Inside Bitcoin

1251    mov y0, a       ; y0 = a
1252    add h, y2       ; h = h + S1 + CH + k + w
1253    mov y2, a       ; y2 = a
1254    pshufb  XTMP4, SHUF_00BA    ; XTMP4 = s1 {00BA}
1255    or  y0, c       ; y0 = a|c
1256    add d, h        ; d = d + h + S1 + CH + k + w
1257    and y2, c       ; y2 = a&c
1258    paddd   XTMP0, XTMP4    ; XTMP0 = {..., ..., W[1], W[0]}
1259    and y0, b       ; y0 = (a|c)&b
1260    add h, y1       ; h = h + S1 + CH + k + w + S0
1261    ;; compute high s1
1262    pshufd  XTMP2, XTMP0, 01010000b ; XTMP2 = W[-2] {DDCC}
1263    or  y0, y2      ; y0 = MAJ = (a|c)&b)|(a&c)
1264    add h, y0       ; h = h + S1 + CH + k + w + S0 + MAJ
1265
1266ROTATE_ARGS
1267    movdqa  XTMP3, XTMP2    ; XTMP3 = W[-2] {DDCC}
1268    mov y0, e       ; y0 = e
1269    ror y0, (25-11) ; y0 = e >> (25-11)
1270    mov y1, a       ; y1 = a
1271    movdqa  X0,    XTMP2    ; X0    = W[-2] {DDCC}
1272    ror y1, (22-13) ; y1 = a >> (22-13)
1273    xor y0, e       ; y0 = e ^ (e >> (25-11))
1274    mov y2, f       ; y2 = f
1275    ror y0, (11-6)  ; y0 = (e >> (11-6)) ^ (e >> (25-6))
1276    psrlq   XTMP2, 17   ; XTMP2 = W[-2] ror 17 {xDxC}
1277    xor y1, a       ; y1 = a ^ (a >> (22-13)
1278    xor y2, g       ; y2 = f^g
1279    psrlq   XTMP3, 19   ; XTMP3 = W[-2] ror 19 {xDxC}
1280    xor y0, e       ; y0 = e ^ (e >> (11-6)) ^ (e >> (25-6))
1281    and y2, e       ; y2 = (f^g)&e
1282    ror y1, (13-2)  ; y1 = (a >> (13-2)) ^ (a >> (22-2))
1283    psrld   X0,    10   ; X0 = W[-2] >> 10 {DDCC}
1284    xor y1, a       ; y1 = a ^ (a >> (13-2)) ^ (a >> (22-2))
1285    ror y0, 6       ; y0 = S1 = (e>>6) & (e>>11) ^ (e>>25)
1286    xor y2, g       ; y2 = CH = ((f^g)&e)^g
1287    pxor    XTMP2, XTMP3
1288    ror y1, 2       ; y1 = S0 = (a>>2) ^ (a>>13) ^ (a>>22)
1289    add y2, y0      ; y2 = S1 + CH
1290    add y2, [rsp + _XFER + 3*4] ; y2 = k + w + S1 + CH
1291    pxor    X0, XTMP2   ; X0 = s1 {xDxC}
1292    mov y0, a       ; y0 = a
1293    add h, y2       ; h = h + S1 + CH + k + w
1294    mov y2, a       ; y2 = a
1295    pshufb  X0, SHUF_DC00   ; X0 = s1 {DC00}
1296    or  y0, c       ; y0 = a|c
1297    add d, h        ; d = d + h + S1 + CH + k + w
1298    and y2, c       ; y2 = a&c
1299    paddd   X0, XTMP0   ; X0 = {W[3], W[2], W[1], W[0]}
1300    and y0, b       ; y0 = (a|c)&b
1301    add h, y1       ; h = h + S1 + CH + k + w + S0
1302    or  y0, y2      ; y0 = MAJ = (a|c)&b)|(a&c)
1303    add h, y0       ; h = h + S1 + CH + k + w + S0 + MAJ
1304
1305ROTATE_ARGS
1306rotate_Xs
1307%endm
1308
1309;; input is [rsp + _XFER + %1 * 4]
1310%macro DO_ROUND 1
1311    mov y0, e       ; y0 = e
1312    ror y0, (25-11) ; y0 = e >> (25-11)
1313    mov y1, a       ; y1 = a
1314    xor y0, e       ; y0 = e ^ (e >> (25-11))
1315    ror y1, (22-13) ; y1 = a >> (22-13)
1316    mov y2, f       ; y2 = f
1317    xor y1, a       ; y1 = a ^ (a >> (22-13)
1318    ror y0, (11-6)  ; y0 = (e >> (11-6)) ^ (e >> (25-6))
1319    xor y2, g       ; y2 = f^g
1320    xor y0, e       ; y0 = e ^ (e >> (11-6)) ^ (e >> (25-6))
1321    ror y1, (13-2)  ; y1 = (a >> (13-2)) ^ (a >> (22-2))
1322    and y2, e       ; y2 = (f^g)&e
1323    xor y1, a       ; y1 = a ^ (a >> (13-2)) ^ (a >> (22-2))
1324    ror y0, 6       ; y0 = S1 = (e>>6) & (e>>11) ^ (e>>25)
1325    xor y2, g       ; y2 = CH = ((f^g)&e)^g
1326    add y2, y0      ; y2 = S1 + CH
1327    ror y1, 2       ; y1 = S0 = (a>>2) ^ (a>>13) ^ (a>>22)
1328    add y2, [rsp + _XFER + %1 * 4]  ; y2 = k + w + S1 + CH
1329    mov y0, a       ; y0 = a
1330    add h, y2       ; h = h + S1 + CH + k + w
1331    mov y2, a       ; y2 = a
1332    or  y0, c       ; y0 = a|c
1333    add d, h        ; d = d + h + S1 + CH + k + w
1334    and y2, c       ; y2 = a&c
1335    and y0, b       ; y0 = (a|c)&b
1336    add h, y1       ; h = h + S1 + CH + k + w + S0
1337    or  y0, y2      ; y0 = MAJ = (a|c)&b)|(a&c)
1338    add h, y0       ; h = h + S1 + CH + k + w + S0 + MAJ
1339    ROTATE_ARGS
1340%endm
1341
1342;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
1343;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
1344;; void sha256_sse4(void *input_data, UINT32 digest[8], UINT64 num_blks)
1345;; arg 1 : pointer to input data
1346;; arg 2 : pointer to digest
1347;; arg 3 : Num blocks
1348section .text
1349global sha256_sse4
1350align 32
1351sha256_sse4:
1352    push    rbx
1353%ifndef LINUX
1354    push    rsi
1355    push    rdi
1356%endif
1357    push    rbp
1358    push    r13
1359    push    r14
1360    push    r15
1361
1362    sub rsp,STACK_SIZE
1363%ifndef LINUX
1364    movdqa  [rsp + _XMM_SAVE + 0*16],xmm6
1365    movdqa  [rsp + _XMM_SAVE + 1*16],xmm7
1366    movdqa  [rsp + _XMM_SAVE + 2*16],xmm8
1367    movdqa  [rsp + _XMM_SAVE + 3*16],xmm9
1368    movdqa  [rsp + _XMM_SAVE + 4*16],xmm10
1369    movdqa  [rsp + _XMM_SAVE + 5*16],xmm11
1370    movdqa  [rsp + _XMM_SAVE + 6*16],xmm12
1371%endif
1372
1373    shl NUM_BLKS, 6 ; convert to bytes
1374    jz  done_hash
1375    add NUM_BLKS, INP   ; pointer to end of data
1376    mov [rsp + _INP_END], NUM_BLKS
1377
1378    ;; load initial digest
1379    mov a,[4*0 + CTX]
1380    mov b,[4*1 + CTX]
1381    mov c,[4*2 + CTX]
1382    mov d,[4*3 + CTX]
1383    mov e,[4*4 + CTX]
1384    mov f,[4*5 + CTX]
1385    mov g,[4*6 + CTX]
1386    mov h,[4*7 + CTX]
1387
1388    movdqa  BYTE_FLIP_MASK, [PSHUFFLE_BYTE_FLIP_MASK wrt rip]
1389    movdqa  SHUF_00BA, [_SHUF_00BA wrt rip]
1390    movdqa  SHUF_DC00, [_SHUF_DC00 wrt rip]
1391
1392loop0:
1393    lea TBL,[K256 wrt rip]
1394
1395    ;; byte swap first 16 dwords
1396    COPY_XMM_AND_BSWAP  X0, [INP + 0*16], BYTE_FLIP_MASK
1397    COPY_XMM_AND_BSWAP  X1, [INP + 1*16], BYTE_FLIP_MASK
1398    COPY_XMM_AND_BSWAP  X2, [INP + 2*16], BYTE_FLIP_MASK
1399    COPY_XMM_AND_BSWAP  X3, [INP + 3*16], BYTE_FLIP_MASK
1400
1401    mov [rsp + _INP], INP
1402
1403    ;; schedule 48 input dwords, by doing 3 rounds of 16 each
1404    mov SRND, 3
1405align 16
1406loop1:
1407    movdqa  XFER, [TBL + 0*16]
1408    paddd   XFER, X0
1409    movdqa  [rsp + _XFER], XFER
1410    FOUR_ROUNDS_AND_SCHED
1411
1412    movdqa  XFER, [TBL + 1*16]
1413    paddd   XFER, X0
1414    movdqa  [rsp + _XFER], XFER
1415    FOUR_ROUNDS_AND_SCHED
1416
1417    movdqa  XFER, [TBL + 2*16]
1418    paddd   XFER, X0
1419    movdqa  [rsp + _XFER], XFER
1420    FOUR_ROUNDS_AND_SCHED
1421
1422    movdqa  XFER, [TBL + 3*16]
1423    paddd   XFER, X0
1424    movdqa  [rsp + _XFER], XFER
1425    add TBL, 4*16
1426    FOUR_ROUNDS_AND_SCHED
1427
1428    sub SRND, 1
1429    jne loop1
1430
1431    mov SRND, 2
1432loop2:
1433    paddd   X0, [TBL + 0*16]
1434    movdqa  [rsp + _XFER], X0
1435    DO_ROUND    0
1436    DO_ROUND    1
1437    DO_ROUND    2
1438    DO_ROUND    3
1439    paddd   X1, [TBL + 1*16]
1440    movdqa  [rsp + _XFER], X1
1441    add TBL, 2*16
1442    DO_ROUND    0
1443    DO_ROUND    1
1444    DO_ROUND    2
1445    DO_ROUND    3
1446
1447    movdqa  X0, X2
1448    movdqa  X1, X3
1449
1450    sub SRND, 1
1451    jne loop2
1452
1453    addm    [4*0 + CTX],a
1454    addm    [4*1 + CTX],b
1455    addm    [4*2 + CTX],c
1456    addm    [4*3 + CTX],d
1457    addm    [4*4 + CTX],e
1458    addm    [4*5 + CTX],f
1459    addm    [4*6 + CTX],g
1460    addm    [4*7 + CTX],h
1461
1462    mov INP, [rsp + _INP]
1463    add INP, 64
1464    cmp INP, [rsp + _INP_END]
1465    jne loop0
1466
1467done_hash:
1468%ifndef LINUX
1469    movdqa  xmm6,[rsp + _XMM_SAVE + 0*16]
1470    movdqa  xmm7,[rsp + _XMM_SAVE + 1*16]
1471    movdqa  xmm8,[rsp + _XMM_SAVE + 2*16]
1472    movdqa  xmm9,[rsp + _XMM_SAVE + 3*16]
1473    movdqa  xmm10,[rsp + _XMM_SAVE + 4*16]
1474    movdqa  xmm11,[rsp + _XMM_SAVE + 5*16]
1475    movdqa  xmm12,[rsp + _XMM_SAVE + 6*16]
1476%endif
1477
1478    add rsp, STACK_SIZE
1479
1480    pop r15
1481    pop r14
1482    pop r13
1483    pop rbp
1484%ifndef LINUX
1485    pop rdi
1486    pop rsi
1487%endif
1488    pop rbx
1489
1490    ret
1491
1492
1493section .data
1494align 64
1495K256:
1496    dd  0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
1497    dd  0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
1498    dd  0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
1499    dd  0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
1500    dd  0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc