X86: Lower a select directly to a setcc_carry if possible.

  int test(unsigned long a, unsigned long b) { return -(a < b); }
compiles to
  _test:                              ## @test
    cmpq  %rsi, %rdi                  ## encoding: [0x48,0x39,0xf7]
    sbbl  %eax, %eax                  ## encoding: [0x19,0xc0]
    ret                               ## encoding: [0xc3]
instead of
  _test:                              ## @test
    xorl  %ecx, %ecx                  ## encoding: [0x31,0xc9]
    cmpq  %rsi, %rdi                  ## encoding: [0x48,0x39,0xf7]
    movl  $-1, %eax                   ## encoding: [0xb8,0xff,0xff,0xff,0xff]
    cmovael %ecx, %eax                ## encoding: [0x0f,0x43,0xc1]
    ret                               ## encoding: [0xc3]

llvm-svn: 122451

2 files changed, 17 insertions, 52 deletions

diff --git a/llvm/lib/Target/X86/README.txt b/llvm/lib/Target/X86/README.txt
index 181c1bd4a9a..ff32d38005a 100644
--- a/llvm/lib/Target/X86/README.txt
+++ b/llvm/lib/Target/X86/README.txt
@@ -1165,58 +1165,6 @@ abs:
 
 //===---------------------------------------------------------------------===//
 
-Consider:
-int test(unsigned long a, unsigned long b) { return -(a < b); }
-
-We currently compile this to:
-
-define i32 @test(i32 %a, i32 %b) nounwind  {
-	%tmp3 = icmp ult i32 %a, %b		; <i1> [#uses=1]
-	%tmp34 = zext i1 %tmp3 to i32		; <i32> [#uses=1]
-	%tmp5 = sub i32 0, %tmp34		; <i32> [#uses=1]
-	ret i32 %tmp5
-}
-
-and
-
-_test:
-	movl	8(%esp), %eax
-	cmpl	%eax, 4(%esp)
-	setb	%al
-	movzbl	%al, %eax
-	negl	%eax
-	ret
-
-Several deficiencies here.  First, we should instcombine zext+neg into sext:
-
-define i32 @test2(i32 %a, i32 %b) nounwind  {
-	%tmp3 = icmp ult i32 %a, %b		; <i1> [#uses=1]
-	%tmp34 = sext i1 %tmp3 to i32		; <i32> [#uses=1]
-	ret i32 %tmp34
-}
-
-However, before we can do that, we have to fix the bad codegen that we get for
-sext from bool:
-
-_test2:
-	movl	8(%esp), %eax
-	cmpl	%eax, 4(%esp)
-	setb	%al
-	movzbl	%al, %eax
-	shll	$31, %eax
-	sarl	$31, %eax
-	ret
-
-This code should be at least as good as the code above.  Once this is fixed, we
-can optimize this specific case even more to:
-
-	movl	8(%esp), %eax
-	xorl	%ecx, %ecx
-        cmpl    %eax, 4(%esp)
-        sbbl    %ecx, %ecx
-
-//===---------------------------------------------------------------------===//
-
 Take the following code (from 
 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16541):
 
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5290453f05a..fca7b02c40a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -7313,6 +7313,23 @@ SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
     Cond = EmitTest(Cond, X86::COND_NE, DAG);
   }
 
+  // a <  b ? -1 :  0 -> RES = ~setcc_carry
+  // a <  b ?  0 : -1 -> RES = setcc_carry
+  // a >= b ? -1 :  0 -> RES = setcc_carry
+  // a >= b ?  0 : -1 -> RES = ~setcc_carry
+  if (Cond.getOpcode() == X86ISD::CMP) {
+    unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
+
+    if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
+        (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
+      SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
+                                DAG.getConstant(X86::COND_B, MVT::i8), Cond);
+      if (isAllOnes(Op1) != (CondCode == X86::COND_B))
+        return DAG.getNOT(DL, Res, Res.getValueType());
+      return Res;
+    }
+  }
+
   // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
   // condition is true.
   SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);