[FFmpeg-devel] [PATCH] lavc: avoid rounding errors in float constants

[Andreas Rheinhardt]

Rémi Denis-Courmont:
> Le tiistaina 13. syyskuuta 2022, 18.11.35 EEST Andreas Rheinhardt a écrit :
>> remi at remlab.net:
>>> From: Rémi Denis-Courmont <remi at remlab.net>
>>>
>>> INT_MAX is (typically) a value with 31 significant bits but float can
>>> only represent 23 significant bits, leading to a rounding error.
>>>
>>> This substitutes the actual rounded value to avoid a clang warning:
>>>  warning: implicit conversion from 'int' to 'float' changes value from
>>>  
>>>   2147483647 to 2147483648 [-Wimplicit-const-int-float-conversion]
>>>
>>> ---
>>>
>>>  libavcodec/aaccoder.c | 2 +-
>>>  libavcodec/imc.c      | 2 +-
>>>  2 files changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
>>> index e3b6b2f02c..877558c91c 100644
>>> --- a/libavcodec/aaccoder.c
>>> +++ b/libavcodec/aaccoder.c
>>> @@ -531,7 +531,7 @@ static void search_for_quantizers_anmr(AVCodecContext
>>> *avctx, AACEncContext *s,> 
>>>              int nz = 0;
>>>              
>>>              bandaddr[idx] = w * 16 + g;
>>>
>>> -            qmin = INT_MAX;
>>> +            qmin = -INT_MIN;
>>>
>>>              qmax = 0.0f;
>>>              for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
>>>              
>>>                  FFPsyBand *band =
>>>                  &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
>>>
>>> diff --git a/libavcodec/imc.c b/libavcodec/imc.c
>>> index 92f9980ded..d4dfe3222c 100644
>>> --- a/libavcodec/imc.c
>>> +++ b/libavcodec/imc.c
>>> @@ -917,7 +917,7 @@ static int imc_decode_block(AVCodecContext *avctx,
>>> IMCContext *q, int ch)> 
>>>                                         chctx->flcoeffs1,
>>>                                         chctx->flcoeffs2);
>>>      
>>>      for(i=0; i<BANDS; i++) {
>>>
>>> -        if(chctx->flcoeffs1[i] > INT_MAX) {
>>> +        if(chctx->flcoeffs1[i] > -INT_MIN) {
>>>
>>>              av_log(avctx, AV_LOG_ERROR, "scalefactor out of range\n");
>>>              return AVERROR_INVALIDDATA;
>>>          
>>>          }
>>
>> -INT_MIN can't be represented in an int
> 
> Sure, but that's irrelevant.
> 
>> and therefore -INT_MIN on the right is UB by C11 6.5 (5).
> 
> Of course not. The type of an integer constant is always large enough to fit 
> the value. In this case, it will either be long int or long long int, the 
> later being large enough on any platform.
> 
> See C11 §6.4.4.1.
> 

Incorrect. -INT_MIN is the unary minus operator applied to INT_MIN.
Regardless of how INT_MIN is defined, it is not an integer constant,
because they just don't start with '-'. See the syntax in 6.4.4.1.
Anyway, my limits.h here defines INT_MIN as follows: "#define INT_MIN
(-__INT_MAX__  -1)". As you can see, this is definitely not even close
to an integer constant as defined in 6.4.4.1.

- Andreas