crypto/intel/sha1edp.s

101 #define	sp			%rsp							// unifying architectural stack pointer representation
113 #define	sp			%esp							// unifying architectural stack pointer representation
114 #define HASH_PTR	stack_size+16+4(sp)				// use 1st input argument from caller function, 16 for (esi/edi/ebx/ebp)
115 #define BUFFER_PTR	stack_size+16+8(sp)				// use 2nd input argument from caller function
116 #define cnt			stack_size+16+12(sp)			// use 3rd input argument from caller function
117 #define K_BASE		stack_size-4(sp)				// use for K_BASE
136 #define W24     12*16(sp)
137 #define W28     13*16(sp)
138 #define XMM_SHUFB_BSWAP 14*16(sp)			// used only when ssse3 is supported
155 #define	WK(t)	(t&15)*4(sp)
249 	mov		T1, 14*16(sp)
254 	mov		T1, 4+14*16(sp)
259 	mov		T1, 8+14*16(sp)
264 	mov		T1, 12+14*16(sp)
268 	xmov	14*16(sp), W_TMP				// load the bswapped 16-bytes from the aligned stack memory
534 	W_PRECALC_00_15_3	3					// (sp) = W_TMP = W0 + K
540 	W_PRECALC_00_15_3	7					// 16(sp) = W_TMP = W28 + K
546 	W_PRECALC_00_15_3	11					// 32(sp) = W_TMP = W24 + K
552 	W_PRECALC_00_15_3	15					// 48(sp) = W_TMP = W20 + K
771 	W_PRECALC_00_15_3	3					// (sp) = W_TMP = W0 + K
781 	W_PRECALC_00_15_3	7					// 16(sp) = W_TMP = W28 + K[0]
791 	W_PRECALC_00_15_3	11					// 32(sp) = W_TMP = W24 + K
801 	W_PRECALC_00_15_3	15					// 48(sp) = W_TMP = W20 + K
839 	W_PRECALC_00_15_3	3					// (sp) = W_TMP = W0 + K
845 	W_PRECALC_00_15_3	7					// 16(sp) = W_TMP = W28 + K
851 	W_PRECALC_00_15_3	11					// 32(sp) = W_TMP = W24 + K
857 	W_PRECALC_00_15_3	15					// 48(sp) = W_TMP = W20 + K
1090 	W_PRECALC_00_15_3	3					// (sp) = W_TMP = W0 + K
1100 	W_PRECALC_00_15_3	7					// 16(sp) = W_TMP = W28 + K[0]
1110 	W_PRECALC_00_15_3	11					// 32(sp) = W_TMP = W24 + K
1120 	W_PRECALC_00_15_3	15					// 48(sp) = W_TMP = W20 + K
1290 	sub		$stack_size, sp					// allocate stack memory for use
1294 	xmov	%xmm0, 4*16(sp)
1295 	xmov	%xmm1, 5*16(sp)
1296 	xmov	%xmm2, 6*16(sp)
1297 	xmov	%xmm3, 7*16(sp)
1298 	xmov	%xmm4, 8*16(sp)
1299 	xmov	%xmm5, 9*16(sp)
1300 	xmov	%xmm6, 10*16(sp)
1301 	xmov	%xmm7, 11*16(sp)
1303 	xmov	%xmm8, 12*16(sp)
1304 	xmov	%xmm9, 13*16(sp)
1305 	xmov	%xmm10, 14*16(sp)
1340 	xmov	4*16(sp), %xmm0
1341 	xmov	5*16(sp), %xmm1
1342 	xmov	6*16(sp), %xmm2
1343 	xmov	7*16(sp), %xmm3
1344 	xmov	8*16(sp), %xmm4
1345 	xmov	9*16(sp), %xmm5
1346 	xmov	10*16(sp), %xmm6
1347 	xmov	11*16(sp), %xmm7
1349 	xmov	12*16(sp), %xmm8
1350 	xmov	13*16(sp), %xmm9
1351 	xmov	14*16(sp), %xmm10
1355 	add		$stack_size, sp		// deallocate stack memory
1387 	sub		$stack_size, sp			// allocate stack memory for local use
1391 	xmov	%xmm0, 4*16(sp)
1392 	xmov	%xmm1, 5*16(sp)
1393 	xmov	%xmm2, 6*16(sp)
1394 	xmov	%xmm3, 7*16(sp)
1395 	xmov	%xmm4, 8*16(sp)
1396 	xmov	%xmm5, 9*16(sp)
1397 	xmov	%xmm6, 10*16(sp)
1398 	xmov	%xmm7, 11*16(sp)
1400 	xmov	%xmm8, 12*16(sp)
1401 	xmov	%xmm9, 13*16(sp)
1433 	xmov	4*16(sp), %xmm0
1434 	xmov	5*16(sp), %xmm1
1435 	xmov	6*16(sp), %xmm2
1436 	xmov	7*16(sp), %xmm3
1437 	xmov	8*16(sp), %xmm4
1438 	xmov	9*16(sp), %xmm5
1439 	xmov	10*16(sp), %xmm6
1440 	xmov	11*16(sp), %xmm7
1442 	xmov	12*16(sp), %xmm8
1443 	xmov	13*16(sp), %xmm9
1447 	add		$stack_size, sp		// deallocate stack memory