[InstCombine] Clear sign-bit of the constant magnitude in copysign

[dtcxzyw]

Alive2: https://alive2.llvm.org/ce/z/vFykcZ
Address the comment #85772 (comment).

Unfortunately, non-splat vector constants are not supported because we haven't implemented constant folding of fabs with vector operands.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/85787.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp (+10)
• (modified) llvm/test/Transforms/InstCombine/copysign-fneg-fabs.ll (+46)
• (modified) llvm/test/Transforms/InstCombine/fcmp.ll (+5-5)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 8537dbc6fe531b..3eafcb03644ecd 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -2473,6 +2473,16 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
     if (match(Sign, m_Intrinsic<Intrinsic::copysign>(m_Value(), m_Value(X))))
       return replaceOperand(*II, 1, X);
 
+    // Clear sign-bit of constant magnitude:
+    // copysign -MagC, X --> copysign MagC, X
+    // TODO: Support constant folding for fabs
+    const APFloat *MagC;
+    if (match(Mag, m_APFloat(MagC)) && MagC->isNegative()) {
+      APFloat PosMagC = *MagC;
+      PosMagC.clearSign();
+      return replaceOperand(*II, 0, ConstantFP::get(Mag->getType(), PosMagC));
+    }
+
     // Peek through changes of magnitude's sign-bit. This call rewrites those:
     // copysign (fabs X), Sign --> copysign X, Sign
     // copysign (fneg X), Sign --> copysign X, Sign
diff --git a/llvm/test/Transforms/InstCombine/copysign-fneg-fabs.ll b/llvm/test/Transforms/InstCombine/copysign-fneg-fabs.ll
index af939cf74399ab..84cb252a150d7a 100644
--- a/llvm/test/Transforms/InstCombine/copysign-fneg-fabs.ll
+++ b/llvm/test/Transforms/InstCombine/copysign-fneg-fabs.ll
@@ -275,6 +275,52 @@ define half @fneg_fabs_copysign_multi_use_fabs(half %x, half %y, ptr %ptr) {
   ret half %fabs.copysign
 }
 
+define half @copysign_pos(half %a) {
+; CHECK-LABEL: @copysign_pos(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[RET:%.*]] = call half @llvm.copysign.f16(half 0xH3C00, half [[A:%.*]])
+; CHECK-NEXT:    ret half [[RET]]
+;
+entry:
+  %ret = call half @llvm.copysign.f16(half 0xH3C00, half %a)
+  ret half %ret
+}
+
+define half @copysign_neg(half %a) {
+; CHECK-LABEL: @copysign_neg(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[RET:%.*]] = call half @llvm.copysign.f16(half 0xH3C00, half [[A:%.*]])
+; CHECK-NEXT:    ret half [[RET]]
+;
+entry:
+  %ret = call half @llvm.copysign.f16(half 0xHBC00, half %a)
+  ret half %ret
+}
+
+define <4 x half> @copysign_splat(<4 x half> %a) {
+; CHECK-LABEL: @copysign_splat(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[RET:%.*]] = call <4 x half> @llvm.copysign.v4f16(<4 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00>, <4 x half> [[A:%.*]])
+; CHECK-NEXT:    ret <4 x half> [[RET]]
+;
+entry:
+  %ret = call <4 x half> @llvm.copysign.v4f16(<4 x half> splat(half 0xHBC00), <4 x half> %a)
+  ret <4 x half> %ret
+}
+
+; TODO: Support constant folding of fabs
+
+define <4 x half> @copysign_vec4(<4 x half> %a) {
+; CHECK-LABEL: @copysign_vec4(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[RET:%.*]] = call <4 x half> @llvm.copysign.v4f16(<4 x half> <half 0xH3C00, half 0xHBC00, half undef, half poison>, <4 x half> [[A:%.*]])
+; CHECK-NEXT:    ret <4 x half> [[RET]]
+;
+entry:
+  %ret = call <4 x half> @llvm.copysign.v4f16(<4 x half> <half 0xH3C00, half 0xHBC00, half undef, half poison>, <4 x half> %a)
+  ret <4 x half> %ret
+}
+
 declare half @llvm.fabs.f16(half)
 declare <2 x half> @llvm.fabs.v2f16(<2 x half>)
 declare half @llvm.copysign.f16(half, half)
diff --git a/llvm/test/Transforms/InstCombine/fcmp.ll b/llvm/test/Transforms/InstCombine/fcmp.ll
index 159c84d0dd8aa9..f2701d16d0f3d1 100644
--- a/llvm/test/Transforms/InstCombine/fcmp.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp.ll
@@ -644,7 +644,7 @@ define <2 x i1> @is_signbit_set_anyzero(<2 x double> %x) {
 
 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:    [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
 ; CHECK-NEXT:    [[R:%.*]] = fcmp ogt double [[S]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
@@ -655,7 +655,7 @@ 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:    [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
 ; CHECK-NEXT:    [[R:%.*]] = fcmp ugt double [[S]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
@@ -666,7 +666,7 @@ 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:    [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
 ; CHECK-NEXT:    [[R:%.*]] = fcmp oge double [[S]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
@@ -677,7 +677,7 @@ 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:    [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
 ; CHECK-NEXT:    [[R:%.*]] = fcmp uge double [[S]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
@@ -705,7 +705,7 @@ define i1 @is_signbit_set_extra_use(double %x, ptr %p) {
 
 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:    [[S:%.*]] = call double @llvm.copysign.f64(double 4.200000e+01, double [[X:%.*]])
 ; CHECK-NEXT:    [[R:%.*]] = fcmp ogt double [[S]], 1.000000e+00
 ; CHECK-NEXT:    ret i1 [[R]]
 ;