On Sat, Jul 07, 2012 at 03:15:36AM -0700, Albert ARIBAUD wrote:
Hi Allen,
On Fri, 6 Jul 2012 16:17:19 -0700, Allen Martin amartin@nvidia.com wrote:
On Fri, Jul 06, 2012 at 01:44:32PM -0700, Stephen Warren wrote:
On 07/06/2012 02:33 PM, Allen Martin wrote:
On Fri, Jul 06, 2012 at 12:09:43PM -0700, Stephen Warren wrote:
On 07/06/2012 12:08 PM, Allen Martin wrote:
Rearrange the link order of libraries to avoid out of bound relocations in thumb mode. I have no idea how to fix this for real.
Are the relocations branches or something else? It looks like unconditional jump range is +/-4MB for Thumb1 and +/-16MB for Thumb2, so I'm surprised we'd be exceeding that, considering the U-boot binary is on the order of 256KB on Tegra right now.
This is the relcation type:
arch/arm/lib/libarm.o: In function `__flush_dcache_all': /home/arm/u-boot/arch/arm/lib/cache.c:52: relocation truncated to fit: R_ARM_THM_JUMP11 against symbol `flush_cache' defined in .text section in arch/arm/cpu/armv7/libarmv7.o
The instruction is a "b.n" not a "b", which is what is causing the problem.
I think because of the weak alias the compiler used a short jump to the local function, but when it got linked it resolved to a function that was too far away for the short jump:
void flush_cache(unsigned long start, unsigned long size) __attribute__((weak, alias("__flush_cache")));
00000002 <__flush_dcache_all>: 2: 2000 movs r0, #0 4: f04f 31ff mov.w r1, #4294967295 ; 0xffffffff 8: e7fe b.n 0 <__flush_cache>
Ah, that explanation makes sense.
It looks like there's a "-fno-optimize-sibling-calls" option to gcc to avoid this problem. Seems a shame to disable all short jumps for this one case though.
It seems like a bug that the b-vs-b.n optimization is applied to a weak symbol, since the compiler can't possibly know the range of the jump.
Also, I've seen ld for some architectures rewrite the equivalent of b.n to plain b when needing to expand the branch target range; IIRC a process known as "relaxing"? Perhaps gcc is expecting ld to do that, but ld isn't?
And I forgot to mention, the code bloat from disabling the optimization is about 400 bytes (185136 -> 185540), so it's not bad, but it still seems a shame to disable all short branches because of one misoptimized one.
Can this not be limited to compiling the object files which are known to be sensitive to the problem?
I understand this issue fairly well now. It's a known bug in the assembler that has already been fixed:
http://sourceware.org/bugzilla/show_bug.cgi?id=12532
It only impacts preembtable symbols, and since u-boot doesn't have any dynamic loadable objects it's only explictly defined weak symbols that should trigger the bug.
I built a new toolchain with binutils 2.22 and verified the bug is no longer present there, and -fno-optimize-sibling-calls is the correct workaround for toolchains that do have the bug, so conditionally disabling the optimization for binutils < 2.22 seems like the right fix.
I ran a quick scrub of the u-boot tree and there's 195 instances of __attribute__((weak)) spread across 123 source files, so I think just disabling optimization on the failing object files may be too fragile, as code movement could cause others to crop up.
-Allen