[PATCH] Fix bug on soft emulation of float point in gcc/longlong.c
There are 2 bugs existing in __udiv_qrnnd_c:
1. remainder could not be counted correctly
__r1 and __r0 are not adding enough, so I use do..while to replace former if
2. the case of overflow of __m are not considered
so I add a couple of lines to handle it
I found the bugs from Linux kernel,
for PowerPC math-emu use __udiv_qrnnd_c to emulate float div.
I sent the patch to the maintainers of kernel, but they replied to me that
this code came from gcc and should better keep coincident.
So I think I should send patch to here.
I don't know how to trigger __udiv_qrnnd_c in gcc.
but if the argument of __udiv_qrnnd_c are like
0x07f8 0x07f8 0x00210fff 0xc000 0x07f8,
the bug will be reproduced.
Also, you can try the program below, I have test it on Linux kernel math-emu
but haven't test on gcc.
---
#include
union fu {
unsigned int u;
float f;
};
union fu fdiv(union fu a, union fu b, union fu expected)
{
union fu z;
z.f = a.f / b.f;
printf("fdiv %08x / %08x = %08x expected %08x %s\n",
a.u, b.u, z.u, expected.u,
(z.u == expected.u) ? "(PASS)" : "(FAIL)");
printf(" %e / %e = %e\n", a.f, b.f, z.f);
}
int main()
{
union fu fa,fb,fe;
fa.u = 0xc0843fff;
fb.u = 0x80ff;
fe.u = 0x7f044000;
fdiv(fa,fb,fe);
}
---
Signed-off-by: Liu Yu <[EMAIL PROTECTED]>
---
gcc/longlong.h | 18 --
1 files changed, 12 insertions(+), 6 deletions(-)
diff --git a/gcc/longlong.h b/gcc/longlong.h
index 6c24564..256399b 100644
--- a/gcc/longlong.h
+++ b/gcc/longlong.h
@@ -1368,12 +1368,15 @@ UDItype __umulsidi3 (USItype, USItype);
__q1 = (n1) / __d1;
\
__m = (UWtype) __q1 * __d0;
\
__r1 = __r1 * __ll_B | __ll_highpart (n0); \
+if (__d0 && __q1 > __m / __d0) \
+ do {\
+ __q1--, __r1 += (d);\
+ } while (__r1 >= (d)); \
if (__r1 < __m)\
{
\
- __q1--, __r1 += (d);\
- if (__r1 >= (d)) /* i.e. we didn't get carry when adding to __r1 */\
- if (__r1 < __m) \
+ do {\
__q1--, __r1 += (d);\
+ } while (__r1 >= (d) && __r1 < __m);\
}
\
__r1 -= __m; \
\
@@ -1381,12 +1384,15 @@ UDItype __umulsidi3 (USItype, USItype);
__q0 = __r1 / __d1;
\
__m = (UWtype) __q0 * __d0;
\
__r0 = __r0 * __ll_B | __ll_lowpart (n0); \
+if (__d0 && __q0 > __m / __d0) \
+ do {\
+ __q0--, __r0 += (d);\
+ } while (__r0 >= (d)); \
if (__r0 < __m)\
{
\
- __q0--, __r0 += (d);\
- if (__r0 >= (d))\
- if (__r0 < __m) \
+ do {\
__q0--, __r0 += (d);\
+ } while (__r0 >= (d) && __r0 < __m);\
}
\
__r0 -= __m; \
\
--
1.5.2
[PATCH] Fix bug on soft emulation of float point in gcc/longlong.c
There are 2 bugs existing in __udiv_qrnnd_c:
1. remainder could not be counted correctly
__r1 and __r0 are not adding enough, so I use do..while to replace
former if
2. the case of overflow of __m are not considered
so I add a couple of lines to handle it
I found the bugs from Linux kernel,
for PowerPC math-emu use __udiv_qrnnd_c to emulate float div.
I sent the patch to the maintainers of kernel, but they replied to me
that
this code came from gcc and should better keep coincident.
So I think I should send patch to here.
I don't know how to trigger __udiv_qrnnd_c in gcc.
but if the argument of __udiv_qrnnd_c are like
0x07f8 0x07f8 0x00210fff 0xc000 0x07f8,
the bug will be reproduced.
Also, you can try the program below, I have test it on Linux kernel
math-emu but haven't test on gcc.
---
#include
union fu {
unsigned int u;
float f;
};
union fu fdiv(union fu a, union fu b, union fu expected)
{
union fu z;
z.f = a.f / b.f;
printf("fdiv %08x / %08x = %08x expected %08x %s\n",
a.u, b.u, z.u, expected.u,
(z.u == expected.u) ? "(PASS)" : "(FAIL)");
printf(" %e / %e = %e\n", a.f, b.f, z.f);
}
int main()
{
union fu fa,fb,fe;
fa.u = 0xc0843fff;
fb.u = 0x80ff;
fe.u = 0x7f044000;
fdiv(fa,fb,fe);
}
--
---
gcc/longlong.h | 18 --
1 files changed, 12 insertions(+), 6 deletions(-)
diff --git a/gcc/longlong.h b/gcc/longlong.h
index 6c24564..256399b 100644
--- a/gcc/longlong.h
+++ b/gcc/longlong.h
@@ -1368,12 +1368,15 @@ UDItype __umulsidi3 (USItype, USItype);
__q1 = (n1) / __d1;
\
__m = (UWtype) __q1 * __d0;
\
__r1 = __r1 * __ll_B | __ll_highpart (n0);
\
+if (__d0 && __q1 > __m / __d0)
\
+ do {
\
+ __q1--, __r1 += (d);
\
+ } while (__r1 >= (d));
\
if (__r1 < __m)
\
{
\
- __q1--, __r1 += (d);
\
- if (__r1 >= (d)) /* i.e. we didn't get carry when adding to __r1
*/\
- if (__r1 < __m)
\
+ do {
\
__q1--, __r1 += (d);
\
+ } while (__r1 >= (d) && __r1 < __m);
\
}
\
__r1 -= __m;
\
\
@@ -1381,12 +1384,15 @@ UDItype __umulsidi3 (USItype, USItype);
__q0 = __r1 / __d1;
\
__m = (UWtype) __q0 * __d0;
\
__r0 = __r0 * __ll_B | __ll_lowpart (n0);
\
+if (__d0 && __q0 > __m / __d0)
\
+ do {
\
+ __q0--, __r0 += (d);
\
+ } while (__r0 >= (d));
\
if (__r0 < __m)
\
{
\
- __q0--, __r0 += (d);
\
- if (__r0 >= (d))
\
- if (__r0 < __m)
\
+ do {
\
__q0--, __r0 += (d);
\
+ } while (__r0 >= (d) && __r0 < __m);
\
}
\
__r0 -= __m;
\
\
--
RE: [PATCH] Fix bug on soft emulation of float point in gcc/longlong.c
Hi Ian, Thanks a lot for your detailed explanation. > -Original Message- > From: Ian Lance Taylor [mailto:[EMAIL PROTECTED] > Sent: Thursday, March 13, 2008 1:45 PM > To: Liu Yu > Cc: gcc@gcc.gnu.org > Subject: Re: [PATCH] Fix bug on soft emulation of float point > in gcc/longlong.c > > Liu Yu <[EMAIL PROTECTED]> writes: > > > There are 2 bugs existing in __udiv_qrnnd_c: > > > > 1. remainder could not be counted correctly > > __r1 and __r0 are not adding enough, so I use do..while to replace > > former if > > > > 2. the case of overflow of __m are not considered so I add > a couple of > > lines to handle it > > > > I found the bugs from Linux kernel, > > for PowerPC math-emu use __udiv_qrnnd_c to emulate float div. > > Is the kernel code testing UDIV_NEEDS_NORMALIZATION and > implementing the normalization required before calling __udiv_qrnnd_c? > > > > I don't know how to trigger __udiv_qrnnd_c in gcc. > > gcc will use __udiv_qrnnd_c in -lgcc for division and modulos > of 64-bit values on processors which can't do that directly > and for which there is no alternate implementation in longlong.h. > > > > but if the argument of __udiv_qrnnd_c are like > > 0x07f8 0x07f8 0x00210fff 0xc000 0x07f8, the > bug will > > be reproduced. > > Those arguments to __udiv_qrnnd_c are not valid. For correct > operation, the most significant bit of the last argument is > required to be 1. UDIV_NEEDS_NORMALIZATION signals this > fact. In gcc the required normalization is implemented > before calling __udiv_qrnnd_c, in __udivmoddi4 in libgcc2.c. > > This normalization approach is taken, rather than introducing > while loops as you suggest, because the while loops can run > for quite a long time on small numbers. > > Ian >
