>> ! float4 tmpsml = cnt_sml(qtrg, key, *recheck);
>>
>> /* strange bug at freebsd 5.2.1 and gcc 3.3.3 */
>> ! res = (*(int *) &tmpsml == *(int *) &nlimit || tmpsml > nlimit) ? true : false;
>
> What's the compiler bug about? This code and comment were introduced in
> cbfa4092bb (May 2004) without any explanation. Do we still need to keep
> it, if &nlimit is now a local variable instead of a global? FWIW the
> oldest GCC in the buildfarm is 3.4.2/3.4.3 (except for Gaur which uses
As I remeber, the problem was with x87 math coprocessor. Compiler (suppose,
modern compiler could do it too) keeps tmpsml in internal 80-bit wide register
and compares with 32-bit wide float. Of course, it depends on level of
optimization and so, result of comparison was differ in optimization enabled and
disabled instances. Such strange way I choose to force compiler to use main
memory for tmpsml variable. Actually, I don't know better way even now.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/