Re: Two observations using GDB 13 snapshot

[Torbjorn SVENSSON <torbjorn.svensson@foss.st.com>]

On 2023-01-04 19:10, Eli Zaretskii via Gdb-patches wrote:
>> Date: Tue, 3 Jan 2023 16:34:53 -0500
>> Cc: tom@tromey.com, gdb-patches@sourceware.org, luis.machado@arm.com
>> From: Simon Marchi <simark@simark.ca>
>>
>>> In terms of the code, may be worth trying TOLOWER from
>>> include/safe-ctype.h instead of tolower()
>>
>> The tolower call is inside strcasecmp, we don't call tolower directly:
>>
>> #0  0x77c348d5 in msvcrt!__crtLCMapStringA ()
>>     from C:\WINDOWS\system32\msvcrt.dll
>> #1  0x77c348cd in msvcrt!__crtLCMapStringA ()
>>     from C:\WINDOWS\system32\msvcrt.dll
>> #2  0x77c30045 in wmktemp () from C:\WINDOWS\system32\msvcrt.dll
>> #3  0x77c1c992 in tolower () from C:\WINDOWS\system32\msvcrt.dll
>> #4  0x77c462a1 in stricmp () from C:\WINDOWS\system32\msvcrt.dll
>> #5  0x005107d3 in strcasecmp (__s2=<optimized out>, __s1=<optimized out>)
>>      at d:/usr/include/strings.h:92
>> #6  cooked_index_entry::operator< (this=<optimized out>, other=...)
>>      at ./dwarf2/cooked-index.h:150
>>
>> It would be interesting to change that strcasecmp call to strcmp, just
>> to see if it makes an impact on the performance.  Whether or not that
>> would be correct is another thing, but it would help see if that
>> strcasecmp / tolower call is really at fault here.
> 
> Looks like indeed strcasecmp is the culprit.  With the patch below,
> which replaces strcasecmp with a simple case-insensitive comparison
> that only works with ASCII, the phase of reading symbols from gdb.exe
> goes down to just 6 seconds, which is basically the same time as with
> GDB 12.
> 
> --- gdb/dwarf2/cooked-index.h~0	2022-12-17 03:47:12.000000000 +0200
> +++ gdb/dwarf2/cooked-index.h	2023-01-04 20:00:04.052250000 +0200
> @@ -35,6 +35,8 @@
>   #include "dwarf2/tag.h"
>   #include "gdbsupport/range-chain.h"
>   
> +#define my_tolower(c) (('A' <= (c) && (c) <= 'Z') ? ((c) - 'A' + 'a') : (c))
> +
>   struct dwarf2_per_cu_data;
>   
>   /* Flags that describe an entry in the index.  */
> @@ -147,7 +149,20 @@ struct cooked_index_entry : public alloc
>        entries.  */
>     bool operator< (const cooked_index_entry &other) const
>     {
> +#if 0
>       return strcasecmp (canonical, other.canonical) < 0;
> +#else
> +    const unsigned char *s1 = (unsigned char *)canonical, *s2 = (unsigned char *)other.canonical;
> +
> +    while (my_tolower(*s1) == my_tolower(*s2))
> +      {
> +	if (*s1 == 0)
> +	  return false;

Without any knowledge of the cooked_index_entry type or what it's 
supposed to do, I think the return statement here is problematic in the 
case where there are 2 cooked_index_entry instances contains the same 
case insensitive canonical strings. I'm not sure if this can happen, but 
it's just a thought.
It also appears that the old strcasecmp implementation would suffer the 
same limitation. To avoid the issue, I guess the address of the 2 
objects could be compared in order to get a stable result if there is no 
other property that is guaranteed to be unique for the 2 instances.

Kind regards,
Torbjörn

> +	s1++;
> +	s2++;
> +      }
> +    return (int)my_tolower(*s1) < (int)my_tolower(*s2);
> +#endif
>     }
>   
>     /* The name as it appears in DWARF.  This always points into one of