[InstCombine] Relax one-use requirement for add iN (sext i1 X), (sext i1 Y) --> sext (X | Y) to iN

[AZero13]

Since these remove instructions as long as at least one of X or Y is one-use, we don't need to check one-use for both.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: AtariDreams (AtariDreams)

Changes

Since these remove instructions, we can forgo the one-use check.

---

Full diff: https://github.com/llvm/llvm-project/pull/90509.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp (+1-3)
• (modified) llvm/test/Transforms/InstCombine/add.ll (+11-17)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 6739b8745d74e4..953b0da564a48f 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -510,9 +510,7 @@ Value *InstCombinerImpl::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
       }
 
       // add iN (sext i1 X), (sext i1 Y) --> sext (X | Y) to iN
-      // TODO: Relax the one-use checks because we are removing an instruction?
-      if (match(I, m_Add(m_OneUse(m_SExt(m_Value(X))),
-                         m_OneUse(m_SExt(m_Value(Y))))) &&
+      if (match(I, m_Add(m_SExt(m_Value(X)), m_SExt(m_Value(Y)))) &&
           X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType()) {
         // Truth table for inputs and output signbits:
         //       X:0 | X:1
diff --git a/llvm/test/Transforms/InstCombine/add.ll b/llvm/test/Transforms/InstCombine/add.ll
index 56ee54d351e762..5c7826f33074c1 100644
--- a/llvm/test/Transforms/InstCombine/add.ll
+++ b/llvm/test/Transforms/InstCombine/add.ll
@@ -1435,15 +1435,12 @@ define i32 @and31_add_sexts(i1 %x, i1 %y) {
   ret i32 %r
 }
 
-; Negative test - extra use
-
 define i32 @lshr_add_use_sexts(i1 %x, i1 %y, ptr %p) {
 ; CHECK-LABEL: @lshr_add_use_sexts(
 ; CHECK-NEXT:    [[XS:%.*]] = sext i1 [[X:%.*]] to i32
 ; CHECK-NEXT:    store i32 [[XS]], ptr [[P:%.*]], align 4
-; CHECK-NEXT:    [[YS:%.*]] = sext i1 [[Y:%.*]] to i32
-; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[XS]], [[YS]]
-; CHECK-NEXT:    [[R:%.*]] = lshr i32 [[SUB]], 31
+; CHECK-NEXT:    [[TMP1:%.*]] = or i1 [[X]], [[Y:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = zext i1 [[TMP1]] to i32
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %xs = sext i1 %x to i32
@@ -1454,15 +1451,12 @@ define i32 @lshr_add_use_sexts(i1 %x, i1 %y, ptr %p) {
   ret i32 %r
 }
 
-; Negative test - extra use
-
 define i32 @lshr_add_use2_sexts(i1 %x, i1 %y, ptr %p) {
 ; CHECK-LABEL: @lshr_add_use2_sexts(
-; CHECK-NEXT:    [[XS:%.*]] = sext i1 [[X:%.*]] to i32
 ; CHECK-NEXT:    [[YS:%.*]] = sext i1 [[Y:%.*]] to i32
 ; CHECK-NEXT:    store i32 [[YS]], ptr [[P:%.*]], align 4
-; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[XS]], [[YS]]
-; CHECK-NEXT:    [[R:%.*]] = lshr i32 [[SUB]], 31
+; CHECK-NEXT:    [[TMP1:%.*]] = or i1 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[R:%.*]] = zext i1 [[TMP1]] to i32
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %xs = sext i1 %x to i32
@@ -4018,8 +4012,8 @@ define i32 @add_reduce_sqr_sum_varC_invalid2(i32 %a, i32 %b) {
 
 define i32 @fold_sext_addition_or_disjoint(i8 %x) {
 ; CHECK-LABEL: @fold_sext_addition_or_disjoint(
-; CHECK-NEXT:    [[SE:%.*]] = sext i8 [[XX:%.*]] to i32
-; CHECK-NEXT:    [[R:%.*]] = add nsw i32 [[SE]], 1246
+; CHECK-NEXT:    [[TMP1:%.*]] = sext i8 [[X:%.*]] to i32
+; CHECK-NEXT:    [[R:%.*]] = add nsw i32 [[TMP1]], 1246
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %xx = or disjoint i8 %x, 12
@@ -4043,8 +4037,8 @@ define i32 @fold_sext_addition_fail(i8 %x) {
 
 define i32 @fold_zext_addition_or_disjoint(i8 %x) {
 ; CHECK-LABEL: @fold_zext_addition_or_disjoint(
-; CHECK-NEXT:    [[SE:%.*]] = zext i8 [[XX:%.*]] to i32
-; CHECK-NEXT:    [[R:%.*]] = add nuw nsw i32 [[SE]], 1246
+; CHECK-NEXT:    [[TMP1:%.*]] = zext i8 [[X:%.*]] to i32
+; CHECK-NEXT:    [[R:%.*]] = add nuw nsw i32 [[TMP1]], 1246
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %xx = or disjoint i8 %x, 12
@@ -4055,9 +4049,9 @@ define i32 @fold_zext_addition_or_disjoint(i8 %x) {
 
 define i32 @fold_zext_addition_or_disjoint2(i8 %x) {
 ; CHECK-LABEL: @fold_zext_addition_or_disjoint2(
-; CHECK-NEXT:    [[XX:%.*]] = add nuw i8 [[X:%.*]], 4
-; CHECK-NEXT:    [[SE:%.*]] = zext i8 [[XX]] to i32
-; CHECK-NEXT:    ret i32 [[SE]]
+; CHECK-NEXT:    [[TMP1:%.*]] = add nuw i8 [[X:%.*]], 4
+; CHECK-NEXT:    [[R:%.*]] = zext i8 [[TMP1]] to i32
+; CHECK-NEXT:    ret i32 [[R]]
 ;
   %xx = or disjoint i8 %x, 18
   %se = zext i8 %xx to i32

[nikic]

Increases instruction count if both sexts have extra uses.

[AZero13]

I guess I'm gonna test that then

[AZero13]

Increases instruction count if both sexts have extra uses.

Just checked and at worst the count stays the same.

[topperc]

Increases instruction count if both sexts have extra uses.

Just checked and at worst the count stays the same.

In the worst case doesn't it remove 1 Add and insert an Or and a SExt. That's an increase of 1.

[nikic]

The case with two extra uses is still untested?

[goldsteinn]

The the title should the "Relax one-use..." instead of "Remove one-use..."

[AZero13]

The the title should the "Relax one-use..." instead of "Remove one-use..."

Fixed!

[nikic]

LGTM

[goldsteinn]

This appears to have a pretty significant compile time impact: https://llvm-compile-time-tracker.com/compare.php?from=519e0bb094bd9444b64ac62f2e8192543fdb94b9&to=dd0245e3d74955ace14c95d288abf72346ad52e8&stat=instructions:u

[nikic]

@goldsteinn That range also contains other commits. The regression is mostly #96851.