Support specifying the assembly that should be generated

[tibbe]

We want to make sure that native code generator optimizations, like primop unrolling, always fire. We can do this by allowing the programmer to specify what assembly should be generated for a particular snippet of Cmm. The programmer gives a snippet of Cmm:

#include "Cmm.h"

// Large memcpy's should lower to calls.
callMemcpy
{
  W_ dst, src;
  prim %memcpy(dst "ptr", src "ptr", 1024, 4) [];
}

and the expected assembly output:

callMemcpy:
movq  ; Move arguments into place
movq
movl
movl
call memcpy

The expected output shouldn't mention specific register names as these are likely to change.

[tibbe]

I've attached a patch that implements this behavior, together with a first test case.

[nicolast]

This looks somewhat like LLVM's "FileCheck?" system [1]. Might be useful?

Next to that: maybe removing all register names from the asm to be generated could be replaced by using placeholders (like "Variables" in FileCheck?).

[1]  http://llvm.org/docs/TestingGuide.html#FileCheck

[simonmar]

Committed, thanks:

commit a351a67a242f8484764fe450b973b2d4ac9d098e
Author: Johan Tibell <johan.tibell@gmail.com>
Date:   Tue May 31 11:56:58 2011 +0200

    Make it possible to test the generated assembly
    
    This test framework feature is inspired by a similar feature in LLVM.
    The programmer writes a bit of Cmm
    
        #include "Cmm.h"
    
        // Large memcpy's should lower to calls.
        callMemcpy
        {
          W_ dst, src;
          prim %memcpy(dst "ptr", src "ptr", 1024, 4) [];
        }
    
    and asserts what the generated assembly should look like, modulo
    register naming.
    
        callMemcpy:
        movq  ; Move arguments into place
        movq
        movl
        movl
        call memcpy
    
    Patch edited and updated by Simon Marlow, and I also added a test for
    unrolling memcpy and a simple constant-propagation test.