On Mon, Mar 19, 2012 at 10:14 AM, Ronald S. Bultje <[email protected]> wrote: > Hi, > > On Mon, Mar 19, 2012 at 9:48 AM, Jason Garrett-Glaser <[email protected]> wrote: >> On Sat, Mar 17, 2012 at 9:34 AM, Ronald S. Bultje <[email protected]> wrote: >>> --- >>> libavcodec/x86/cabac.h | 17 ++++++++++------- >>> 1 files changed, 10 insertions(+), 7 deletions(-) >>> >>> diff --git a/libavcodec/x86/cabac.h b/libavcodec/x86/cabac.h >>> index 3c3652d..c4832c3 100644 >>> --- a/libavcodec/x86/cabac.h >>> +++ b/libavcodec/x86/cabac.h >>> @@ -105,8 +105,8 @@ static av_always_inline int >>> get_cabac_bypass_sign_x86(CABACContext *c, int val) >>> { >>> x86_reg tmp; >>> __asm__ volatile( >>> - "movl %4, %k1 \n\t" >>> - "movl %2, %%eax \n\t" >>> + "movl %c5(%2), %k1 \n\t" >>> + "movl %c3(%2), %%eax \n\t" >>> "shl $17, %k1 \n\t" >>> "add %%eax, %%eax \n\t" >>> "sub %k1, %%eax \n\t" >>> @@ -117,7 +117,7 @@ static av_always_inline int >>> get_cabac_bypass_sign_x86(CABACContext *c, int val) >>> "sub %%edx, %%ecx \n\t" >>> "test %%ax, %%ax \n\t" >>> " jnz 1f \n\t" >>> - "mov %3, %1 \n\t" >>> + "mov %c4(%2), %1 \n\t" >>> "subl $0xFFFF, %%eax \n\t" >>> "movzwl (%1), %%edx \n\t" >>> "bswap %%edx \n\t" >>> @@ -126,11 +126,14 @@ static av_always_inline int >>> get_cabac_bypass_sign_x86(CABACContext *c, int val) >>> "addl %%edx, %%eax \n\t" >>> "mov %1, %3 \n\t" >>> "1: \n\t" >>> - "movl %%eax, %2 \n\t" >>> + "movl %%eax, %c4(%2) \n\t" >>> >>> - :"+c"(val), "=&r"(tmp), "+m"(c->low), "+m"(c->bytestream) >>> - :"m"(c->range) >>> - : "%eax", "%edx" >>> + : "+c"(val), "=&r"(tmp) >>> + : "r"(c), >>> + "i"(offsetof(CABACContext, low)), >>> + "i"(offsetof(CABACContext, bytestream)), >>> + "i"(offsetof(CABACContext, range)) >>> + : "%eax", "%edx", "memory" >> >> IMO clobbering memory looks very very hacky, and I don't like it. If >> you need to clobber something, it'd be much better if we could clobber >> exactly what needs clobbering, and nothing more. > > Well, I don't think inline assembly supports explicitely clobbering > variables without marking them as "+m" or "+r", which messes up the > register allocator for at least gcc-4.2.1 (it uses a different > register for each "m"(c->...), thus running out of registers; yes, > there's many things wrong there).
You can clobber a memory location without referencing it in the asm, and thus without allocating a register for it. Jason _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
