[llvm] 0f8b668 - [InstCombine] narrow popcount with zext operand

[Sanjay Patel via llvm-commits]

Author: Sanjay Patel
Date: 2021-04-29T15:07:16-04:00
New Revision: 0f8b6686ac288cda8d14d2ec5b8ca98d188b0684

URL: https://github.com/llvm/llvm-project/commit/0f8b6686ac288cda8d14d2ec5b8ca98d188b0684
DIFF: https://github.com/llvm/llvm-project/commit/0f8b6686ac288cda8d14d2ec5b8ca98d188b0684.diff

LOG: [InstCombine] narrow popcount with zext operand

https://llvm.org/PR50141

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
    llvm/test/Transforms/InstCombine/ctpop.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 262c2aa9694fb..345290aac384e 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -528,6 +528,13 @@ static Instruction *foldCtpop(IntrinsicInst &II, InstCombinerImpl &IC) {
     return CallInst::Create(F, {X, IC.Builder.getFalse()});
   }
 
+  // Zext doesn't change the number of set bits, so narrow:
+  // ctpop (zext X) --> zext (ctpop X)
+  if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
+    Value *NarrowPop = IC.Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, X);
+    return CastInst::Create(Instruction::ZExt, NarrowPop, Ty);
+  }
+
   KnownBits Known(BitWidth);
   IC.computeKnownBits(Op0, Known, 0, &II);
 

diff  --git a/llvm/test/Transforms/InstCombine/ctpop.ll b/llvm/test/Transforms/InstCombine/ctpop.ll
index 617506fc7f6a3..d1e6699994d9a 100644
--- a/llvm/test/Transforms/InstCombine/ctpop.ll
+++ b/llvm/test/Transforms/InstCombine/ctpop.ll
@@ -349,8 +349,8 @@ define <2 x i32> @sub_ctpop_vec_extra_use(<2 x i32> %a, <2 x i32>* %p) {
 
 define i32 @zext_ctpop(i16 %x) {
 ; CHECK-LABEL: @zext_ctpop(
-; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i32
-; CHECK-NEXT:    [[P:%.*]] = call i32 @llvm.ctpop.i32(i32 [[Z]]), !range [[RNG3:![0-9]+]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call i16 @llvm.ctpop.i16(i16 [[X:%.*]]), !range [[RNG3:![0-9]+]]
+; CHECK-NEXT:    [[P:%.*]] = zext i16 [[TMP1]] to i32
 ; CHECK-NEXT:    ret i32 [[P]]
 ;
   %z = zext i16 %x to i32
@@ -360,8 +360,8 @@ define i32 @zext_ctpop(i16 %x) {
 
 define <2 x i32> @zext_ctpop_vec(<2 x i7> %x) {
 ; CHECK-LABEL: @zext_ctpop_vec(
-; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i7> [[X:%.*]] to <2 x i32>
-; CHECK-NEXT:    [[P:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[Z]])
+; CHECK-NEXT:    [[TMP1:%.*]] = call <2 x i7> @llvm.ctpop.v2i7(<2 x i7> [[X:%.*]])
+; CHECK-NEXT:    [[P:%.*]] = zext <2 x i7> [[TMP1]] to <2 x i32>
 ; CHECK-NEXT:    ret <2 x i32> [[P]]
 ;
   %z = zext <2 x i7> %x to <2 x i32>
@@ -373,7 +373,7 @@ define i32 @zext_ctpop_extra_use(i16 %x, i32* %q) {
 ; CHECK-LABEL: @zext_ctpop_extra_use(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i32
 ; CHECK-NEXT:    store i32 [[Z]], i32* [[Q:%.*]], align 4
-; CHECK-NEXT:    [[P:%.*]] = call i32 @llvm.ctpop.i32(i32 [[Z]]), !range [[RNG3]]
+; CHECK-NEXT:    [[P:%.*]] = call i32 @llvm.ctpop.i32(i32 [[Z]]), !range [[RNG4:![0-9]+]]
 ; CHECK-NEXT:    ret i32 [[P]]
 ;
   %z = zext i16 %x to i32