Skip to content

Commit f9c8d18

Browse files
floodyberryhackmod
authored andcommitted
avx uop fusing
1 parent 45a5305 commit f9c8d18

File tree

3 files changed

+5
-5
lines changed

3 files changed

+5
-5
lines changed

src/scryptjane/scrypt-jane-mix_chacha-avx.h

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -91,7 +91,6 @@ asm_naked_fn(scrypt_ChunkMix_avx)
9191
a3(vpsrld xmm6,xmm1,25)
9292
a3(vpslld xmm1,xmm1,7)
9393
a3(vpxor xmm1,xmm1,xmm6)
94-
a2(sub eax,2)
9594
a3(vpaddd xmm0,xmm0,xmm1)
9695
a3(vpxor xmm3,xmm3,xmm0)
9796
a3(vpshufb xmm3,xmm3,xmm4)
@@ -111,6 +110,7 @@ asm_naked_fn(scrypt_ChunkMix_avx)
111110
a3(vpsrld xmm6,xmm1,25)
112111
a3(vpslld xmm1,xmm1,7)
113112
a3(vpxor xmm1,xmm1,xmm6)
113+
a2(sub eax,2)
114114
aj(ja scrypt_chacha_avx_loop)
115115
a3(vpaddd xmm0,xmm0,[esp+0])
116116
a3(vpaddd xmm1,xmm1,[esp+16])
@@ -213,7 +213,6 @@ asm_naked_fn(scrypt_ChunkMix_avx)
213213
a3(vpsrld xmm12,xmm1,25)
214214
a3(vpslld xmm1,xmm1,7)
215215
a3(vpxor xmm1,xmm1,xmm12)
216-
a2(sub rax,2)
217216
a3(vpaddd xmm0,xmm0,xmm1)
218217
a3(vpxor xmm3,xmm3,xmm0)
219218
a3(vpshufb xmm3,xmm3,xmm4)
@@ -233,6 +232,7 @@ asm_naked_fn(scrypt_ChunkMix_avx)
233232
a3(vpsrld xmm12,xmm1,25)
234233
a3(vpslld xmm1,xmm1,7)
235234
a3(vpxor xmm1,xmm1,xmm12)
235+
a2(sub rax,2)
236236
aj(ja scrypt_chacha_avx_loop)
237237
a3(vpaddd xmm0,xmm0,xmm8)
238238
a3(vpaddd xmm1,xmm1,xmm9)

src/scryptjane/scrypt-jane-mix_salsa-avx.h

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -72,7 +72,6 @@ asm_naked_fn(scrypt_ChunkMix_avx)
7272
a3(vpxor xmm0, xmm0, xmm5)
7373
a3(pshufd xmm2, xmm2, 0x4e)
7474
a3(vpxor xmm0, xmm0, xmm4)
75-
a2(sub eax, 2)
7675
a3(vpaddd xmm4, xmm3, xmm0)
7776
a3(pshufd xmm1, xmm1, 0x39)
7877
a3(vpsrld xmm5, xmm4, 25)
@@ -97,6 +96,7 @@ asm_naked_fn(scrypt_ChunkMix_avx)
9796
a3(pshufd xmm2, xmm2, 0x4e)
9897
a3(vpxor xmm0, xmm0, xmm4)
9998
a3(pshufd xmm3, xmm3, 0x39)
99+
a2(sub eax, 2)
100100
aj(ja scrypt_salsa_avx_loop)
101101
a3(vpaddd xmm0,xmm0,[esp+0])
102102
a3(vpaddd xmm1,xmm1,[esp+16])
@@ -192,7 +192,6 @@ asm_naked_fn(scrypt_ChunkMix_avx)
192192
a3(vpxor xmm0, xmm0, xmm5)
193193
a3(pshufd xmm2, xmm2, 0x4e)
194194
a3(vpxor xmm0, xmm0, xmm4)
195-
a2(sub rax, 2)
196195
a3(vpaddd xmm4, xmm3, xmm0)
197196
a3(pshufd xmm1, xmm1, 0x39)
198197
a3(vpsrld xmm5, xmm4, 25)
@@ -217,6 +216,7 @@ asm_naked_fn(scrypt_ChunkMix_avx)
217216
a3(pshufd xmm2, xmm2, 0x4e)
218217
a3(vpxor xmm0, xmm0, xmm4)
219218
a3(pshufd xmm3, xmm3, 0x39)
219+
a2(sub rax, 2)
220220
aj(ja scrypt_salsa_avx_loop)
221221
a3(vpaddd xmm0,xmm0,xmm8)
222222
a3(vpaddd xmm1,xmm1,xmm9)

src/scryptjane/scrypt-jane-mix_salsa64-avx.h

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -103,7 +103,6 @@ asm_naked_fn(scrypt_ChunkMix_avx)
103103
a4(vpalignr xmm3, xmm7, xmm6, 8)
104104
a4(vpalignr xmm6, xmm9, xmm8, 8)
105105
a4(vpalignr xmm7, xmm8, xmm9, 8)
106-
a2(sub rax, 2)
107106
a3(vpaddq xmm10, xmm0, xmm2)
108107
a3(vpaddq xmm11, xmm1, xmm3)
109108
a3(vpshufd xmm10, xmm10, 0xb1)
@@ -142,6 +141,7 @@ asm_naked_fn(scrypt_ChunkMix_avx)
142141
a4(vpalignr xmm3, xmm7, xmm6, 8)
143142
a4(vpalignr xmm6, xmm11, xmm10, 8)
144143
a4(vpalignr xmm7, xmm10, xmm11, 8)
144+
a2(sub rax, 2)
145145
aj(ja scrypt_salsa64_avx_loop)
146146
a3(vpaddq xmm0,xmm0,[rsp+0])
147147
a3(vpaddq xmm1,xmm1,[rsp+16])

0 commit comments

Comments
 (0)