X-Authentication-Warning: delorie.com: mailnull set sender to djgpp-bounces using -f
From: pavenis AT lanet DOT lv
To: djgpp AT delorie DOT com
Date: Thu, 6 Dec 2001 18:16:35 +0200
MIME-Version: 1.0
Subject: Re: Is it a gcc bug?
Message-ID: <3C0FB603.18944.1396CC5@localhost>
In-reply-to: <9uo3dc$i1i$1@nntp9.atl.mindspring.net>
X-mailer: Pegasus Mail for Windows (v4.01)
Content-type: text/plain; charset=US-ASCII
Content-transfer-encoding: 7BIT
Content-description: Mail message body
Reply-To: djgpp AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

On 6 Dec 2001 at 10:34, Marp wrote:

> I think I might have run into an optimizer bug in GCC 3.0.2, but I need some
> insight. I have the version dated Nov 21, 2001 on the ftp server.
> 
> Consider the following program:
> 
> #include <stdio.h>
> #include <dos.h>
> 
> int main(void) {
>   union {
>     unsigned long long full;
>     struct {
>       unsigned long lo, hi;
>     } half;
>   } before, after;
> 
>   asm("rdtsc" : "=a" (before.half.lo), "=d" (before.half.hi));
>   delay(1000); // pause for 1 second
>   asm("rdtsc" : "=a" (after.half.lo), "=d" (after.half.hi));
> 
>   printf("CPU Speed: %llu MHz\n", (after.full - before.full) / 1000000llu);
> 
>   return 0;
> }
> 
> If I compile with no optimization (gcc foo.c -o foo.exe -s) the program
> outputs the correct result (400). However, if I turn on the optimizer (gcc
> foo.c -o foo.exe -s -O1) the program outputs an incorrect result (0). When I
> inspect the assembly output of the compiler, the one difference that caught
> my attention is that the second asm statement appears to be missing when the
> optimizer is turned on. Now I'm not exactly an expert with assembly, but I
> don't see how a correct result could be computed without that second asm
> statement.
> 
> So can someone explain to me what is wrong? Is it a gcc bug, or my mistake?

GCC tries to be too wise and optimizes one rdtsc out. Use 
'asm volatile ("rdtsc" : ... )' to force it not do wrong optimization.
Anyway I would suggest a different aproach fo measuring clock frequency
like reading rdtsc from timer interrupt procedure (or RTC interrupt procedure), 
using delay together with  uclock() or something similar. 

Andris