Bug 96173 - double to _Decimal64 or _Decimal128 conversion with BID generates 3 MB of code
Summary: double to _Decimal64 or _Decimal128 conversion with BID generates 3 MB of code
Status: RESOLVED WONTFIX
Alias: None
Product: gcc
Classification: Unclassified
Component: libgcc (show other bugs)
Version: 11.0
: P3 normal
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2020-07-11 23:51 UTC by Vincent Lefèvre
Modified: 2020-07-13 21:45 UTC (History)
0 users

See Also:
Host:
Target:
Build:
Known to work:
Known to fail:
Last reconfirmed:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Vincent Lefèvre 2020-07-11 23:51:06 UTC
Consider the following code:

int main (void)
{
  volatile double x = 0.0;
  volatile _Decimal128 i = x;
  return i != 0;
}

On x86_64:

zira:~> gcc-snapshot -O2 tst.c -o tst
zira:~> ll --human-readable tst
-rwxr-xr-x 1 vinc17 vinc17 3.3M 2020-07-12 01:44:05 tst*

With _Decimal64 instead of _Decimal128, tst is a bit smaller: 2.9M

Tested with gcc (Debian 20200616-1) 11.0.0 20200616 (experimental) [master revision beaf12b49ae:aed76232726:b70eeb248efe2b3e9bdb5e26b490e3d8aa07022d]
Comment 1 Andrew Pinski 2020-07-12 00:58:22 UTC
_Decimal128 and _Decimal64 are all software based so there are functions which implement everything.  The main thing which is actually taking up the space is the functions for the conversions.

The functions for the conversions are all stored in one .o file:
 .text          0x00000000004005d0    0x10567 /home/apinski/upstream-gcc/lib/gcc/x86_64-pc-linux-gnu/11.0.0/libgcc.a(bid_binarydecimal.o)
                0x00000000004005d0                __bid32_to_binary32
                0x0000000000400bf0                __bid64_to_binary32
                0x0000000000401460                __bid128_to_binary32
                0x0000000000401f10                __bid32_to_binary64
                0x0000000000402460                __bid64_to_binary64
                0x0000000000402b60                __bid128_to_binary64
                0x0000000000403680                __bid32_to_binary80
                0x0000000000403e30                __bid64_to_binary80
                0x0000000000404870                __bid128_to_binary80
                0x0000000000405640                __bid32_to_binary128
                0x0000000000405c10                __bid64_to_binary128
                0x0000000000406430                __bid128_to_binary128
                0x0000000000406fc0                __binary32_to_bid32
                0x0000000000407650                __binary64_to_bid32
                0x0000000000408100                __binary80_to_bid32
                0x0000000000408be0                __binary128_to_bid32
                0x0000000000409990                __binary32_to_bid64
                0x0000000000409fb0                __binary64_to_bid64
                0x000000000040a730                __binary80_to_bid64
                0x000000000040b280                __binary128_to_bid64
                0x000000000040c040                __binary32_to_bid128
                0x000000000040cfa0                __binary64_to_bid128
                0x000000000040e160                __binary80_to_bid128
                0x000000000040f330                __binary128_to_bid128

Since most of the time when using _Decimal128/_Decimal64, you will be using many of these files, it won't change the size overall to have these functions in different .o file.
Comment 2 Vincent Lefèvre 2020-07-12 02:20:45 UTC
IMHO, the implementation is highly inefficient. Even with all these functions (which are similar, thus should share most code), 3 MB seems a lot to me.

In particular, some user complained that the size of the GNU MPFR library (which now uses such conversions) has been multiplied by 5. This is even worse with the GCC 11 snapshot, using ./configure CC=gcc-snapshot CFLAGS="-O2":

 663880 with --disable-decimal-float
4836016 with --enable-decimal-float
1914376 with --enable-decimal-float and hardcoded values instead of conversions
 668240 with --enable-decimal-float and even more hardcoded values

Note that this is MPFR that does the binary-to-decimal conversion itself (MPFR uses _Decimal128 operations just for the format conversion, to generate either NaN/±Inf/±0 from a double or some regular value from a decimal character sequence). If MPFR can do this conversion within its few hundreds of KB[*], I don't see why this can't be done by GCC.

[*] This does not include the small part of GMP on which MPFR is based, but this includes much unrelated code, for all the functions MPFR implements.
Comment 3 Andrew Pinski 2020-07-12 02:51:42 UTC
The code is all located in libgcc/config/libbid/bid_binarydecimal.c

It looks to be precalulated tables which increase the size.
Comment 4 Andrew Pinski 2020-07-12 02:54:28 UTC
(In reply to Andrew Pinski from comment #3)
> The code is all located in libgcc/config/libbid/bid_binarydecimal.c
> 
> It looks to be precalulated tables which increase the size.

That is the code size is simple, just the tables are huge and go into the readonly section (text section).
Comment 5 joseph@codesourcery.com 2020-07-13 21:45:53 UTC
I expect there's a speed/space trade-off here.  You can use large tables 
for the conversions with less computation, or small tables with more 
computation (and the BID implementation in libgcc uses large tables).

The DPD implementation avoids the whole question of how to convert 
efficiently between decimal and binary FP by doing such conversions via 
strings (which may end up using large tables or less efficient algorithms 
in the libc code used for binary FP / string conversions; if you know the 
source and target formats in advance, there's more scope for statically 
determined bounds on how much internal precision is needed to get 
correctly rounded results for all inputs of the given floating-point 
format).