Fix some simple copy-paste errors in MBlaze ASM Parser and Makefile.

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

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 07722f44ec39dcd96b06798829290af14f63806b..8d9eabad0be0a046ca1a00827bbc273ac96ef314 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -7,14 +7,6 @@ copy (3-addr bswap + memory support?)  This is available on Atom processors.
 
 //===---------------------------------------------------------------------===//
 
-CodeGen/X86/lea-3.ll:test3 should be a single LEA, not a shift/move.  The X86
-backend knows how to three-addressify this shift, but it appears the register
-allocator isn't even asking it to do so in this case.  We should investigate
-why this isn't happening, it could have significant impact on other important
-cases for X86 as well.
-
-//===---------------------------------------------------------------------===//
-
 This should be one DIV/IDIV instruction, not a libcall:
 
 unsigned test(unsigned long long X, unsigned Y) {
@@ -1572,7 +1564,7 @@ Implement processor-specific optimizations for parity with GCC on these
 processors.  GCC does two optimizations:
 
 1. ix86_pad_returns inserts a noop before ret instructions if immediately
-   preceeded by a conditional branch or is the target of a jump.
+   preceded by a conditional branch or is the target of a jump.
 2. ix86_avoid_jump_misspredicts inserts noops in cases where a 16-byte block of
    code contains more than 3 branches.
    
@@ -1735,26 +1727,6 @@ are functionally identical.
 
 //===---------------------------------------------------------------------===//
 
-Take the following C code:
-int x(int y) { return (y & 63) << 14; }
-
-Code produced by gcc:
-       andl    $63, %edi
-       sall    $14, %edi
-       movl    %edi, %eax
-       ret
-
-Code produced by clang:
-       shll    $14, %edi
-       movl    %edi, %eax
-       andl    $1032192, %eax
-       ret
-
-The code produced by gcc is 3 bytes shorter.  This sort of construct often
-shows up with bitfields.
-
-//===---------------------------------------------------------------------===//
-
 Take the following C code:
 int f(int a, int b) { return (unsigned char)a == (unsigned char)b; }
 
@@ -2040,3 +2012,88 @@ clamp_float:                            # @clamp_float
 with -ffast-math.
 
 //===---------------------------------------------------------------------===//
+
+This function (from PR9803):
+
+int clamp2(int a) {
+        if (a > 5)
+                a = 5;
+        if (a < 0) 
+                return 0;
+        return a;
+}
+
+Compiles to:
+
+_clamp2:                                ## @clamp2
+        pushq   %rbp
+        movq    %rsp, %rbp
+        cmpl    $5, %edi
+        movl    $5, %ecx
+        cmovlel %edi, %ecx
+        testl   %ecx, %ecx
+        movl    $0, %eax
+        cmovnsl %ecx, %eax
+        popq    %rbp
+        ret
+
+The move of 0 could be scheduled above the test to make it is xor reg,reg.
+
+//===---------------------------------------------------------------------===//
+
+GCC PR48986.  We currently compile this:
+
+void bar(void);
+void yyy(int* p) {
+    if (__sync_fetch_and_add(p, -1) == 1)
+      bar();
+}
+
+into:
+       movl    $-1, %eax
+       lock
+       xaddl   %eax, (%rdi)
+       cmpl    $1, %eax
+       je      LBB0_2
+
+Instead we could generate:
+
+       lock
+       dec %rdi
+       je LBB0_2
+
+The trick is to match "fetch_and_add(X, -C) == C".
+
+//===---------------------------------------------------------------------===//
+
+unsigned log2(unsigned x) {
+  return x > 1 ? 32-__builtin_clz(x-1) : 0;
+}
+
+generates (x86_64):
+       xorl    %eax, %eax
+       cmpl    $2, %edi
+       jb      LBB0_2
+## BB#1:
+       decl    %edi
+       movl    $63, %ecx
+       bsrl    %edi, %eax
+       cmovel  %ecx, %eax
+       xorl    $-32, %eax
+       addl    $33, %eax
+LBB0_2:
+       ret
+
+The cmov and the early test are redundant:
+       xorl    %eax, %eax
+       cmpl    $2, %edi
+       jb      LBB0_2
+## BB#1:
+       decl    %edi
+       bsrl    %edi, %eax
+       xorl    $-32, %eax
+       addl    $33, %eax
+LBB0_2:
+       ret
+
+//===---------------------------------------------------------------------===//