Teach instcombine to transform

(X >s -1) ? C1 : C2 and (X <s  0) ? C2 : C1
into ((X >>s 31) & (C2 - C1)) + C1, avoiding the conditional.

This optimization could be extended to take non-const C1 and C2 but we better
stay conservative to avoid code size bloat for now.

for
int sel(int n) {
     return n >= 0 ? 60 : 100;
}

we now generate
  sarl  $31, %edi
  andl  $40, %edi
  leal  60(%rdi), %eax

instead of
  testl %edi, %edi
  movl  $60, %ecx
  movl  $100, %eax
  cmovnsl %ecx, %eax

llvm-svn: 107866

2 files changed, 31 insertions, 11 deletions

diff --git a/llvm/lib/Target/README.txt b/llvm/lib/Target/README.txt
index 90afe3640ce..4d7ee08de1d 100644
--- a/llvm/lib/Target/README.txt
+++ b/llvm/lib/Target/README.txt
@@ -906,17 +906,6 @@ The expression should optimize to something like
 
 //===---------------------------------------------------------------------===//
 
-From GCC Bug 3756:
-int
-pn (int n)
-{
- return (n >= 0 ? 1 : -1);
-}
-Should combine to (n >> 31) | 1.  Currently not optimized with "clang
--emit-llvm-bc | opt -std-compile-opts | llc".
-
-//===---------------------------------------------------------------------===//
-
 void a(int variable)
 {
  if (variable == 4 || variable == 6)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index c958cde0917..c44fe9db6e3 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -329,6 +329,37 @@ Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
       }
     }
 
+  // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1
+  // and       (X <s  0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1
+  // FIXME: Type and constness constraints could be lifted, but we have to
+  //        watch code size carefully. We should consider xor instead of
+  //        sub/add when we decide to do that.
+  if (const IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) {
+    if (TrueVal->getType() == Ty) {
+      if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) {
+        ConstantInt *C1 = NULL, *C2 = NULL;
+        if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) {
+          C1 = dyn_cast<ConstantInt>(TrueVal);
+          C2 = dyn_cast<ConstantInt>(FalseVal);
+        } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) {
+          C1 = dyn_cast<ConstantInt>(FalseVal);
+          C2 = dyn_cast<ConstantInt>(TrueVal);
+        }
+        if (C1 && C2) {
+          // This shift results in either -1 or 0.
+          Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1);
+
+          // Check if we can express the operation with a single or.
+          if (C2->isAllOnesValue())
+            return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1));
+
+          Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue());
+          return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1));
+        }
+      }
+    }
+  }
+
   if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
     // Transform (X == Y) ? X : Y  -> Y
     if (Pred == ICmpInst::ICMP_EQ)