(1 - X) * (-2) -> (x - 1) * 2, for all positive nonzero powers of 2

rdar://problem/6501862

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132316 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index a9b60d3..94b619b 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -135,6 +135,23 @@ Instruction *InstCombiner::visitMul(BinaryOperator &I) {
         return BinaryOperator::CreateAdd(Add, Builder->CreateMul(C1, CI));
       }
     }
+
+    // (1 - X) * (-2) -> (x - 1) * 2, for all positive nonzero powers of 2
+    // The "* 2" thus becomes a potential shifting opportunity.
+    {
+      const APInt &   Val = CI->getValue();
+      const APInt &PosVal = Val.abs();
+      if (Val.isNegative() && PosVal.isPowerOf2()) {
+        Value *X = 0;
+        if (match(Op0, m_Sub(m_One(), m_Value(X)))) {
+          // ConstantInt::get(Op0->getType(), 2);
+          Value *Sub = Builder->CreateSub(X, ConstantInt::get(X->getType(), 1),
+                                          "dec1");
+          return BinaryOperator::CreateMul(Sub, ConstantInt::get(X->getType(),
+                                                                 PosVal));
+        }
+      }
+    }
   }
   
   // Simplify mul instructions with a constant RHS.

diff --git a/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll b/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll
new file mode 100644 (file)
index 0000000..e4417ef
--- /dev/null
+++ b/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll
@@ -0,0 +1,24 @@
+; ModuleID = 'test1.c'
+; RUN: opt -S -instcombine < %s | FileCheck %s
+target triple = "x86_64-apple-macosx10.6.6"
+
+define zeroext i16 @foo(i32 %on_off, i16* %puls) nounwind uwtable ssp {
+entry:
+  %on_off.addr = alloca i32, align 4
+  %puls.addr = alloca i16*, align 8
+  %a = alloca i32, align 4
+  store i32 %on_off, i32* %on_off.addr, align 4
+  store i16* %puls, i16** %puls.addr, align 8
+  %tmp = load i32* %on_off.addr, align 4
+; CHECK-NOT: sub
+; CHECK-NOT: mul
+; (1 - %tmp) * (-2) -> (%tmp - 1) * 2
+  %sub = sub i32 1, %tmp
+  %mul = mul i32 %sub, -2
+; CHECK: shl
+; CHECK-NEXT: add
+  store i32 %mul, i32* %a, align 4
+  %tmp1 = load i32* %a, align 4
+  %conv = trunc i32 %tmp1 to i16
+  ret i16 %conv
+}