[llvm] 10e7671 - Reland "[InstCombine] Fold (sub nuw X, (Y << nuw Z)) >>u exact Z --> (X >>u exact Z) sub nuw Y" (#93571)

via llvm-commits llvm-commits at lists.llvm.org
Sun Jun 2 23:24:59 PDT 2024


Author: AtariDreams
Date: 2024-06-03T08:24:54+02:00
New Revision: 10e7671d9adf1da2f08d67fde6585c707ac4c1d2

URL: https://github.com/llvm/llvm-project/commit/10e7671d9adf1da2f08d67fde6585c707ac4c1d2
DIFF: https://github.com/llvm/llvm-project/commit/10e7671d9adf1da2f08d67fde6585c707ac4c1d2.diff

LOG: Reland "[InstCombine] Fold (sub nuw X, (Y << nuw Z)) >>u exact Z --> (X >>u exact Z) sub nuw Y" (#93571)

This is the same fold as ((X << nuw Z) sub nuw Y) >>u exact Z --> X sub
nuw (Y >>u exact Z), but with the sub operands swapped.

Alive2 Proof:
https://alive2.llvm.org/ce/z/pT-RxG

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
    llvm/test/Transforms/InstCombine/lshr.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
index fbc02cddfb005..9ff817da79368 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -1275,6 +1275,17 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
     return NewSub;
   }
 
+  // (sub nuw X, (Y << nuw Z)) >>u exact Z --> (X >>u exact Z) sub nuw Y
+  if (I.isExact() &&
+      match(Op0, m_OneUse(m_NUWSub(m_Value(X),
+                                   m_NUWShl(m_Value(Y), m_Specific(Op1)))))) {
+    Value *NewLshr = Builder.CreateLShr(X, Op1, "", /*isExact=*/true);
+    auto *NewSub = BinaryOperator::CreateNUWSub(NewLshr, Y);
+    NewSub->setHasNoSignedWrap(
+        cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap());
+    return NewSub;
+  }
+
   auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) {
     switch (BinOpcode) {
     default:

diff  --git a/llvm/test/Transforms/InstCombine/lshr.ll b/llvm/test/Transforms/InstCombine/lshr.ll
index 0392764490945..60f03733ebd99 100644
--- a/llvm/test/Transforms/InstCombine/lshr.ll
+++ b/llvm/test/Transforms/InstCombine/lshr.ll
@@ -464,6 +464,119 @@ define i32 @shl_sub_lshr(i32 %x, i32 %c, i32 %y) {
   ret i32 %lshr
 }
 
+define i32 @shl_sub_lshr_reverse(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse(
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr exact i32 [[Y:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[LSHR:%.*]] = sub nuw nsw i32 [[TMP1]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw i32 %x, %c
+  %sub = sub nuw nsw i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+define i32 @shl_sub_lshr_reverse_no_nsw(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_no_nsw(
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr exact i32 [[Y:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[LSHR:%.*]] = sub nuw i32 [[TMP1]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw i32 %x, %c
+  %sub = sub nuw i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+define i32 @shl_sub_lshr_reverse_nsw_on_op1(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_nsw_on_op1(
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr exact i32 [[Y:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[LSHR:%.*]] = sub nuw i32 [[TMP1]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw nsw i32 %x, %c
+  %sub = sub nuw i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+; Negative test
+
+define i32 @shl_sub_lshr_reverse_no_exact(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_no_exact(
+; CHECK-NEXT:    [[SHL:%.*]] = shl nuw i32 [[X:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub nuw nsw i32 [[Y:%.*]], [[SHL]]
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr i32 [[SUB]], [[C]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw i32 %x, %c
+  %sub = sub nuw nsw i32 %y, %shl
+  %lshr = lshr i32 %sub, %c
+  ret i32 %lshr
+}
+
+; Negative test
+
+define i32 @shl_sub_lshr_reverse_multiuse(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_multiuse(
+; CHECK-NEXT:    [[SHL:%.*]] = shl nuw i32 [[X:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub nuw i32 [[Y:%.*]], [[SHL]]
+; CHECK-NEXT:    call void @use(i32 [[SUB]])
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr exact i32 [[SUB]], [[C]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw i32 %x, %c
+  %sub = sub nuw i32 %y, %shl
+  call void @use(i32 %sub)
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+define i32 @shl_sub_lshr_reverse_multiuse2(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_multiuse2(
+; CHECK-NEXT:    [[SHL:%.*]] = shl nuw i32 [[X:%.*]], [[C:%.*]]
+; CHECK-NEXT:    call void @use(i32 [[SHL]])
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr exact i32 [[Y:%.*]], [[C]]
+; CHECK-NEXT:    [[LSHR:%.*]] = sub nuw i32 [[TMP1]], [[X]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw i32 %x, %c
+  call void @use(i32 %shl)
+  %sub = sub nuw i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+; Negative test
+
+define i32 @shl_sub_lshr_reverse_no_nuw(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_no_nuw(
+; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[X:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub nuw i32 [[Y:%.*]], [[SHL]]
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr exact i32 [[SUB]], [[C]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl i32 %x, %c
+  %sub = sub nuw i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
+; Negative test
+
+define i32 @shl_sub_lshr_reverse_no_nsw_2(i32 %x, i32 %c, i32 %y) {
+; CHECK-LABEL: @shl_sub_lshr_reverse_no_nsw_2(
+; CHECK-NEXT:    [[SHL:%.*]] = shl nuw nsw i32 [[X:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 [[Y:%.*]], [[SHL]]
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr exact i32 [[SUB]], [[C]]
+; CHECK-NEXT:    ret i32 [[LSHR]]
+;
+  %shl = shl nuw nsw i32 %x, %c
+  %sub = sub i32 %y, %shl
+  %lshr = lshr exact i32 %sub, %c
+  ret i32 %lshr
+}
+
 define i32 @shl_or_lshr(i32 %x, i32 %c, i32 %y) {
 ; CHECK-LABEL: @shl_or_lshr(
 ; CHECK-NEXT:    [[TMP1:%.*]] = lshr i32 [[Y:%.*]], [[C:%.*]]


        


More information about the llvm-commits mailing list