root/src/lib/amd64_blend.S

/* [<][>][^][v][top][bottom][index][help] */
#include <config.h>
#include "asm.h"

#ifdef DO_AMD64_ASM

/*\ 
|*| AMD64 SSE2 assembly blending routines for Imlib2
|*| Written by John Slaten <zartheenumerator@comcast.net>
|*| Based on MMX routines written by Willem Monsuwe <willem@stack.nl>
\*/

/*\ Some useful masks \*/
.data
m0X000000: .byte   0,   0,   0,   0,   0,   0, 255,   0
           .byte   0,   0,   0,   0,   0,   0, 255,   0
m10000000: .byte   0,   0,   0,   0,   0,   0,   0,   1
           .byte   0,   0,   0,   0,   0,   0,   0,   1
m00XXXXXX: .byte 255, 255, 255, 255, 255, 255,   0,   0
           .byte 255, 255, 255, 255, 255, 255,   0,   0
mVX000000: .byte   0,   0,   0,   0,   0,   0, 255, 127
           .byte   0,   0,   0,   0,   0,   0, 255, 127
mV0000000: .byte   0,   0,   0,   0,   0,   0,   0, 128
           .byte   0,   0,   0,   0,   0,   0,   0, 128
mX000X000: .byte   0,   0,   0,   0,   0,   0, 255, 255
           .byte   0,   0,   0,   0,   0,   0, 255, 255
m0XXX0XXX0XXX0XXX: .byte 255, 255, 255,   0, 255, 255, 255,   0
                   .byte 255, 255, 255,   0, 255, 255, 255,   0
m0XXX0XXX00000000: .byte 255, 255, 255,   0, 255, 255, 255,   0
                   .byte   0,   0,   0,   0,   0,   0,   0,   0
m0XXX000000000000: .byte 255, 255, 255,   0,   0,   0,   0,   0
                   .byte   0,   0,   0,   0,   0,   0,   0,   0
mX000X000X000X000: .byte   0,   0,   0, 255,   0,   0,   0, 255
                   .byte   0,   0,   0, 255,   0,   0,   0, 255
mX000X00000000000: .byte   0,   0,   0, 255,   0,   0,   0, 255
                   .byte   0,   0,   0, 255,   0,   0,   0, 255
mX000000000000000: .byte   0,   0,   0, 255,   0,   0,   0, 255
                   .byte   0,   0,   0, 255,   0,   0,   0, 255
m1000100010001000: .byte   0,   0,   0,   1,   0,   0,   0,   1
                   .byte   0,   0,   0,   1,   0,   0,   0,   1
m000V0V0V000V0V0V: .byte 127,   0, 127,   0, 127,   0,   0,   0
                   .byte 127,   0, 127,   0, 127,   0,   0,   0
mI0000000I0000000: .byte   0,   0,   0,   0,   0,   0,   0,  64
                   .byte   0,   0,   0,   0,   0,   0,   0,  64
m0VVV0VVV0VVV0VVV: .byte 127, 127, 127,   0, 127, 127, 127,   0
                   .byte 127, 127, 127,   0, 127, 127, 127,   0
c1: .word 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1

/*\ All functions have the same calling convention:
|*|  __imlib_amd64_<op>_rgba_to_rgb[A](void *src, int sw, void *dst, int dw,
|*|                                    int w, int h, ImlibColorModifier *cm)
|*| AMD64 GCC passes paramters by register, so no aliases exist in this version.
\*/

.text
        .align 16
FN_(imlib_amd64_blend_rgba_to_rgb)
FN_(imlib_amd64_blend_rgba_to_rgba)
FN_(imlib_amd64_copy_rgba_to_rgb)
FN_(imlib_amd64_copy_rgba_to_rgba)

FN_(imlib_amd64_copy_rgb_to_rgba)
FN_(imlib_amd64_add_blend_rgba_to_rgb)
FN_(imlib_amd64_add_blend_rgba_to_rgba)
FN_(imlib_amd64_add_copy_rgba_to_rgb)
FN_(imlib_amd64_add_copy_rgba_to_rgba)
FN_(imlib_amd64_add_copy_rgb_to_rgba)

FN_(imlib_amd64_subtract_blend_rgba_to_rgb)
FN_(imlib_amd64_subtract_blend_rgba_to_rgba)
FN_(imlib_amd64_subtract_copy_rgba_to_rgb)
FN_(imlib_amd64_subtract_copy_rgba_to_rgba)
FN_(imlib_amd64_subtract_copy_rgb_to_rgba)

FN_(imlib_amd64_reshade_blend_rgba_to_rgb)
FN_(imlib_amd64_reshade_blend_rgba_to_rgba)
FN_(imlib_amd64_reshade_copy_rgba_to_rgb)
FN_(imlib_amd64_reshade_copy_rgba_to_rgba)
FN_(imlib_amd64_reshade_copy_rgb_to_rgba)

.extern pow_lut
        
/*\ SSE register use:
|*| %xmm1 = Source value
|*| %xmm2 = Destination value
|*| %xmm3 = Alpha value
|*| %xmm4 = 0
|*| %xmm5-%xmm7 = masks
\*/

/*\ Variables:
|*| %rsi = src
|*| %rdi = dst
|*| %r8d = w
|*| %r9d = h
|*| %r10d = sw
|*| %r11d = dw
\*/
        




#define ENTER           \
        pushq %rbp      ; \
        movq %rsp, %rbp ; \
        pushq %rbx      ; \
        pushq %r13      ; \
        pushq %r14      ; \
        movq %rsi, %r10 ; \
        movq %rcx, %r11 ; \
        movq %rdi, %rsi ; \
        movq %rdx, %rdi ; \
        movq 16(%rbp), %r14 ; \
                        ; \
        /* param sanity check */ ; \
        testq %r8, %r8  ; \
        jz 9f           ; \
        testq %r9, %r9  ; \
        jz 9f
        
#define LEAVE           \
        popq %r14       ; \
        popq %r13       ; \
        popq %rbx       ; \
        movq %rbp, %rsp ; \
        popq %rbp       ; \
        ret


PR_(imlib_amd64_blend_rgba_to_rgb):
        ENTER

        pxor %xmm4, %xmm4
        movdqu c1(%rip), %xmm5
        movdqu m00XXXXXX(%rip), %xmm6

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_blend_rgba_to_rgb)
PR_(imlib_amd64_blend_rgba_to_rgba):
        ENTER

        pxor %xmm4, %xmm4
        movdqu c1(%rip), %xmm5
        xorq %rax, %rax
        movdqu mX000X000X000X000(%rip), %xmm6
        movq pow_lut@GOTPCREL(%rip), %r13

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Load one pixel as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %eax, %xmm3 
        /* override source alpha to 255 */
        por %xmm6, %xmm1

        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* unpack source and dest */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * ((s - d) + 0.5)) */
        psubw %xmm2, %xmm1
        psllw $1, %xmm1
        paddw %xmm5, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* repack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_blend_rgba_to_rgba)
PR_(imlib_amd64_copy_rgba_to_rgb):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5
        movdqu mX000X000X000X000(%rip), %xmm6

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (s & 0x00ffffff) | (d & 0xff000000) */
        pand %xmm5, %xmm1
        pand %xmm6, %xmm2
        por %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_copy_rgba_to_rgb)
PR_(imlib_amd64_copy_rgba_to_rgba):
        ENTER


        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd %xmm1, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd %xmm1, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_copy_rgba_to_rgba)
PR_(imlib_amd64_copy_rgb_to_rgba):
        ENTER

        movdqu mX000X000X000X000(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movd %xmm1, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movdqa %xmm1, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        /* d = s | 0xff000000 */
        por %xmm5, %xmm1
        movd %xmm1, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_copy_rgb_to_rgba)
PR_(imlib_amd64_add_blend_rgba_to_rgb):
        ENTER

        pxor %xmm4, %xmm4
        movdqu m00XXXXXX(%rip), %xmm6

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (a * s) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_add_blend_rgba_to_rgb)

PR_(imlib_amd64_add_blend_rgba_to_rgba):
        ENTER

        pxor %xmm4, %xmm4
        movdqu c1(%rip), %xmm5
        xorq %rax, %rax
        movdqu mX000X000X000X000(%rip), %xmm6
        movq pow_lut@GOTPCREL(%rip), %r13

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Load one pixel as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %eax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        por %xmm6, %xmm1
        pand %xmm6, %xmm0
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (s * ca) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_add_blend_rgba_to_rgba)

PR_(imlib_amd64_add_copy_rgba_to_rgb):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d + (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        paddusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_add_copy_rgba_to_rgb)

PR_(imlib_amd64_add_copy_rgba_to_rgba):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d & 0x00ffffff) + s */
        pand %xmm5, %xmm2
        paddusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_add_copy_rgba_to_rgba)

PR_(imlib_amd64_add_copy_rgb_to_rgba):
        ENTER

        movdqu mX000X000X000X000(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d + s) | 0xff000000 */  
        paddusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_add_copy_rgb_to_rgba)

PR_(imlib_amd64_subtract_blend_rgba_to_rgb):
        ENTER

        pxor %xmm4, %xmm4
        movdqu m00XXXXXX(%rip), %xmm6

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Get alpha from source and unpack to words
         * Result ranges is [0, 0x7fff], and is mapped to
         *  point values in [0.0, 1.0) by using the high word
         *  of the 32 bit multiplication result.
         * Because we want the unsigned value, we shift right one 
         *  here and also shift left the other factors to compensate.
         */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero out the alpha channel of the source to leave the
         * destination alpha unchanged.
         */
        pand %xmm6, %xmm3

        /* Unpack src and dst to words */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - (s * a) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        psubsw %xmm1, %xmm2

        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_subtract_blend_rgba_to_rgb)

PR_(imlib_amd64_subtract_blend_rgba_to_rgba):
        ENTER

        movq pow_lut@GOTPCREL(%rip), %r13
        pxor %xmm4, %xmm4
        movdqu c1(%rip), %xmm5
        movdqu mX000X000X000X000(%rip), %xmm6
        movdqu mX000X000(%rip), %xmm7
        xorq %rax, %rax

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Load two pixels as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %rax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Load one pixel as 00, 00, src alpha, combined alpha 
         * Combined alpha is derived from the pow_lut table in blend.c 
         */
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        movd %eax, %xmm3 
        /* unpack alpha to src alpha, combined alpha x 3 */
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        /* src alpha = 255 - dst alpha */
        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        /* unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d - ((s * a) ^ 0xff000000) */
        psllw $1, %xmm1
        pmulhw %xmm3, %xmm1
        pxor %xmm7, %xmm1
        psubsw %xmm1, %xmm2
        
        /* pack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_subtract_blend_rgba_to_rgba)

PR_(imlib_amd64_subtract_copy_rgba_to_rgb):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d - (s & 0x00ffffff) */
        pand %xmm5, %xmm1
        psubusb %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_subtract_copy_rgba_to_rgb)

PR_(imlib_amd64_subtract_copy_rgba_to_rgba):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5
        movdqu mX000X000X000X000(%rip), %xmm6

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = d - s, d alpha = s alpha */
        psubusb %xmm1, %xmm2
        pand %xmm6, %xmm1
        pand %xmm5, %xmm2
        por %xmm1, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_subtract_copy_rgba_to_rgba)

PR_(imlib_amd64_subtract_copy_rgb_to_rgba):
        ENTER

        movdqu mX000X000X000X000(%rip), %xmm5

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* d = (d - s) | 0xff000000 */  
        psubusb %xmm1, %xmm2
        por %xmm5, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_subtract_copy_rgb_to_rgba)

PR_(imlib_amd64_reshade_blend_rgba_to_rgb):
        ENTER

        pxor %xmm4, %xmm4
        movdqu m000V0V0V000V0V0V(%rip), %xmm6
        movdqu m00XXXXXX(%rip), %xmm7

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* Unpack alpha */
        movq %xmm1, %xmm3
        punpcklbw %xmm3, %xmm3
        pshufhw $0xFF, %xmm3, %xmm3
        pshuflw $0xFF, %xmm3, %xmm3 
        psrlw $1, %xmm3

        /* Zero blending alpha */
        pand %xmm7, %xmm3

        /* Unpack src and dst */
        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        /* d = d + (2 * a * (s - 127)) */
        psubw %xmm6, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2

        /* Repack new pixels */
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_reshade_blend_rgba_to_rgb)

PR_(imlib_amd64_reshade_blend_rgba_to_rgba):
        ENTER

        movq pow_lut@GOTPCREL(%rip), %r13
        pxor %xmm4, %xmm4
        movdqu c1(%rip), %xmm5
        movdqu mX000X000X000X000(%rip), %xmm6
        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm7
        movdqu m000V0V0V000V0V0V(%rip), %xmm8
        xorq %rax, %rax

        /* Move right to left across each line, */ 
        /* processing in two pixel chunks */ 
        leaq (%rsi, %r8, 4), %rsi       
        leaq (%rdi, %r8, 4), %rdi       
                                        
        /* Last instruction is %rcx = 0 */ 
        subq $4, %rsi                   
        subq $4, %rdi                   
                                        
        negq %r8                        
0:                                      
        movq %r8, %rcx                  
                                        
        incq %rcx                       

        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        jz 2f /* one pixel line */      
1:
        /* main loop, unrolled to work on 64 byte chunks */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        jz 2f
        jns 3f                          

        movq (%rsi, %rcx, 4), %xmm1
        movq (%rdi, %rcx, 4), %xmm2
        movzbq 7(%rdi, %rcx, 4), %rdx
        movb 7(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        shlq $32, %rax
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %rax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movq %xmm2, (%rdi, %rcx, 4)

        incq %rcx                       
        incq %rcx                       
        js 1b                           
        jnz 3f                          
2:
        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        movzbq 3(%rdi, %rcx, 4), %rdx
        movb 3(%rsi, %rcx, 4), %dh
        movb (%r13, %rdx), %al
        movb %dh, %ah
        shrb $1, %ah
        movd %eax, %xmm3 
        punpcklbw %xmm3, %xmm3
        pshufhw $0x40, %xmm3, %xmm3
        pshuflw $0x40, %xmm3, %xmm3     
        psrlw $1, %xmm3

        movdqa %xmm2, %xmm0
        pand %xmm6, %xmm0
        por %xmm6, %xmm1
        psubusb %xmm0, %xmm1

        punpcklbw %xmm4, %xmm1
        punpcklbw %xmm4, %xmm2

        psubw %xmm8, %xmm1
        psllw $2, %xmm1
        pmulhw %xmm3, %xmm1
        paddsw %xmm1, %xmm2
        
        packuswb %xmm4, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
3:                                      
        leaq (%rsi, %r10, 4), %rsi      
        leaq (%rdi, %r11, 4), %rdi      
        decq %r9                        
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_reshade_blend_rgba_to_rgba)

PR_(imlib_amd64_reshade_copy_rgba_to_rgb):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5
        movdqu m0VVV0VVV0VVV0VVV(%rip), %xmm6

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* To take advantage of saturation and be able to do 8 bytes
         *  at a time, we divide reshading into two separate steps:
         *  adding values above 128, and subtracting values below 128
         * These values go into %mm1 and %mm3 respectively
         * - %xmm1 becomes (2 * (s - 127))
         * - %xmm3 becomes (2 * (255 - (127 + s))) = (2 * (128 - s))
         */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* dest alpha should not be changed in this func */
        pand %xmm5, %xmm1
        pand %xmm5, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_reshade_copy_rgba_to_rgb)

PR_(imlib_amd64_reshade_copy_rgba_to_rgba):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5
        movdqu m0VVV0VVV0VVV0VVV(%rip), %xmm6
        movdqu mX000X000X000X000(%rip), %xmm7

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        /* This time, dest alpha = src alpha, so src alpha is saved in %xmm0 */
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        movdqa %xmm1, %xmm0
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2
        
        /* d alpha = s alpha */
        pand %xmm5, %xmm2
        pand %xmm7, %xmm0
        por %xmm0, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_reshade_copy_rgba_to_rgba)

PR_(imlib_amd64_reshade_copy_rgb_to_rgba):
        ENTER

        movdqu m0XXX0XXX0XXX0XXX(%rip), %xmm5
        movdqu m0VVV0VVV0VVV0VVV(%rip), %xmm6
        movdqu mX000X000X000X000(%rip), %xmm7

        leaq (%rsi, %r8, 4), %rsi
        leaq (%rdi, %r8, 4), %rdi

        subq $12, %rsi
        subq $12, %rdi

        negq %r8
0:
        movq %r8, %rcx

        /* if < 4 pixels left, goto end */
        addq $3, %rcx
        jns 4f
1:
        /* 16 byte align dst ptr */
        leaq (%rdi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jz 1f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jz 4f
        jmp 1b
1:
        /* prefetch a couple cache lines ahead */
        prefetchnta (%rsi, %rcx, 4)
        prefetcht0 (%rdi, %rcx, 4)
        prefetchnta 64(%rsi, %rcx, 4)
        prefetcht0 64(%rdi, %rcx, 4)

        /* test if 16 byte aligned src ptr */
        leaq (%rsi, %rcx, 4), %rdx
        test $0x0f, %rdx
        jnz 3f
2:
        /* main loop, unrolled to work on 64 byte chunks */
        /* aligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqa (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 2b
        jmp 4f
3:
        /* main loop, unrolled to work on 64 byte chunks */
        /* unaligned src, aligned dst */
        prefetchnta 128(%rsi, %rcx, 4)
        prefetcht0 128(%rdi, %rcx, 4)

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        jns 4f

        movdqu (%rsi, %rcx, 4), %xmm1
        movdqa (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movdqa %xmm2, (%rdi, %rcx, 4)
        addq $4, %rcx
        js 3b
4:
        /* finish loop */
        cmp $2, %rcx
        jg 5f

        movd (%rsi, %rcx, 4), %xmm1
        movd (%rdi, %rcx, 4), %xmm2
        movdqa %xmm1, %xmm3
        psubusb %xmm6, %xmm1
        paddusb %xmm1, %xmm1
        paddusb %xmm6, %xmm3
        pxor %xmm5, %xmm3
        paddusb %xmm3, %xmm3

        /* d = d + s1 - s2, unsigned saturation */
        paddusb %xmm1, %xmm2
        psubusb %xmm3, %xmm2

        /* d alpha = 0xff */
        por %xmm7, %xmm2
        movd %xmm2, (%rdi, %rcx, 4)
        incq %rcx
        jmp 4b
5:
        /* finish line */
        leaq (%rsi, %r10, 4), %rsi
        leaq (%rdi, %r11, 4), %rdi
        decq %r9
        jnz 0b

9:
        LEAVE
SIZE(imlib_amd64_reshade_copy_rgb_to_rgba)

#endif

#ifdef __ELF__
.section .note.GNU-stack,"",@progbits
#endif

/* [<][>][^][v][top][bottom][index][help] */