fix rdar://9289583 - fast isel should handle non-canonical commutative binops

allowing us to fold the immediate into the 'and' in this case:

int test1(int i) {
  return 8&i;
}

llvm-svn: 129653

2 files changed, 37 insertions, 4 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp b/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
index aab23ea1c77..df75c4a88f7 100644
--- a/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -122,10 +122,9 @@ unsigned FastISel::getRegForValue(const Value *V) {
   // only locally. This is because Instructions already have the SSA
   // def-dominates-use requirement enforced.
   DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
-  if (I != FuncInfo.ValueMap.end()) {
-    unsigned Reg = I->second;
-    return Reg;
-  }
+  if (I != FuncInfo.ValueMap.end())
+    return I->second;
+
   unsigned Reg = LocalValueMap[V];
   if (Reg != 0)
     return Reg;
@@ -331,6 +330,26 @@ bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
       return false;
   }
 
+  // Check if the first operand is a constant, and handle it as "ri".  At -O0,
+  // we don't have anything that canonicalizes operand order.
+  if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0)))
+    if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) {
+      unsigned Op1 = getRegForValue(I->getOperand(1));
+      if (Op1 == 0) return false;
+
+      bool Op1IsKill = hasTrivialKill(I->getOperand(1));
+      
+      unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
+                                       ISDOpcode, Op1, Op1IsKill,
+                                       CI->getZExtValue());
+      if (ResultReg != 0) {
+        // We successfully emitted code for the given LLVM Instruction.
+        UpdateValueMap(I, ResultReg);
+        return true;
+      }
+    }
+  
+  
   unsigned Op0 = getRegForValue(I->getOperand(0));
   if (Op0 == 0)
     // Unhandled operand. Halt "fast" selection and bail.
diff --git a/llvm/test/CodeGen/X86/fast-isel-x86-64.ll b/llvm/test/CodeGen/X86/fast-isel-x86-64.ll
new file mode 100644
index 00000000000..b2d1263ca77
--- /dev/null
+++ b/llvm/test/CodeGen/X86/fast-isel-x86-64.ll
@@ -0,0 +1,14 @@
+; RUN: llc < %s | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-darwin10.0.0"
+
+; Make sure that fast-isel folds the immediate into the binop even though it
+; is non-canonical.
+define i32 @test1(i32 %i) nounwind ssp {
+  %and = and i32 8, %i
+  ret i32 %and
+}
+
+; CHECK: test1:
+; CHECK: andl	$8,