[PATCH] D127398: [InstCombine] add vector support for (A >> C) == (B >> C) --> (A^B) u< (1 << C)

[Chenbing.Zheng via Phabricator via llvm-commits]

Chenbing.Zheng created this revision.
Chenbing.Zheng added reviewers: spatel, RKSimon, benshi001.
Chenbing.Zheng added a project: LLVM.
Herald added a subscriber: hiraditya.
Herald added a project: All.
Chenbing.Zheng requested review of this revision.
Herald added subscribers: llvm-commits, jacquesguan.

Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D127398

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


Index: llvm/test/Transforms/InstCombine/compare-signs.ll
===================================================================
--- llvm/test/Transforms/InstCombine/compare-signs.ll
+++ llvm/test/Transforms/InstCombine/compare-signs.ll
@@ -47,17 +47,32 @@
   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:    [[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]]
+;
+  %t0 = lshr <2 x i32> %a, <i32 31, i32 31>
+  %t1 = lshr <2 x i32> %b, <i32 31, i32 31>
+  %t2 = icmp eq <2 x i32> %t0, %t1
+  %t3 = zext <2 x i1> %t2 to <2 x i32>
+  ret <2 x i32> %t3
+}
+
+; negative test
+
+define <2 x i32> @test4vec(<2 x i32> %a, <2 x i32> %b) nounwind readnone {
+; CHECK-LABEL: @test4vec(
 ; 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:    [[T1:%.*]] = lshr <2 x i32> [[B:%.*]], <i32 30, i32 30>
 ; CHECK-NEXT:    [[T2:%.*]] = icmp eq <2 x i32> [[T0]], [[T1]]
 ; CHECK-NEXT:    [[T3:%.*]] = zext <2 x i1> [[T2]] to <2 x i32>
 ; CHECK-NEXT:    ret <2 x i32> [[T3]]
 ;
   %t0 = lshr <2 x i32> %a, <i32 31, i32 31>
-  %t1 = lshr <2 x i32> %b, <i32 31, i32 31>
+  %t1 = lshr <2 x i32> %b, <i32 30, i32 30>
   %t2 = icmp eq <2 x i32> %t0, %t1
   %t3 = zext <2 x i1> %t2 to <2 x i32>
   ret <2 x i32> %t3
Index: llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -4607,12 +4607,15 @@
 
   // (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_APInt(AP1)))) &&
+       match(Op1, m_OneUse(m_LShr(m_Value(B), m_APInt(AP2))))) ||
+      (match(Op0, m_OneUse(m_AShr(m_Value(A), m_APInt(AP1)))) &&
+       match(Op1, m_OneUse(m_AShr(m_Value(B), m_APInt(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;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D127398.435503.patch
Type: text/x-patch
Size: 3041 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220609/77ebfeb2/attachment.bin>