Skip to contentSkip to conversation

rc4-mmotm1110 - Another build error

DISCUSSION [email protected] 3 participants 6 messages Nov 11, 2008

lore
Share & tools
lore
#0 lore raw

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed only inside a function make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

#1 lore raw
Quoted 7 lines · On Mon, 10 Nov 2008 22:20:33 -0500 [email protected] wrote:

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

Quoted 9 lines
                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed only inside a function make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

I hadn't got around to testing that one yet.

So ug. ALIGN() is used in array sizing and hence has to be a compile-time thing. But ALIGN(foo, bar()) will call bar() twice.

Now how do we fix that?

#2 lore raw
Quoted 19 lines · On Mon, 10 Nov 2008 19:43:53 -0800 Andrew Morton <[email protected]>…
Quoted 7 lines · On Mon, 10 Nov 2008 22:20:33 -0500 [email protected] wrote:

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

Quoted 9 lines
                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed only inside a function make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

Quoted 6 lines

I hadn't got around to testing that one yet.

So ug. ALIGN() is used in array sizing and hence has to be a compile-time thing. But ALIGN(foo, bar()) will call bar() twice.

Now how do we fix that?

How about provoding
- ALIGN() - args can be evaluated twice.
and
- static inline unsigned long align()

messy ?

Thanks,
-Kame

#3 lore raw
Quoted 24 lines · On Tue, 11 Nov 2008 15:54:38 +0900 KAMEZAWA Hiroyuki…
Quoted 21 lines · On Mon, 10 Nov 2008 19:43:53 -0800 Andrew Morton <[email protected]>…
Quoted 7 lines · On Mon, 10 Nov 2008 22:20:33 -0500 [email protected] wrote:

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

Quoted 9 lines
                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed only inside a function make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

I hadn't got around to testing that one yet.

Quoted 7 lines
Quoted 4 lines

So ug. ALIGN() is used in array sizing and hence has to be a compile-time thing. But ALIGN(foo, bar()) will call bar() twice.

Now how do we fix that?

How about provoding
- ALIGN() - args can be evaluated twice.

Well. The problem here is the risk of mistakes, and such an ALIGN() could still be called by mistake.

Quoted 2 lines

and
- static inline unsigned long align()

ALIGN() is designed so that it can take any scalar types as arguments and still do the right thing, so it can't really be implemented as an inline.

Quoted 1 line

messy ?

I suppose we could switch all those ALIGN calls which must occur at compile-time over to a new STATIC_ALIGN() (?) which is allowed to evaluate its args twice, then fix ALIGN().

A bit of a pain, but any missed conversions would reliably fail to compile and would get fixed quickly.

hm, ALIGN() is used in assembly too, but it means something totally different, sigh.

<greps>

./arch/x86/mm/hugetlbpage.c: addr = ALIGN(start_addr, huge_page_size(h));
./arch/x86/mm/hugetlbpage.c: addr = ALIGN(vma->vm_end, huge_page_size(h));
./arch/x86/mm/hugetlbpage.c: addr = ALIGN(addr, huge_page_size(h));

These will presently be calling huge_page_size() twice.

Will these:

./drivers/mtd/ubi/build.c: ubi->ec_hdr_alsize = ALIGN(UBI_EC_HDR_SIZE, ubi->hdrs_min_io_size);
./drivers/mtd/ubi/build.c: ubi->vid_hdr_alsize = ALIGN(UBI_VID_HDR_SIZE, ubi->hdrs_min_io_size);

do the dereference twice? Probably not.

I didn't immediately spot any bugs, but I didn't look very closely. Hopefully if there are any there, we'd have heard about it.

#4 lore raw
In reply to Andrew Morton
Quoted 5 lines · Andrew Morton
Quoted 3 lines · On Mon, 10 Nov 2008 22:20:33 -0500 [email protected] wrote:

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

Quoted 7 lines
Quoted 7 lines

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed

only inside a function

Quoted 7 lines
Quoted 6 lines

make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

Quoted 6 lines

I hadn't got around to testing that one yet.

So ug. ALIGN() is used in array sizing and hence has to be a compile-time thing. But ALIGN(foo, bar()) will call bar() twice.

Now how do we fix that?

Can we abuse __builtin_constant_p somehow? Maybe we can make a definition something like:

#define __align_mask(x,mask) (
	__builtin_constant_p(mask) ? {
		(((x)+(mask))&~(mask)) :
		{              \
		      typeof(mask) __mask = mask;             \
		       (((x) + __mask) & ~__mask);             \
		}
	})

If it's a compile-time constant, it's safe to evaluate mask twice. And if it's *not* a constant, it shouldn't be used for sizing arrays in open code. And if it's an array local to a procedure, it probably should be kmalloc()'ed instead.

#5 lore raw
Quoted 40 lines · On Tue, 11 Nov 2008 14:04:57 -0500 [email protected] wrote:
Quoted 5 lines · Andrew Morton
Quoted 3 lines · On Mon, 10 Nov 2008 22:20:33 -0500 [email protected] wrote:

gcc --version says:
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)

Quoted 7 lines
Quoted 7 lines

ka-blam number 2:

CC kernel/audit.o
In file included from include/net/dst.h:15,

                 from include/net/sock.h:57,
                 from kernel/audit.c:54:

include/net/neighbour.h:114: error: braced-group within expression allowed

only inside a function

Quoted 7 lines
Quoted 6 lines

make[1]: *** [kernel/audit.o] Error 1

I'm placing bets on patches/align-avoid-evaluating-its-argument-twice.patch

Yep, revert that patch, and audit.o compiles again.

Quoted 6 lines

I hadn't got around to testing that one yet.

So ug. ALIGN() is used in array sizing and hence has to be a compile-time thing. But ALIGN(foo, bar()) will call bar() twice.

Now how do we fix that?

Can we abuse __builtin_constant_p somehow? Maybe we can make a definition something like:

#define __align_mask(x,mask) (
	__builtin_constant_p(mask) ? {
		(((x)+(mask))&~(mask)) :
		{              \
		      typeof(mask) __mask = mask;             \
		       (((x) + __mask) & ~__mask);             \
		}
	})
Quoted 1 line

If it's a compile-time constant, it's safe to evaluate mask twice.

This seems to compile:

int y;
char x[__builtin_constant_p(1) ? 42 : y + 1];

so yes, that might work. For this gcc version. However I suspect that it would blow up on us. We keep on having problems with the __builtin_constant_p() in kmalloc() doing wrong things (with gcc-3.2.x, iirc).

It'd be worth trying though.

DIV_ROUND_UP() and roundup() are busted too.