InstCombine: Try harder to combine icmp instructions

consider: (and (icmp X, Y), (and Z, (icmp A, B)))
It may be possible to combine (icmp X, Y) with (icmp A, B).
If we successfully combine, create an 'and' instruction with Z.

This fixes PR20814.

N.B. There is room for improvement after this change but I'm not
convinced it's worth chasing yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216814 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index 2920c88..b19e8b4 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -1101,7 +1101,6 @@ Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
   return nullptr;
 }
 
-
 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
   bool Changed = SimplifyAssociativeOrCommutative(I);
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
@@ -1307,11 +1306,34 @@ Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
       return BinaryOperator::CreateAnd(A, B);
   }
 
-  if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
-    if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
+  {
+    ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
+    ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
+    if (LHS && RHS)
       if (Value *Res = FoldAndOfICmps(LHS, RHS))
         return ReplaceInstUsesWith(I, Res);
 
+    // TODO: Make this recursive; it's a little tricky because an arbitrary
+    // number of 'and' instructions might have to be created.
+    Value *X, *Y;
+    if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
+      if (auto *Cmp = dyn_cast<ICmpInst>(X))
+        if (Value *Res = FoldAndOfICmps(LHS, Cmp))
+          return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
+      if (auto *Cmp = dyn_cast<ICmpInst>(Y))
+        if (Value *Res = FoldAndOfICmps(LHS, Cmp))
+          return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
+    }
+    if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
+      if (auto *Cmp = dyn_cast<ICmpInst>(X))
+        if (Value *Res = FoldAndOfICmps(Cmp, RHS))
+          return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
+      if (auto *Cmp = dyn_cast<ICmpInst>(Y))
+        if (Value *Res = FoldAndOfICmps(Cmp, RHS))
+          return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
+    }
+  }
+
   // If and'ing two fcmp, try combine them into one.
   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))

diff --git a/test/Transforms/InstCombine/and2.ll b/test/Transforms/InstCombine/and2.ll
index e88fd59..c76bb44 100644 (file)
--- a/test/Transforms/InstCombine/and2.ll
+++ b/test/Transforms/InstCombine/and2.ll
@@ -45,7 +45,7 @@ define <4 x i32> @test5(<4 x i32> %A) {
 
 ; Check that we combine "if x!=0 && x!=-1" into "if x+1u>1"
 define i32 @test6(i64 %x) nounwind {
-; CHECK: @test6
+; CHECK-LABEL: @test6(
 ; CHECK-NEXT: add i64 %x, 1
 ; CHECK-NEXT: icmp ugt i64 %x.off, 1
   %cmp1 = icmp ne i64 %x, -1
@@ -54,3 +54,15 @@ define i32 @test6(i64 %x) nounwind {
   %land.ext = zext i1 %.cmp1 to i32
   ret i32 %land.ext
 }
+
+define i1 @test7(i32 %i, i1 %b) {
+; CHECK-LABEL: @test7(
+; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 %i, 0
+; CHECK-NEXT: [[AND:%.*]] = and i1 [[CMP]], %b
+; CHECK-NEXT: ret i1 [[AND]]
+  %cmp1 = icmp slt i32 %i, 1
+  %cmp2 = icmp sgt i32 %i, -1
+  %and1 = and i1 %cmp1, %b
+  %and2 = and i1 %and1, %cmp2
+  ret i1 %and2
+}