Reapply 132348 with fixes. rdar://problem/6501862

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132402 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 94b619b2037b2639b716cec9b7e59656aefe8cd5..2d29403097ce79384f76d37f092ab04bce233e49 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -136,19 +136,25 @@ Instruction *InstCombiner::visitMul(BinaryOperator &I) {
       }
     }
 
-    // (1 - X) * (-2) -> (x - 1) * 2, for all positive nonzero powers of 2
-    // The "* 2" thus becomes a potential shifting opportunity.
+    // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n
+    // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n
+    // The "* (2**n)" 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));
+        Value *X = 0, *Y = 0;
+        if (Op0->hasOneUse()) {
+          ConstantInt *C1;
+          Value *Sub = 0;
+          if (match(Op0, m_Sub(m_Value(Y), m_Value(X))))
+            Sub = Builder->CreateSub(X, Y, "suba");
+          else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1))))
+            Sub = Builder->CreateSub(Builder->CreateNeg(C1), Y, "subc");
+          if (Sub)
+            return
+              BinaryOperator::CreateMul(Sub,
+                                        ConstantInt::get(Y->getType(), PosVal));
         }
       }
     }

diff --git a/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll b/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll
index e4417ef96a049b116225e0d78e627e874ec750ca..b096d1f952f9dd6c75a068defe13bb17b3d1689a 100644 (file)
--- a/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll
+++ b/test/Transforms/InstCombine/2011-05-28-swapmulsub.ll
@@ -2,18 +2,14 @@
 ; 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 {
+define zeroext i16 @foo1(i32 %on_off) 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
+; CHECK-NOT: mul i32
   %mul = mul i32 %sub, -2
 ; CHECK: shl
 ; CHECK-NEXT: add
@@ -22,3 +18,40 @@ entry:
   %conv = trunc i32 %tmp1 to i16
   ret i16 %conv
 }
+
+define zeroext i16 @foo2(i32 %on_off, i32 %q) nounwind uwtable ssp {
+entry:
+  %on_off.addr = alloca i32, align 4
+  %q.addr = alloca i32, align 4
+  %a = alloca i32, align 4
+  store i32 %on_off, i32* %on_off.addr, align 4
+  store i32 %q, i32* %q.addr, align 4
+  %tmp = load i32* %q.addr, align 4
+  %tmp1 = load i32* %on_off.addr, align 4
+  %sub = sub i32 %tmp, %tmp1
+; CHECK-NOT: mul i32
+  %mul = mul i32 %sub, -4
+; CHECK: sub i32
+; CHECK-NEXT: shl
+  store i32 %mul, i32* %a, align 4
+  %tmp2 = load i32* %a, align 4
+  %conv = trunc i32 %tmp2 to i16
+  ret i16 %conv
+}
+
+define zeroext i16 @foo3(i32 %on_off) nounwind uwtable ssp {
+entry:
+  %on_off.addr = alloca i32, align 4
+  %a = alloca i32, align 4
+  store i32 %on_off, i32* %on_off.addr, align 4
+  %tmp = load i32* %on_off.addr, align 4
+  %sub = sub i32 7, %tmp
+; CHECK-NOT: mul i32
+  %mul = mul i32 %sub, -4
+; 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
+}