llvm-gcc no longer accepts some assembly: Invalid or unsupported inline assembly!

[edwintorok]

Bugzilla Link	2267
Resolution	FIXED
Resolved on	May 21, 2008 23:47
Version	unspecified
OS	Linux
CC	@asl,@EdSchouten

Extended Description

I tried to build glibc with llvm-gcc4.2, and I get lots of there errors:

abort.c: In function ‘abort’:
abort.c:86: error: Invalid or unsupported inline assembly!

This errors is shown many times, I picked the error from abort.c:

$ grep "unsupported inline assembly" ../../i386-libc/log|wc -l
1043

The reduced testcase shows it is a libc asm low-level lock function that is not accepted by llvm-gcc. A while ago (I don't remember which revision) I was able to compile glibc without these errors.

To reproduce:
$ gcc testcase-min.i -fsyntax-only
$ llvm-gcc testcase-min.i -fsyntax-only
testcase-min.i: In function ‘abort’:
testcase-min.i:10: error: Invalid or unsupported inline assembly!

$ uname -a
Linux thunder 2.6.25-rc6-00268-gd2532dd-dirty #​23 SMP PREEMPT Sun Mar 23 18:34:55 EET 2008 i686 GNU/Linux

testcase-min.i:
typedef struct {
int lock;
int cnt;
}
__libc_lock_recursive_t;
static int stage;
static __libc_lock_recursive_t lock;
void abort (void) {
if (stage == 2) {
do { if (--(lock).cnt == 0) { (void) ({ int ignore; if (__builtin_constant_p (0) && (0) == 0) __asm __volatile ("cmpl $0, %%gs:%P3\n\t" "je 0f\n\t" "lock\n" "0:\tsubl $1,%0\n\t" "jne L_unlock%=\n\t" ".subsection 1\n\t" ".type L_unlock%=,@function\n" "L_unlock%=:\n" "1:\tleal %0, %%eax\n" "2:\tcall __lll_unlock_wake_private\n" "3:\tjmp 18f\n" "4:\t.size L_unlock%=, 4b-1b\n\t" ".previous\n" ".section .eh_frame,"a",@progbits\n" "5:\t" ".long 7f-6f # Length of Common Information Entry\n" "6:\t" ".long 0x0 # CIE Identifier Tag\n\t" ".byte 0x1 # CIE Version\n\t" ".ascii "zR\0" # CIE Augmentation\n\t" ".uleb128 0x1 # CIE Code Alignment Factor\n\t" ".sleb128 -4 # CIE Data Alignment Factor\n\t" ".byte 0x8 # CIE RA Column\n\t" ".uleb128 0x1 # Augmentation size\n\t" ".byte 0x1b # FDE Encoding (pcrel sdata4)\n\t" ".byte 0xc # DW_CFA_def_cfa\n\t" ".uleb128 0x4\n\t" ".uleb128 0x0\n\t" ".align 4\n" "7:\t" ".long 17f-8f # FDE Length\n" "8:\t" ".long 8b-5b # FDE CIE offset\n\t" ".long 1b-.# FDE initial location\n\t" ".long 4b-1b # FDE address range\n\t" ".uleb128 0x0 # Augmentation size\n\t" ".byte 0x16 # DW_CFA_val_expression\n\t" ".uleb128 0x8\n\t" ".uleb128 10f-9f\n" "9:\t" ".byte 0x78 # DW_OP_breg8\n\t" ".sleb128 3b-1b\n" "10:\t" ".byte 0x40 + (2b-1b) # DW_CFA_advance_loc\n\t" ".byte 0x16 # DW_CFA_val_expression\n\t" ".uleb128 0x8\n\t" ".uleb128 12f-11f\n" "11:\t" ".byte 0x78 # DW_OP_breg8\n\t" ".sleb128 3b-2b\n" "12:\t" ".byte 0x40 + (3b-2b-1) # DW_CFA_advance_loc\n\t" ".byte 0x16 # DW_CFA_val_expression\n\t" ".uleb128 0x8\n\t" ".uleb128 16f-13f\n" "13:\t" ".byte 0x78 # DW_OP_breg8\n\t" ".sleb128 15f-14f\n\t" ".byte 0x0d # DW_OP_const4s\n" "14:\t" ".4byte 3b-.\n\t" ".byte 0x1c # DW_OP_minus\n\t" ".byte 0x0d # DW_OP_const4s\n" "15:\t" ".4byte 18f-.\n\t" ".byte 0x22 # DW_OP_plus\n" "16:\t" ".align 4\n" "17:\t" ".previous\n" "18:" : "=m" ((lock).lock), "=&a" (ignore) : "m" ((lock).lock), "i" (12) : "memory"); } ); } }
while (0);
}
while (1) asm ("hlt");
}

[edwintorok]

Using gcc -dD -E these are the macros that assembly comes from:
__libc_lock_unlock_recursive (lock);

#define __libc_lock_unlock_recursive(NAME) do { if (--(NAME).cnt == 0) { (NAME).owner = NULL; lll_unlock ((NAME).lock, LLL_PRIVATE); } } while (0)

449 "../nptl/sysdeps/unix/sysv/linux/i386/lowlevellock.h"

#define __lll_unlock_asm "cmpl $0, %%gs:%P3\n\t" "je 0f\n\t" "lock\n" "0:\tsubl $1,%0\n\t"

#define lll_unlock(futex,private) (void) ({ int ignore; if (__builtin_constant_p (private) && (private) == LLL_PRIVATE) __asm __volatile (__lll_unlock_asm "jne L_unlock%=\n\t" ".subsection 1\n\t" ".type L_unlock%=,@function\n" "L_unlock%=:\n" "1:\tleal %0, %%eax\n" "2:\tcall __lll_unlock_wake_private\n" "3:\tjmp 18f\n" "4:\t.size L_unlock%=, 4b-1b\n\t" ".previous\n" LLL_STUB_UNWIND_INFO_3 "18:" : "=m" (futex), "=&a" (ignore) : "m" (futex), "i" (MULTIPLE_THREADS_OFFSET) : "memory"); else { int ignore2; __asm __volatile (__lll_unlock_asm "jne L_unlock%=\n\t" ".subsection 1\n\t" ".type L_unlock%=,@function\n" "L_unlock%=:\n" "1:\tleal %0, %%eax\n" "0:\tmovl %5, %%ecx\n" "2:\tcall __lll_unlock_wake\n" "3:\tjmp 18f\n" "4:\t.size L_unlock%=, 4b-1b\n\t" ".previous\n" LLL_STUB_UNWIND_INFO_4 "18:" : "=m" (futex), "=&a" (ignore), "=&c" (ignore2) : "i" (MULTIPLE_THREADS_OFFSET), "m" (futex), "g" (private) : "memory"); } })

[EdSchouten]

Just want to say "me too". I know I'm a little late, but it would be great if at least some more constructs would be supported for the upcoming 2.3 release.

I've been trying LLVM on FreeBSD/i386 at the office. I was capable of compiling a kernel, except for some of the atomic macro's in sys/i386/include/atomic.h, see:

http://cvsweb.freebsd.org/src/sys/i386/include/atomic.h

LLVM-GCC bails out on lines 263-266 (the four ATOMIC_STORE_LOAD statements), giving the same error message.

If this could be fixed before the 2.3 release, we can compile FreeBSD SMP kernels using LLVM! :-)

[EdSchouten]

Testcase for bug 2267, based on the FreeBSD source code
I've trimmed down the root of the cause, based on the FreeBSD atomic instruction in the kernel header file.

[llvmbot]

I can't reproduce on x86-linux (32 and 64 bit) at -O0, ..., -O3
using latest svn. That's odd because I don't think anything
changed in this area since the release was branched off.

[lattner]

I just verified that the attach testcase works fine on llvm-2.3 on darwin/x86. The testcase in the first comment doesn't work because it is munged and doesn't compile.

[EdSchouten]

Proper testcase for bug 2267
I initially though the lock; cmpxchg instructions caused compilation to fail. After performing some more experiments, it turned out to be related to xchg.

[lattner]

verified, investigating.

[lattner]

Fixed, patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080519/062746.html
Testcase here: test/CodeGen/X86/asm-indirect-mem.ll