Add a target-specific branchless method for double-width relational

comparisons on x86.  Essentially, the way this works is that SUB+SBB sets
the relevant flags the same way a double-width CMP would.

This is a substantial improvement over the generic lowering in LLVM. The output
is also shorter than the gcc-generated output; I haven't done any detailed
benchmarking, though.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@127852 91177308-0d34-0410-b5e6-96231b3b80d8

test/CodeGen/X86/long-setcc.ll

@@ -1,18 +1,35 @@
-; RUN: llc < %s -march=x86 | grep cmp | count 1
-; RUN: llc < %s -march=x86 | grep shr | count 1
-; RUN: llc < %s -march=x86 | grep xor | count 1
+; RUN: llc < %s -march=x86 | FileCheck %s
 
-define i1 @t1(i64 %x) nounwind {
-	%B = icmp slt i64 %x, 0
-	ret i1 %B
+; General case
+define i1 @t1(i64 %x, i64 %y) nounwind {
+; CHECK: @t1
+; CHECK: subl
+; CHECK: sbbl
+; CHECK: setl %al
+  %B = icmp slt i64 %x, %y
+  ret i1 %B
 }
 
+; Some special cases
 define i1 @t2(i64 %x) nounwind {
-	%tmp = icmp ult i64 %x, 4294967296
-	ret i1 %tmp
+; CHECK: @t2
+; CHECK: shrl $31, %eax
+  %B = icmp slt i64 %x, 0
+  ret i1 %B
 }
 
-define i1 @t3(i32 %x) nounwind {
-	%tmp = icmp ugt i32 %x, -1
-	ret i1 %tmp
+define i1 @t3(i64 %x) nounwind {
+; CHECK: @t3
+; CHECX: cmpl $0
+; CHECX: sete %al
+  %tmp = icmp ult i64 %x, 4294967296
+  ret i1 %tmp
+}
+
+define i1 @t4(i64 %x) nounwind {
+; CHECK: @t4
+; CHECX: cmpl $0
+; CHECX: setne %al
+  %tmp = icmp ugt i64 %x, 4294967295
+  ret i1 %tmp
 }