stack usage with libgcc math funcs

stack usage with libgcc math funcs

[Mike Frysinger]

[-- Attachment #1: Type: text/plain, Size: 1506 bytes --]

i dont know superh assembly, so i cant really check the code myself.  starting 
with at gcc-4.1.2 (at least, gcc-3.4.6 doesnt seem to have this issue), a few 
objects in libgcc.a lack GNU-stack markings.  here is the scanelf output 
against libgcc.a from gcc-4.3.2.  basically, '!WX' just means that the objects 
do not have a .note.GNU-stack section.

$ scanelf -qeR .
!WX --- ---  ./4.3.2/_ashiftrt.o
!WX --- ---  ./4.3.2/_ashiftrt_n.o
!WX --- ---  ./4.3.2/_ashiftlt.o
!WX --- ---  ./4.3.2/_lshiftrt.o
!WX --- ---  ./4.3.2/_movmem.o
!WX --- ---  ./4.3.2/_movmem_i4.o
!WX --- ---  ./4.3.2/_mulsi3.o
!WX --- ---  ./4.3.2/_sdivsi3.o
!WX --- ---  ./4.3.2/_sdivsi3_i4.o
!WX --- ---  ./4.3.2/_udivsi3.o
!WX --- ---  ./4.3.2/_udivsi3_i4.o
!WX --- ---  ./4.3.2/_set_fpscr.o
!WX --- ---  ./4.3.2/_div_table.o
!WX --- ---  ./4.3.2/_udiv_qrnnd_16.o
!WX --- ---  ./4.3.2/_ic_invalidate.o
!WX --- ---  ./4.3.2/_ic_invalidate_array.o
!WX --- ---  ./4.3.2/linux-atomic.o

assuming libgcc doesnt actually use the stack to exec code, the attached patch 
should fix things i think.
-mike

--- gcc/config/sh/lib1funcs.asm
+++ gcc/config/sh/lib1funcs.asm
@@ -34,6 +34,11 @@
 !! recoded in assembly by Toshiyasu Morita
 !! tm@netcom.com
 
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+.previous
+#endif
+
 /* SH2 optimizations for ___ashrsi3, ___ashlsi3, ___lshrsi3 and
    ELF local label prefixes by J"orn Rennecke
    amylaar@cygnus.com  */

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 835 bytes --]

Re: stack usage with libgcc math funcs

[Mike Frysinger]

[-- Attachment #1: Type: text/plain, Size: 515 bytes --]

and linux-atomic.asm needs a similar change
--- gcc/config/sh/linux-atomic.asm
+++ gcc/config/sh/linux-atomic.asm
@@ -138,3 +138,8 @@
 ATOMIC_FETCH_AND_COMBOP(nand,and,not,4,l,mov)
 
 #endif /* ! __SH5__ */
+
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+.previous
+#endif

as a note, without these changes, the execstack markings bubble up into some 
of the glibc libs as those are built with -static-libgcc ... i can see this 
with glibc-2.9 for sure.
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 835 bytes --]

Re: stack usage with libgcc math funcs

[Kaz Kojima]

Mike Frysinger <vapier@gentoo.org> wrote:
> i dont know superh assembly, so i cant really check the code myself.  starting 
> with at gcc-4.1.2 (at least, gcc-3.4.6 doesnt seem to have this issue), a few 
> objects in libgcc.a lack GNU-stack markings.  here is the scanelf output 
> against libgcc.a from gcc-4.3.2.  basically, '!WX' just means that the objects 
> do not have a .note.GNU-stack section.
> assuming libgcc doesnt actually use the stack to exec code, the attached patch 
> should fix things i think.

Thanks for pointing this out.  Could you please send your
patches to gcc-patches or file this issue to gcc bugzilla?
Unfortunately, now gcc trunk is in stage 4 and any patches
for non strict regressions can't be accepted temporary.
I'll schedule these patches for 4.5, then apply them to
the open branches after 4.5.

Thanks,
	kaz

Re: stack usage with libgcc math funcs

[Mike Frysinger]

[-- Attachment #1: Type: text/plain, Size: 1263 bytes --]

On Wednesday 24 December 2008 17:57:53 Kaz Kojima wrote:
> Mike Frysinger <vapier@gentoo.org> wrote:
> > i dont know superh assembly, so i cant really check the code myself. 
> > starting with at gcc-4.1.2 (at least, gcc-3.4.6 doesnt seem to have this
> > issue), a few objects in libgcc.a lack GNU-stack markings.  here is the
> > scanelf output against libgcc.a from gcc-4.3.2.  basically, '!WX' just
> > means that the objects do not have a .note.GNU-stack section.
> > assuming libgcc doesnt actually use the stack to exec code, the attached
> > patch should fix things i think.
>
> Thanks for pointing this out.

i'm guessing that means they're correct :)

> Could you please send your
> patches to gcc-patches or file this issue to gcc bugzilla?

i dont think people like it when i post patches to bugzilla.  they tell me to 
use gcc-patches.

> Unfortunately, now gcc trunk is in stage 4 and any patches
> for non strict regressions can't be accepted temporary.
> I'll schedule these patches for 4.5, then apply them to
> the open branches after 4.5.

well, i would say this is a long standing regression.  gcc-3.4.6 does not 
result in exec stacks with any of the libgcc.a math functions ... but you're 
the gcc expert.
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 835 bytes --]

Re: stack usage with libgcc math funcs

[Kaz Kojima]

Mike Frysinger <vapier@gentoo.org> wrote:
> well, i would say this is a long standing regression.  gcc-3.4.6 does not 
> result in exec stacks with any of the libgcc.a math functions ... but you're 
> the gcc expert.

.note.GNU-stack wasn't used in gcc-3.4 era.  This type of
problem wouldn't be a "strict" regression, from the compiler's
viewpoint.

Regards,
	kaz