Date: Sun, 30 Nov 1997 10:57:57 -0800 (PST)
Message-Id: <199711301857.KAA28236@adit.ap.net>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
To: Michael Bukin <M DOT A DOT Bukin AT inp DOT nsk DOT su>, djgpp AT delorie DOT com
From: Nate Eldredge <eldredge AT ap DOT net>
Subject: Re: binutils 2.8.1 performance = slow
Precedence: bulk

At 07:59  11/29/1997 +0600, Michael Bukin wrote:

>  Different versions of binutils produce different instructions
>for .align directive.  E.g. for the following function:
>
>int test (void) {
>  __asm__ __volatile__ (".align 4\n\tmovl $0, %eax\n\t.align 4");
>  return 0;
>}
>
>bnu2.7:
>leal 0x0(%esi),%esi
>leal 0x0(%esi,1),%esi
>
>bnu2.8.1:
>leal 0x0(%esi),%esi
>leal 0x0(%edi,1),%edi
>
>This difference may have effects on instructions pipelining or something.
>I think, speed may depend on instructions near .align.
I would not expect this to be a problem in GCC-compiled code (as opposed to
inline asm). Looking at GCC assembler output, whenever GCC wants to align a
particular piece of code, it writes `.align N,0x90'. This pads with the NOP
instruction, which should be sufficiently fast. This avoids wondering what
the assembler will decide to pad with. If the program which experienced the
slowdown used inline asm, this could explain it. The fix is to explicitly
say to pad with NOPs, like GCC does.


Nate Eldredge
eldredge AT ap DOT net