gcc 4.6 miscompiles this source from ffmpeg on x86-64-apple-darwin10, whereas previous compilers worked. I'm not sure if the asm is legal, but it's existed in the wild for a long time. const unsigned long long __attribute__((aligned(8))) ff_bgr24toUV[2][4] = { {0x38380000DAC83838ULL, 0xECFFDAC80000ECFFULL, 0xF6E40000D0E3F6E4ULL, 0x3838D0E300003838ULL}, {0xECFF0000DAC8ECFFULL, 0x3838DAC800003838ULL , 0x38380000D0E33838ULL, 0xF6E4D0E30000F6E4ULL}, }; static void bgr24ToUV_mmx_MMX2(int f) { __asm__ volatile( "movq 24+%0, %%mm6 \n\t" :: "m"(ff_bgr24toUV[f == 0][0])); } void rgb24ToUV_MMX2() { bgr24ToUV_mmx_MMX2(1); } > gcc -v Using built-in specs. COLLECT_GCC=/usr/local/gcc46/bin/gcc COLLECT_LTO_WRAPPER=/usr/local/gcc46/libexec/gcc/x86_64-apple-darwin10.5.0/4.6.0/lto-wrapper Target: x86_64-apple-darwin10.5.0 Configured with: ../../src/gcc/configure --prefix=/usr/local/gcc46 --with-arch=native --with-tune=native --disable-nls --with-gmp=/sw --disable-bootstrap --enable-checking --enable-languages=c,c++,lto,objc,obj-c++ Thread model: posix gcc version 4.6.0 20101122 (experimental) (GCC) > gcc -O -o swscale-fails.s -S swscale.i swscale.i: In function 'rgb24ToUV_MMX2': swscale.i:10:2: warning: use of memory input without lvalue in asm operand 0 is deprecated [enabled by default] Working asm (4.2): _rgb24ToUV_MMX2: pushq %rbp movq %rsp, %rbp movq 24+_ff_bgr24toUV(%rip), %mm6 leave ret .globl _ff_bgr24toUV .const .align 3 _ff_bgr24toUV: .quad 4050987868490315832 .quad -1369135209168966401 .quad -656399642184648988 .quad 4051217538195929144 .quad -1369375758026740481 .quad 4051228417348089912 .quad 4050987868324313144 .quad -656169972313032988 .section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support Non-working asm (4.6): _rgb24ToUV_MMX2: movq 24+LC0(%rip), %mm6 ret .globl _ff_bgr24toUV .const .align 3 _ff_bgr24toUV: .quad 4050987868490315832 .quad -1369135209168966401 .quad -656399642184648988 .quad 4051217538195929144 .quad -1369375758026740481 .quad 4051228417348089912 .quad 4050987868324313144 .quad -656169972313032988 .literal8 .align 3 LC0: .quad 4050987868490315832 .section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support 24+_ff_bgr24toUV(%rip) is fine, but 24+LC0(%rip) is a pointer to nothing, and ld breaks: ld: in /var/folders/MY/MYkVh2TwHgKZhNFIG8M3wU+++TI/-Tmp-//cc9dJIWa.o, in section __TEXT,__text reloc 0: local relocation for address 0x0000000C in section __text does not target section __literal8 I'm going to fix the asm since it looks fragile anyway, but that won't fix existing releases of ffmpeg. Note that creating LC0 is not even an optimization since it doesn't save any space (because the array is __attribute__((used))).
That is really invalid. The "m" operand is just long long, so there are no guarantees that you get anything around it.
I added a patch to mpeg2dec in Mandriva by removing the const modifier of the vectors. It was basic hacking with some trial&error to correct https://bugzilla.gnome.org/show_bug.cgi?id=649930 more information at: https://qa.mandriva.com/show_bug.cgi?id=63279 and the patch being used is: http://svn.mandriva.com/cgi-bin/viewvc.cgi/packages/cooker/mpeg2dec/current/SOURCES/libmpeg2-0.5.1-gcc4.6.patch?view=markup&pathrev=674312 Commenting here because there should exist more code out there that will fail due to the same reason.