InstCombine: fold (A >> C) == (B >> C) --> (A^B) < (1 << C) for constant Cs.

This is common in bitfield code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194925 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCompares.cpp b/lib/Transforms/InstCombine/InstCombineCompares.cpp
index e624572e041e26c59272512788dbffb3c2c6af1a..226126b39ed733f48cde3d50261817d07638d107 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -2948,6 +2948,24 @@ Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
                             Builder->CreateTrunc(B, A->getType()));
     }
 
+    // (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);
+      if (ShAmt < TypeBits && ShAmt != 0) {
+        ICmpInst::Predicate Pred = I.getPredicate() == 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(Pred, Xor, Builder->getInt(CmpVal));
+      }
+    }
+
     // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
     // "icmp (and X, mask), cst"
     uint64_t ShAmt = 0;

diff --git a/test/Transforms/InstCombine/compare-signs.ll b/test/Transforms/InstCombine/compare-signs.ll
index cdf95abcec506258297caf95264f5c343b261299..62cd5b3f94d56ec924b42fb910dc6ce438389a13 100644 (file)
--- a/test/Transforms/InstCombine/compare-signs.ll
+++ b/test/Transforms/InstCombine/compare-signs.ll
@@ -24,9 +24,9 @@
 define i32 @test3(i32 %a, i32 %b) nounwind readnone {
 ; CHECK-LABEL: @test3(
 entry:
-; CHECK: xor i32 %a, %b
-; CHECK: lshr i32 %0, 31
-; CHECK: xor i32 %1, 1
+; CHECK: [[XOR1:%.*]] = xor i32 %a, %b
+; CHECK: [[SHIFT:%.*]] = lshr i32 [[XOR1]], 31
+; CHECK: [[XOR2:%.*]] = xor i32 [[SHIFT]], 1
         %0 = lshr i32 %a, 31            ; <i32> [#uses=1]
         %1 = lshr i32 %b, 31            ; <i32> [#uses=1]
         %2 = icmp eq i32 %0, %1         ; <i1> [#uses=1]
@@ -34,7 +34,7 @@ entry:
         ret i32 %3
 ; CHECK-NOT: icmp
 ; CHECK-NOT: zext
-; CHECK: ret i32 %2
+; CHECK: ret i32 [[XOR2]]
 }
 
 ; Variation on @test3: checking the 2nd bit in a situation where the 5th bit

diff --git a/test/Transforms/InstCombine/icmp.ll b/test/Transforms/InstCombine/icmp.ll
index bd8128ffb1b0411b51f382067205a22ae21c557a..12a4744cc0feb05825409485d7c860b9b0e9a097 100644 (file)
--- a/test/Transforms/InstCombine/icmp.ll
+++ b/test/Transforms/InstCombine/icmp.ll
@@ -523,12 +523,12 @@ define i1 @test52(i32 %x1) nounwind {
 
 ; PR9838
 ; CHECK-LABEL: @test53(
-; CHECK-NEXT: ashr exact
-; CHECK-NEXT: ashr
+; CHECK-NEXT: sdiv exact
+; CHECK-NEXT: sdiv
 ; CHECK-NEXT: icmp
 define i1 @test53(i32 %a, i32 %b) nounwind {
- %x = ashr exact i32 %a, 30
- %y = ashr i32 %b, 30
+ %x = sdiv exact i32 %a, 30
+ %y = sdiv i32 %b, 30
  %z = icmp eq i32 %x, %y
  ret i1 %z
 }
@@ -1336,3 +1336,23 @@ end:
   %res = phi i1 [%restrue, %true], [%resfalse, %false]
   ret i1 %res
 }
+
+; CHECK-LABEL: @icmp_lshr_lshr_eq
+; CHECK: %z.unshifted = xor i32 %a, %b
+; CHECK: %z = icmp ult i32 %z.unshifted, 1073741824
+define i1 @icmp_lshr_lshr_eq(i32 %a, i32 %b) nounwind {
+ %x = lshr i32 %a, 30
+ %y = lshr i32 %b, 30
+ %z = icmp eq i32 %x, %y
+ ret i1 %z
+}
+
+; CHECK-LABEL: @icmp_ashr_ashr_ne
+; CHECK: %z.unshifted = xor i32 %a, %b
+; CHECK: %z = icmp ugt i32 %z.unshifted, 255
+define i1 @icmp_ashr_ashr_ne(i32 %a, i32 %b) nounwind {
+ %x = ashr i32 %a, 8
+ %y = ashr i32 %b, 8
+ %z = icmp ne i32 %x, %y
+ ret i1 %z
+}