[llvm] 0eff6c6 - [InstCombine] add vector support for (A >> C) == (B >> C) --> (A^B) u< (1 << C)

[Chenbing Zheng via llvm-commits]

Author: Chenbing Zheng
Date: 2022-06-20T10:55:47+08:00
New Revision: 0eff6c6ba81c4d5f2a4b60fbc0f44c35529065c5

URL: https://github.com/llvm/llvm-project/commit/0eff6c6ba81c4d5f2a4b60fbc0f44c35529065c5
DIFF: https://github.com/llvm/llvm-project/commit/0eff6c6ba81c4d5f2a4b60fbc0f44c35529065c5.diff

LOG: [InstCombine] add vector support for (A >> C) == (B >> C) --> (A^B) u< (1 << C)

Reviewed By: spatel, RKSimon

Differential Revision: https://reviews.llvm.org/D127398

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
    llvm/test/Transforms/InstCombine/compare-signs.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index c4bfdcf295252..afcd0bed257c1 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -4620,18 +4620,21 @@ Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
 
   // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
   // For lshr and ashr pairs.
-  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
-       match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
-      (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
-       match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
-    unsigned TypeBits = Cst1->getBitWidth();
-    unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
+  const APInt *AP1, *AP2;
+  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_APIntAllowUndef(AP1)))) &&
+       match(Op1, m_OneUse(m_LShr(m_Value(B), m_APIntAllowUndef(AP2))))) ||
+      (match(Op0, m_OneUse(m_AShr(m_Value(A), m_APIntAllowUndef(AP1)))) &&
+       match(Op1, m_OneUse(m_AShr(m_Value(B), m_APIntAllowUndef(AP2)))))) {
+    if (AP1 != AP2)
+      return nullptr;
+    unsigned TypeBits = AP1->getBitWidth();
+    unsigned ShAmt = AP1->getLimitedValue(TypeBits);
     if (ShAmt < TypeBits && ShAmt != 0) {
       ICmpInst::Predicate NewPred =
           Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
       Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
       APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
-      return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
+      return new ICmpInst(NewPred, Xor, ConstantInt::get(A->getType(), CmpVal));
     }
   }
 

diff  --git a/llvm/test/Transforms/InstCombine/compare-signs.ll b/llvm/test/Transforms/InstCombine/compare-signs.ll
index 7b56a7b7bde3f..206fe259580e7 100644
--- a/llvm/test/Transforms/InstCombine/compare-signs.ll
+++ b/llvm/test/Transforms/InstCombine/compare-signs.ll
@@ -47,12 +47,10 @@ define i32 @test3(i32 %a, i32 %b) nounwind readnone {
   ret i32 %t3
 }
 
-; TODO this should optimize but doesn't due to missing vector support in InstCombiner::foldICmpEquality.
 define <2 x i32> @test3vec(<2 x i32> %a, <2 x i32> %b) nounwind readnone {
 ; CHECK-LABEL: @test3vec(
-; CHECK-NEXT:    [[T0:%.*]] = lshr <2 x i32> [[A:%.*]], <i32 31, i32 31>
-; CHECK-NEXT:    [[T1:%.*]] = lshr <2 x i32> [[B:%.*]], <i32 31, i32 31>
-; CHECK-NEXT:    [[T2:%.*]] = icmp eq <2 x i32> [[T0]], [[T1]]
+; CHECK-NEXT:    [[T2_UNSHIFTED:%.*]] = xor <2 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[T2:%.*]] = icmp sgt <2 x i32> [[T2_UNSHIFTED]], <i32 -1, i32 -1>
 ; CHECK-NEXT:    [[T3:%.*]] = zext <2 x i1> [[T2]] to <2 x i32>
 ; CHECK-NEXT:    ret <2 x i32> [[T3]]
 ;
@@ -65,9 +63,8 @@ define <2 x i32> @test3vec(<2 x i32> %a, <2 x i32> %b) nounwind readnone {
 
 define <2 x i32> @test3vec_undef1(<2 x i32> %a, <2 x i32> %b) nounwind readnone {
 ; CHECK-LABEL: @test3vec_undef1(
-; CHECK-NEXT:    [[T0:%.*]] = lshr <2 x i32> [[A:%.*]], <i32 24, i32 undef>
-; CHECK-NEXT:    [[T1:%.*]] = lshr <2 x i32> [[B:%.*]], <i32 24, i32 24>
-; CHECK-NEXT:    [[T2:%.*]] = icmp eq <2 x i32> [[T0]], [[T1]]
+; CHECK-NEXT:    [[T2_UNSHIFTED:%.*]] = xor <2 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[T2:%.*]] = icmp ult <2 x i32> [[T2_UNSHIFTED]], <i32 16777216, i32 16777216>
 ; CHECK-NEXT:    [[T3:%.*]] = zext <2 x i1> [[T2]] to <2 x i32>
 ; CHECK-NEXT:    ret <2 x i32> [[T3]]
 ;
@@ -80,9 +77,8 @@ define <2 x i32> @test3vec_undef1(<2 x i32> %a, <2 x i32> %b) nounwind readnone
 
 define <2 x i32> @test3vec_undef2(<2 x i32> %a, <2 x i32> %b) nounwind readnone {
 ; CHECK-LABEL: @test3vec_undef2(
-; CHECK-NEXT:    [[T0:%.*]] = lshr <2 x i32> [[A:%.*]], <i32 undef, i32 17>
-; CHECK-NEXT:    [[T1:%.*]] = lshr <2 x i32> [[B:%.*]], <i32 undef, i32 17>
-; CHECK-NEXT:    [[T2:%.*]] = icmp eq <2 x i32> [[T0]], [[T1]]
+; CHECK-NEXT:    [[T2_UNSHIFTED:%.*]] = xor <2 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[T2:%.*]] = icmp ult <2 x i32> [[T2_UNSHIFTED]], <i32 131072, i32 131072>
 ; CHECK-NEXT:    [[T3:%.*]] = zext <2 x i1> [[T2]] to <2 x i32>
 ; CHECK-NEXT:    ret <2 x i32> [[T3]]
 ;