Added X86 Atom latencies to X86InstrMMX.td.

[oota-llvm.git] / lib / Target / X86 / README.txt

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 51722b2248380506f369e12646880771acb79c9d..6a8a4fdf2520d81a8b1f5e9828f69cc6b2bcd024 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -56,7 +56,7 @@ cmovs, we should expand to a conditional branch like GCC produces.
 
 Some isel ideas:
 
-1. Dynamic programming based approach when compile time if not an
+1. Dynamic programming based approach when compile time is not an
    issue.
 2. Code duplication (addressing mode) during isel.
 3. Other ideas from "Register-Sensitive Selection, Duplication, and
@@ -2060,3 +2060,21 @@ Instead we could generate:
 The trick is to match "fetch_and_add(X, -C) == C".
 
 //===---------------------------------------------------------------------===//
+
+unsigned t(unsigned a, unsigned b) {
+  return a <= b ? 5 : -5;
+}
+
+We generate:
+       movl    $5, %ecx
+       cmpl    %esi, %edi
+       movl    $-5, %eax
+       cmovbel %ecx, %eax
+
+GCC:
+       cmpl    %edi, %esi
+       sbbl    %eax, %eax
+       andl    $-10, %eax
+       addl    $5, %eax
+
+//===---------------------------------------------------------------------===//