avcodec/x86/mlpdsp_init: Simplify mlp_filter_channel_x86()
authorMichael Niedermayer <michaelni@gmx.at>
Thu, 19 Feb 2015 15:25:29 +0000 (16:25 +0100)
committerMichael Niedermayer <michaelni@gmx.at>
Sat, 21 Mar 2015 00:52:51 +0000 (01:52 +0100)
Based on patch by Francisco Blas Izquierdo Riera
Commit message partly taken from carl

fixes a compilation
error in mlpdsp_init.c with -fstack-check and some gcc compilers (I
reproduced the issue with gcc 4.7.3) by simplifying the code.

See also https://bugs.gentoo.org/show_bug.cgi?id=471756

$ make libavcodec/x86/mlpdsp_init.o
libavcodec/x86/mlpdsp_init.c: In function ‘mlp_filter_channel_x86’:
libavcodec/x86/mlpdsp_init.c:142:5: error: can’t find a register in
class ‘GENERAL_REGS’ while reloading ‘asm’
libavcodec/x86/mlpdsp_init.c:142:5: error: ‘asm’ operand has impossible
constraints

4551 -> 4509 dezicycles

Reviewed-by: Ramiro Polla <ramiro.polla@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
(cherry picked from commit 03f39fbb2a558153a3c464edec1378d637a755fe)

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
libavcodec/x86/mlpdsp_init.c

index dc0bc58..e9d9b1b 100644 (file)
@@ -148,8 +148,8 @@ static void mlp_filter_channel_x86(int32_t *state, const int32_t *coeff,
         FIRMUL   (ff_mlp_firorder_6, 0x14   )
         FIRMUL   (ff_mlp_firorder_5, 0x10   )
         FIRMUL   (ff_mlp_firorder_4, 0x0c   )
-        FIRMULREG(ff_mlp_firorder_3, 0x08,10)
-        FIRMULREG(ff_mlp_firorder_2, 0x04, 9)
+        FIRMUL   (ff_mlp_firorder_3, 0x08   )
+        FIRMUL   (ff_mlp_firorder_2, 0x04   )
         FIRMULREG(ff_mlp_firorder_1, 0x00, 8)
         LABEL_MANGLE(ff_mlp_firorder_0)":\n\t"
         "jmp  *%6                     \n\t"
@@ -178,8 +178,6 @@ static void mlp_filter_channel_x86(int32_t *state, const int32_t *coeff,
         : /* 4*/"r"((x86_reg)mask), /* 5*/"r"(firjump),
           /* 6*/"r"(iirjump)      , /* 7*/"c"(filter_shift)
         , /* 8*/"r"((int64_t)coeff[0])
-        , /* 9*/"r"((int64_t)coeff[1])
-        , /*10*/"r"((int64_t)coeff[2])
         : "rax", "rdx", "rsi"
 #else /* ARCH_X86_32 */
           /* 3*/"+m"(blocksize)