Lines Matching +full:16 +full:g
111 g = %r10d define
121 _XFER_SIZE = 16
144 h = g
145 g = f define
156 ## compute W[-16] + W[-7] 4 at a time
167 xor g, y2 # y2 = f^g
168 vpaddd X0, XTMP0, XTMP0 # XTMP0 = W[-7] + W[-16]
170 and e, y2 # y2 = (f^g)&e
176 xor g, y2 # y2 = CH = ((f^g)&e)^g
203 xor g, y2 # y2 = f^g
207 and e, y2 # y2 = (f^g)&e
211 xor g, y2 # y2 = CH = ((f^g)&e)^g
228 vpaddd XTMP1, XTMP0, XTMP0 # XTMP0 = W[-16] + W[-7] + s0
241 xor g, y2 # y2 = f^g
244 and e, y2 # y2 = (f^g)&e
248 xor g, y2 # y2 = CH = ((f^g)&e)^g
279 xor g, y2 # y2 = f^g
282 and e, y2 # y2 = (f^g)&e
287 xor g, y2 # y2 = CH = ((f^g)&e)^g
319 xor g, y2 # y2 = f^g
322 and e, y2 # y2 = (f^g)&e
325 xor g, y2 # y2 = CH = ((f^g)&e)^g
374 mov 4*6(CTX), g
383 ## byte swap first 16 dwords
384 COPY_XMM_AND_BSWAP X0, 0*16(INP), BYTE_FLIP_MASK
385 COPY_XMM_AND_BSWAP X1, 1*16(INP), BYTE_FLIP_MASK
386 COPY_XMM_AND_BSWAP X2, 2*16(INP), BYTE_FLIP_MASK
387 COPY_XMM_AND_BSWAP X3, 3*16(INP), BYTE_FLIP_MASK
391 ## schedule 48 input dwords, by doing 3 rounds of 16 each
393 .align 16
399 vpaddd 1*16(TBL), X0, XFER
403 vpaddd 2*16(TBL), X0, XFER
407 vpaddd 3*16(TBL), X0, XFER
409 add $4*16, TBL
424 vpaddd 1*16(TBL), X1, XFER
426 add $2*16, TBL
444 addm (4*6)(CTX),g
484 .section .rodata.cst16.PSHUFFLE_BYTE_FLIP_MASK, "aM", @progbits, 16
485 .align 16
489 .section .rodata.cst16._SHUF_00BA, "aM", @progbits, 16
490 .align 16
495 .section .rodata.cst16._SHUF_DC00, "aM", @progbits, 16
496 .align 16