[llvm] r303860 - [InstCombine] make icmp-mul fold more efficient

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Thu May 25 09:13:57 2017
New Revision: 303860

URL: http://llvm.org/viewvc/llvm-project?rev=303860&view=rev
Log:
[InstCombine] make icmp-mul fold more efficient

There's probably a lot more like this (see also comments in D33338 about responsibility), 
but I suspect we don't usually get a visible manifestation.

Given the recent interest in improving InstCombine efficiency, another potential micro-opt
that could be repeated several times in this function: morph the existing icmp pred/operands
instead of creating a new instruction.

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
    llvm/trunk/test/Transforms/InstCombine/icmp.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp?rev=303860&r1=303859&r2=303860&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp Thu May 25 09:13:57 2017
@@ -3057,19 +3057,21 @@ Instruction *InstCombiner::foldICmpBinOp
         break;
 
       const APInt *C;
-      if (match(BO0->getOperand(1), m_APInt(C))) {
+      if (match(BO0->getOperand(1), m_APInt(C)) && *C != 0 && *C != 1) {
         // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
         // Mask = -1 >> count-trailing-zeros(C).
-        if (*C != 0 && *C != 1) {
-          // FIXME: If trailing zeros is 0, don't bother creating Mask.
+        if (unsigned TZs = C->countTrailingZeros()) {
           Constant *Mask = ConstantInt::get(
               BO0->getType(),
-              APInt::getLowBitsSet(C->getBitWidth(),
-                                   C->getBitWidth() - C->countTrailingZeros()));
+              APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
           Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
           Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
           return new ICmpInst(Pred, And1, And2);
         }
+        // If there are no trailing zeros in the multiplier, just eliminate
+        // the multiplies (no masking is needed):
+        // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
+        return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
       }
       break;
     }

Modified: llvm/trunk/test/Transforms/InstCombine/icmp.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/icmp.ll?rev=303860&r1=303859&r2=303860&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/icmp.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/icmp.ll Thu May 25 09:13:57 2017
@@ -2920,7 +2920,7 @@ define i1 @eq_mul_constants(i32 %x, i32
 
 define <2 x i1> @eq_mul_constants_splat(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @eq_mul_constants_splat(
-; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i32> %y, %x
+; CHECK-NEXT:    [[C:%.*]] = icmp ne <2 x i32> %x, %y
 ; CHECK-NEXT:    ret <2 x i1> [[C]]
 ;
   %A = mul <2 x i32> %x, <i32 5, i32 5>