------- Comment #18 from ubizjak at gmail dot com 2008-02-07 20:47 -------
(In reply to comment #17)
> P2 - this should not block the release (it's not that profiledbootstrap was
> never
> broken in released compilers). It's also hard to analyze (no, I'm not on it,
> volunteers welcome).
It looks that ten_to_ptwo() returns wrong result for argumens >= 7.
Try this debug session:
< non-profiled (correct) cc1>:
(gdb) set args -E -quiet /dev/null
(gdb) break real_to_decimal
Breakpoint 3 at 0x599580: file ../../gcc-svn/trunk/gcc/real.c, line 1453.
(gdb) run
Breakpoint ...
(gdb) p *ten_to_ptwo (7)
$1 = {cl = 1, decimal = 0, sign = 0, signalling = 0, canonical = 0,
uexp = 426, sig = {159020156881263929, 14298703881791668535,
10644899600020376799}}
---
<profiled (wrong) cc1>:
<repeat all of the above commands>
(gdb) p *ten_to_ptwo (7)
$1 = {cl = 1, decimal = 0, sign = 0, signalling = 0, canonical = 0,
uexp = 423, sig = {5021769561911101971, 862341187440057448,
11372220470965069550}}
You can see that the results are different.
Suprisingly, result when passing 6 as an argument differs only for one LSB.
(correct):
(gdb) p *ten_to_ptwo (6)
$2 = {cl = 1, decimal = 0, sign = 0, signalling = 0, canonical = 0,
uexp = 213, sig = {5834422113351499776, 4377335499248575995,
14012984643248170709}}
(wrong):
(gdb) p *ten_to_ptwo (6)
$2 = {cl = 1, decimal = 0, sign = 0, signalling = 0, canonical = 0,
uexp = 213, sig = {5834422113351499776, 4377335499248575995,
14012984643248170708}}
Luckily, ten_to_ptwo() from real.c has only ~20 lines.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33992
