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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@107866 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/README.txt b/lib/Target/README.txt
index 90afe3640ce7d961332861fdaff55de1e4c7639b..4d7ee08de1dee3dfc93d4d3e95acfa47bd434f1f 100644 (file)
--- a/lib/Target/README.txt
+++ b/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)
 void a(int variable)
 {
  if (variable == 4 || variable == 6)

diff --git a/lib/Transforms/InstCombine/InstCombineSelect.cpp b/lib/Transforms/InstCombine/InstCombineSelect.cpp
index c958cde0917b5b0f720f14f61121d4d62e13b5be..c44fe9db6e3a7689883f6711af0aca3dab80014f 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/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)
   if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
     // Transform (X == Y) ? X : Y  -> Y
     if (Pred == ICmpInst::ICMP_EQ)

diff --git a/test/Transforms/InstCombine/select.ll b/test/Transforms/InstCombine/select.ll
index 06d5338884ec870fc384dd059e07235fce320448..246a7bc59783eee12d2e830c67c5dec344d9e081 100644 (file)
--- a/test/Transforms/InstCombine/select.ll
+++ b/test/Transforms/InstCombine/select.ll
@@ -438,3 +438,35 @@ define i32 @test34(i32 %x, i32 %y) {
 ; CHECK: @test34
 ; CHECK: ret i32 %x
 }
 ; CHECK: @test34
 ; CHECK: ret i32 %x
 }

+
+define i32 @test35(i32 %x) {
+  %cmp = icmp sge i32 %x, 0
+  %cond = select i1 %cmp, i32 60, i32 100
+  ret i32 %cond
+; CHECK: @test35
+; CHECK: ashr i32 %x, 31
+; CHECK: and i32 {{.*}}, 40
+; CHECK: add i32 {{.*}}, 60
+; CHECK: ret
+}
+
+define i32 @test36(i32 %x) {
+  %cmp = icmp slt i32 %x, 0
+  %cond = select i1 %cmp, i32 60, i32 100
+  ret i32 %cond
+; CHECK: @test36
+; CHECK: ashr i32 %x, 31
+; CHECK: and i32 {{.*}}, -40
+; CHECK: add i32 {{.*}}, 100
+; CHECK: ret
+}
+
+define i32 @test37(i32 %x) {
+  %cmp = icmp sgt i32 %x, -1
+  %cond = select i1 %cmp, i32 1, i32 -1
+  ret i32 %cond
+; CHECK: @test37
+; CHECK: ashr i32 %x, 31
+; CHECK: or i32 {{.*}}, 1
+; CHECK: ret
+}