Re: PATCH: fine-tuning for can_store_by_pieces

[Sandra Loosemore <sandra@codesourcery.com>]

Andrew Pinski wrote:

> It seems if you gave a testcase where this is profitable, it would be
> better to judge this patch (and maybe a testcase for the testsuite
> also).

OK, here's a test case.

void f (char *buf)
{
   __builtin_memcpy (buf, "foo", 3);
}

void g (char *buf)
{
   __builtin_memset (buf, 'f', 3);
}

The MIPS back end has its own custom block move expander, so HAVE_movmemsi is 
true, and the current behavior makes MOVE_RATIO default to 2.  But this only 
handles memcpy, not memset; so with -O2 we end up with f using the custom 
expansion and g using an out-of-line call.

The patch makes STORE_BY_PIECES_P zero so it will always go straight to the 
custom code for memcpy.  The SET_BY_PIECES_P test is now independent of 
MOVE_RATIO and HAVE_movmemsi, so with the patch can_store_by_pieces returns true 
for the memset and it uses the generic expansion instead of a library call.
Before:

g:
         .frame  $sp,0,$31               # vars= 0, regs= 0/0, args= 0, gp= 0
         .mask   0x00000000,0
         .fmask  0x00000000,0
         .set    noreorder
         .set    nomacro

         li      $5,102                  # 0x66
         li      $6,3                    # 0x3
         j       memset
         nop

With the patch:

g:
         .frame  $sp,0,$31               # vars= 0, regs= 0/0, args= 0, gp= 0
         .mask   0x00000000,0
         .fmask  0x00000000,0
         .set    noreorder
         .set    nomacro

         li      $2,102
         sb      $2,2($4)
         sb      $2,0($4)
         j       $31
         sb      $2,1($4)

The code for f, using the custom movmemsi expansion, is unchanged.

-Sandra