DIFF: https://github.com/llvm/llvm-project/commit/7bee94410ce265833695128dfad7fbd7a8eef6ba.diff
LOG: [InstCombine] form copysign from select of FP constants (PR44153)
This should be the last step needed to solve the problem in the
description of PR44153:
https://bugs.llvm.org/show_bug.cgi?id=44153
If we're casting an FP value to int, testing its signbit, and then
choosing between a value and its negated value, that's a
complicated way of saying "copysign":
(bitcast X) < 0 ? -TC : TC --> copysign(TC, X)
Differential Revision: https://reviews.llvm.org/D72643
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
llvm/test/Transforms/InstCombine/select.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index 05a624fde86b..2628e57deba5 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -2317,6 +2317,48 @@ static Instruction *foldSelectRotate(SelectInst &Sel) {
return IntrinsicInst::Create(F, { TVal, TVal, ShAmt });
}
+static Instruction *foldSelectToCopysign(SelectInst &Sel,
+ InstCombiner::BuilderTy &Builder) {
+ Value *Cond = Sel.getCondition();
+ Value *TVal = Sel.getTrueValue();
+ Value *FVal = Sel.getFalseValue();
+ Type *SelType = Sel.getType();
+
+ // Match select ?, TC, FC where the constants are equal but negated.
+ // TODO: Generalize to handle a negated variable operand?
+ const APFloat *TC, *FC;
+ if (!match(TVal, m_APFloat(TC)) || !match(FVal, m_APFloat(FC)) ||
+ !abs(*TC).bitwiseIsEqual(abs(*FC)))
+ return nullptr;
+
+ assert(TC != FC && "Expected equal select arms to simplify");
+
+ Value *X;
+ const APInt *C;
+ bool IsTrueIfSignSet;
+ ICmpInst::Predicate Pred;
+ if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) ||
+ !isSignBitCheck(Pred, *C, IsTrueIfSignSet) || X->getType() != SelType)
+ return nullptr;
+
+ // If needed, negate the value that will be the sign argument of the copysign:
+ // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X)
+ // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X)
+ // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X)
+ // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X)
+ if (IsTrueIfSignSet ^ TC->isNegative())
+ X = Builder.CreateFNegFMF(X, &Sel);
+
+ // Canonicalize the magnitude argument as the positive constant since we do
+ // not care about its sign.
+ Value *MagArg = TC->isNegative() ? FVal : TVal;
+ Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign,
+ Sel.getType());
+ Instruction *CopySign = IntrinsicInst::Create(F, { MagArg, X });
+ CopySign->setFastMathFlags(Sel.getFastMathFlags());
+ return CopySign;
+}
+
Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
Value *CondVal = SI.getCondition();
Value *TrueVal = SI.getTrueValue();
@@ -2785,5 +2827,8 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
if (Instruction *Rot = foldSelectRotate(SI))
return Rot;
+ if (Instruction *Copysign = foldSelectToCopysign(SI, Builder))
+ return Copysign;
+
return nullptr;
}
diff --git a/llvm/test/Transforms/InstCombine/select.ll b/llvm/test/Transforms/InstCombine/select.ll
index 2f04efe44b2a..0f26a93a7bae 100644
--- a/llvm/test/Transforms/InstCombine/select.ll
+++ b/llvm/test/Transforms/InstCombine/select.ll
@@ -1535,9 +1535,7 @@ define <2 x i32> @test_shl_zext_bool_vec(<2 x i1> %t) {
define float @copysign1(float %x) {
; CHECK-LABEL: @copysign1(
-; CHECK-NEXT: [[I:%.*]] = bitcast float [[X:%.*]] to i32
-; CHECK-NEXT: [[ISPOS:%.*]] = icmp sgt i32 [[I]], -1
-; CHECK-NEXT: [[R:%.*]] = select i1 [[ISPOS]], float 1.000000e+00, float -1.000000e+00
+; CHECK-NEXT: [[R:%.*]] = call float @llvm.copysign.f32(float 1.000000e+00, float [[X:%.*]])
; CHECK-NEXT: ret float [[R]]
;
%i = bitcast float %x to i32
@@ -1548,9 +1546,8 @@ define float @copysign1(float %x) {
define <2 x float> @copysign2(<2 x float> %x) {
; CHECK-LABEL: @copysign2(
-; CHECK-NEXT: [[I:%.*]] = bitcast <2 x float> [[X:%.*]] to <2 x i32>
-; CHECK-NEXT: [[ISNEG:%.*]] = icmp slt <2 x i32> [[I]], zeroinitializer
-; CHECK-NEXT: [[R:%.*]] = select nsz <2 x i1> [[ISNEG]], <2 x float> <float 4.200000e+01, float 4.200000e+01>, <2 x float> <float -4.200000e+01, float -4.200000e+01>
+; CHECK-NEXT: [[TMP1:%.*]] = fneg nsz <2 x float> [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = call nsz <2 x float> @llvm.copysign.v2f32(<2 x float> <float 4.200000e+01, float 4.200000e+01>, <2 x float> [[TMP1]])
; CHECK-NEXT: ret <2 x float> [[R]]
;
%i = bitcast <2 x float> %x to <2 x i32>
@@ -1561,9 +1558,8 @@ define <2 x float> @copysign2(<2 x float> %x) {
define float @copysign3(float %x) {
; CHECK-LABEL: @copysign3(
-; CHECK-NEXT: [[I:%.*]] = bitcast float [[X:%.*]] to i32
-; CHECK-NEXT: [[ISPOS:%.*]] = icmp sgt i32 [[I]], -1
-; CHECK-NEXT: [[R:%.*]] = select fast i1 [[ISPOS]], float -4.300000e+01, float 4.300000e+01
+; CHECK-NEXT: [[TMP1:%.*]] = fneg fast float [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = call fast float @llvm.copysign.f32(float 4.300000e+01, float [[TMP1]])
; CHECK-NEXT: ret float [[R]]
;
%i = bitcast float %x to i32
@@ -1572,6 +1568,8 @@ define float @copysign3(float %x) {
ret float %r
}
+; TODO: Allow undefs when matching vectors.
+
define <2 x float> @copysign4(<2 x float> %x) {
; CHECK-LABEL: @copysign4(
; CHECK-NEXT: [[I:%.*]] = bitcast <2 x float> [[X:%.*]] to <2 x i32>
@@ -1587,6 +1585,8 @@ define <2 x float> @copysign4(<2 x float> %x) {
declare void @use1(i1)
+; Negative test
+
define float @copysign_extra_use(float %x) {
; CHECK-LABEL: @copysign_extra_use(
; CHECK-NEXT: [[I:%.*]] = bitcast float [[X:%.*]] to i32
@@ -1602,6 +1602,8 @@ define float @copysign_extra_use(float %x) {
ret float %r
}
+; Negative test
+
define float @copysign_type_mismatch(double %x) {
; CHECK-LABEL: @copysign_type_mismatch(
; CHECK-NEXT: [[I:%.*]] = bitcast double [[X:%.*]] to i64
@@ -1615,6 +1617,8 @@ define float @copysign_type_mismatch(double %x) {
ret float %r
}
+; Negative test
+
define float @copysign_wrong_cmp(float %x) {
; CHECK-LABEL: @copysign_wrong_cmp(
; CHECK-NEXT: [[I:%.*]] = bitcast float [[X:%.*]] to i32
@@ -1628,6 +1632,8 @@ define float @copysign_wrong_cmp(float %x) {
ret float %r
}
+; Negative test
+
define float @copysign_wrong_const(float %x) {
; CHECK-LABEL: @copysign_wrong_const(
; CHECK-NEXT: [[I:%.*]] = bitcast float [[X:%.*]] to i32
More information about the llvm-commits
mailing list