flac dev - Jun 2014 - [PATCH] Re: Bug in FLAC or in GCC or somewhere else?

I noticed that 32-bit flac (from git) compiled with GCC 4.8.3 or 4.9.0
calculates incorrect ReplayGain values. The most common value it produces
is -55.17 dB.

It is possible to avoid this bug by compiling
src/share/replaygain_analysis/replaygain_analysis.c
a) either without -msse2 option
b) or with -O2 instead of -O3
c) another solution is to add -mfpmath=sse option along with -msse2.


For GCC 4.9.0 it's enough to add __attribute__ ((__target__
("arch=i686")))
or __attribute__ ((__target__ ("fpmath=sse"))) to the following
functions
in replaygain_analysis.c: analyzeResult(), GetTitleGain() and GetAlbumGain().

Can anybody confirm this?

lvqcl wrote:
> I noticed that 32-bit flac (from git) compiled with GCC 4.8.3 or 4.9.0
> calculates incorrect ReplayGain values. The most common value it produces
> is -55.17 dB.
It seems that it's a bug in GCC: it cannot correctly calculate the value
of 'elems * (1. - RMS_PERCENTILE)' in the line

     upper = (Int32_t) ceil (elems * (1. - RMS_PERCENTILE));

Since elems * (1. - RMS_PERCENTILE) == elems * (1. - 0.95) =  == elems * 0.05 ==
elems / 20.0   it is possible to replace ceil(...)
with (elems / 20 + (elems % 20) ? 1 : 0).

Patch attached.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: rg_analysis.patch
Type: application/octet-stream
Size: 749 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/flac-dev/attachments/20140601/67c1035e/attachment.obj

lvqcl wrote:
> It seems that it's a bug in GCC: it cannot correctly calculate the
value
> of 'elems * (1. - RMS_PERCENTILE)' in the line
> 
>      upper = (Int32_t) ceil (elems * (1. - RMS_PERCENTILE));
> 
> Since elems * (1. - RMS_PERCENTILE) == elems * (1. - 0.95) =>   == elems
* 0.05 == elems / 20.0   it is possible to replace ceil(...)
> with (elems / 20 + (elems % 20) ? 1 : 0).
> 
> Patch attached.
Patch applied.

Thanks,
Erik
-- 
----------------------------------------------------------------------
Erik de Castro Lopo
http://www.mega-nerd.com/