[llvm-commits] Potential Patch for PR1095

[Chris Lattner]

On Jan 9, 2007, at 12:25 AM, Bill Wendling wrote:
> Hi all,

Hey Bill, sorry for the delay.

> I would like someone to review this patch for PR1095. This gives us  
> a "dialect" for the assembler code if we're given something like:
>
> 	__asm__("{cntlz|cntlzw} ...
>
> Before, we were defaulting to the first one. But there are some  
> platforms which need the second form.

This patch is very close.  Comments below:

Please add a .ll file regression test to llvm/test/Regression/CodeGen/ 
PowerPC.


--- include/llvm/Target/TargetAsmInfo.h	21 Dec 2006 21:24:35 -0000	1.18
+++ include/llvm/Target/TargetAsmInfo.h	9 Jan 2007 08:21:25 -0000
@@ -27,6 +27,11 @@
    /// properties and features specific to the target.
    class TargetAsmInfo {
    protected:
+    enum AsmDialect {
+      ASM_ATT,
+      ASM_INTEL
+    };
+

The meaning of the various dialects vary across the targets.  This  
enum should move into the X86 backend.  The PowerPC backend should  
use PPC_OLD_MNEMONICS = 0 and PPC_NEW_MNEMONICS = 1.  Darwin should  
use new mnemonics of course.


@@ -128,6 +133,11 @@
      /// Otherwise, it emits ".align log2(N)", e.g. 3 to align to an  
8 byte
      /// boundary.
      bool AlignmentIsInBytes;              // Defaults to true
+
+    //===--- Assembler Dialect Information  
--------------------------------===//
+
+    /// AssemblerDialect - Which dialect of an assembler variant to  
use.
+    AsmDialect AssemblerDialect;         // Defaults to ASM_ATT

Please make this just an 'unsigned' and group it with InlineAsmStart/ 
InlineAsmEnd.  You could make the group comment be "Inline Asm  
Directives" or something.

Otherwise, the patch looks great.  Please make these changes and  
apply it.

Thanks Bill!

-Chris