Quoting Borislav Petkov <bp@xxxxxxxxx>: > > Yep, looks better than the previous run and also a bit better or on par > with the initial run I did. > I made few further changes, mainly moving/interleaving 'vmovq/vpextrq' ahead so they should be completed before those target registers are needed. This only gave 0.5% increase on Sandy-bridge, but might help more on Bulldozer. -Jussi --- arch/x86/crypto/twofish-avx-x86_64-asm_64.S | 205 +++++++++++++++++---------- 1 file changed, 130 insertions(+), 75 deletions(-) diff --git a/arch/x86/crypto/twofish-avx-x86_64-asm_64.S b/arch/x86/crypto/twofish-avx-x86_64-asm_64.S index 35f4557..6638a87 100644 --- a/arch/x86/crypto/twofish-avx-x86_64-asm_64.S +++ b/arch/x86/crypto/twofish-avx-x86_64-asm_64.S @@ -4,6 +4,8 @@ * Copyright (C) 2012 Johannes Goetzfried * <Johannes.Goetzfried@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> * + * Copyright © 2012 Jussi Kivilinna <jussi.kivilinna@xxxxxxxx> + * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation; either version 2 of the License, or @@ -47,16 +49,21 @@ #define RC2 %xmm6 #define RD2 %xmm7 -#define RX %xmm8 -#define RY %xmm9 +#define RX0 %xmm8 +#define RY0 %xmm9 + +#define RX1 %xmm10 +#define RY1 %xmm11 -#define RK1 %xmm10 -#define RK2 %xmm11 +#define RK1 %xmm12 +#define RK2 %xmm13 -#define RID1 %rax -#define RID1b %al -#define RID2 %rbx -#define RID2b %bl +#define RT %xmm14 + +#define RID1 %rbp +#define RID1d %ebp +#define RID2 %rsi +#define RID2d %esi #define RGI1 %rdx #define RGI1bl %dl @@ -65,6 +72,13 @@ #define RGI2bl %cl #define RGI2bh %ch +#define RGI3 %rax +#define RGI3bl %al +#define RGI3bh %ah +#define RGI4 %rbx +#define RGI4bl %bl +#define RGI4bh %bh + #define RGS1 %r8 #define RGS1d %r8d #define RGS2 %r9 @@ -73,40 +87,53 @@ #define RGS3d %r10d -#define lookup_32bit(t0, t1, t2, t3, src, dst) \ - movb src ## bl, RID1b; \ - movb src ## bh, RID2b; \ +#define lookup_32bit(t0, t1, t2, t3, src, dst, interleave_op, il_reg) \ + movzbl src ## bl, RID1d; \ + movzbl src ## bh, RID2d; \ + shrq $16, src; \ movl t0(CTX, RID1, 4), dst ## d; \ xorl t1(CTX, RID2, 4), dst ## d; \ - shrq $16, src; \ - movb src ## bl, RID1b; \ - movb src ## bh, RID2b; \ + movzbl src ## bl, RID1d; \ + movzbl src ## bh, RID2d; \ + interleave_op(il_reg); \ xorl t2(CTX, RID1, 4), dst ## d; \ xorl t3(CTX, RID2, 4), dst ## d; -#define G(a, x, t0, t1, t2, t3) \ - vmovq a, RGI1; \ - vpsrldq $8, a, x; \ - vmovq x, RGI2; \ - \ - lookup_32bit(t0, t1, t2, t3, RGI1, RGS1); \ - shrq $16, RGI1; \ - lookup_32bit(t0, t1, t2, t3, RGI1, RGS2); \ - shlq $32, RGS2; \ - orq RGS1, RGS2; \ +#define dummy(d) /* do nothing */ + +#define shr_next(reg) \ + shrq $16, reg; + +#define G(gi1, gi2, x, t0, t1, t2, t3) \ + lookup_32bit(t0, t1, t2, t3, ##gi1, RGS1, shr_next, ##gi1); \ + lookup_32bit(t0, t1, t2, t3, ##gi1, RGS2, dummy, none); \ + shlq $32, RGS2; \ + orq RGS1, RGS2; \ \ - lookup_32bit(t0, t1, t2, t3, RGI2, RGS1); \ - shrq $16, RGI2; \ - lookup_32bit(t0, t1, t2, t3, RGI2, RGS3); \ - shlq $32, RGS3; \ - orq RGS1, RGS3; \ + lookup_32bit(t0, t1, t2, t3, ##gi2, RGS3, shr_next, ##gi2); \ + lookup_32bit(t0, t1, t2, t3, ##gi2, RGS1, dummy, none); \ + shlq $32, RGS1; \ + orq RGS1, RGS3; \ \ - vmovq RGS2, x; \ + vmovq RGS2, x; \ vpinsrq $1, RGS3, x, x; -#define encround(a, b, c, d, x, y) \ - G(a, x, s0, s1, s2, s3); \ - G(b, y, s1, s2, s3, s0); \ +#define encround_head_2(a, b, c, d, x1, y1, x2, y2) \ + vmovq b ## 1, RGI3; \ + vpextrq $1, b ## 1, RGI4; \ + G(RGI1, RGI2, x1, s0, s1, s2, s3); \ + vmovq a ## 2, RGI1; \ + vpextrq $1, a ## 2, RGI2; \ + G(RGI3, RGI4, y1, s1, s2, s3, s0); \ + vmovq b ## 2, RGI3; \ + vpextrq $1, b ## 2, RGI4; \ + G(RGI1, RGI2, x2, s0, s1, s2, s3); \ + G(RGI3, RGI4, y2, s1, s2, s3, s0); + +#define encround_tail(a, b, c, d, x, y) \ + vpslld $1, d, RT; \ + vpsrld $(32 - 1), d, d; \ + vpor d, RT, d; \ vpaddd x, y, x; \ vpaddd y, x, y; \ vpaddd x, RK1, x; \ @@ -115,14 +142,24 @@ vpsrld $1, c, x; \ vpslld $(32 - 1), c, c; \ vpor c, x, c; \ - vpslld $1, d, x; \ - vpsrld $(32 - 1), d, d; \ - vpor d, x, d; \ vpxor d, y, d; -#define decround(a, b, c, d, x, y) \ - G(a, x, s0, s1, s2, s3); \ - G(b, y, s1, s2, s3, s0); \ +#define decround_head_2(a, b, c, d, x1, y1, x2, y2) \ + vmovq b ## 1, RGI3; \ + vpextrq $1, b ## 1, RGI4; \ + G(RGI1, RGI2, x1, s0, s1, s2, s3); \ + vmovq a ## 2, RGI1; \ + vpextrq $1, a ## 2, RGI2; \ + G(RGI3, RGI4, y1, s1, s2, s3, s0); \ + vmovq b ## 2, RGI3; \ + vpextrq $1, b ## 2, RGI4; \ + G(RGI1, RGI2, x2, s0, s1, s2, s3); \ + G(RGI3, RGI4, y2, s1, s2, s3, s0); + +#define decround_tail(a, b, c, d, x, y) \ + vpslld $1, c, RT; \ + vpsrld $(32 - 1), c, c; \ + vpor c, RT, c; \ vpaddd x, y, x; \ vpaddd y, x, y; \ vpaddd y, RK2, y; \ @@ -130,32 +167,44 @@ vpsrld $1, d, y; \ vpslld $(32 - 1), d, d; \ vpor d, y, d; \ - vpslld $1, c, y; \ - vpsrld $(32 - 1), c, c; \ - vpor c, y, c; \ vpaddd x, RK1, x; \ vpxor x, c, c; -#define encrypt_round(n, a, b, c, d) \ - vbroadcastss (k+4*(2*(n)))(CTX), RK1; \ - vbroadcastss (k+4*(2*(n)+1))(CTX), RK2; \ - encround(a ## 1, b ## 1, c ## 1, d ## 1, RX, RY); \ - encround(a ## 2, b ## 2, c ## 2, d ## 2, RX, RY); - -#define decrypt_round(n, a, b, c, d) \ - vbroadcastss (k+4*(2*(n)))(CTX), RK1; \ - vbroadcastss (k+4*(2*(n)+1))(CTX), RK2; \ - decround(a ## 1, b ## 1, c ## 1, d ## 1, RX, RY); \ - decround(a ## 2, b ## 2, c ## 2, d ## 2, RX, RY); +#define preload_rgi(c) \ + vmovq c, RGI1; \ + vpextrq $1, c, RGI2; + +#define encrypt_round(n, a, b, c, d, preload) \ + vbroadcastss (k+4*(2*(n)))(CTX), RK1; \ + vbroadcastss (k+4*(2*(n)+1))(CTX), RK2; \ + encround_head_2(a, b, c, d, RX0, RY0, RX1, RY1); \ + encround_tail(a ## 1, b ## 1, c ## 1, d ## 1, RX0, RY0); \ + preload(c ## 1); \ + encround_tail(a ## 2, b ## 2, c ## 2, d ## 2, RX1, RY1); + +#define decrypt_round(n, a, b, c, d, preload) \ + vbroadcastss (k+4*(2*(n)))(CTX), RK1; \ + vbroadcastss (k+4*(2*(n)+1))(CTX), RK2; \ + decround_head_2(a, b, c, d, RX0, RY0, RX1, RY1); \ + decround_tail(a ## 1, b ## 1, c ## 1, d ## 1, RX0, RY0); \ + preload(c ## 1); \ + decround_tail(a ## 2, b ## 2, c ## 2, d ## 2, RX1, RY1); #define encrypt_cycle(n) \ - encrypt_round((2*n), RA, RB, RC, RD); \ - encrypt_round(((2*n) + 1), RC, RD, RA, RB); + encrypt_round((2*n), RA, RB, RC, RD, preload_rgi); \ + encrypt_round(((2*n) + 1), RC, RD, RA, RB, preload_rgi); + +#define encrypt_cycle_last(n) \ + encrypt_round((2*n), RA, RB, RC, RD, preload_rgi); \ + encrypt_round(((2*n) + 1), RC, RD, RA, RB, dummy); #define decrypt_cycle(n) \ - decrypt_round(((2*n) + 1), RC, RD, RA, RB); \ - decrypt_round((2*n), RA, RB, RC, RD); + decrypt_round(((2*n) + 1), RC, RD, RA, RB, preload_rgi); \ + decrypt_round((2*n), RA, RB, RC, RD, preload_rgi); +#define decrypt_cycle_last(n) \ + decrypt_round(((2*n) + 1), RC, RD, RA, RB, preload_rgi); \ + decrypt_round((2*n), RA, RB, RC, RD, dummy); #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \ vpunpckldq x1, x0, t0; \ @@ -216,17 +265,19 @@ __twofish_enc_blk_8way: * %rcx: bool, if true: xor output */ + pushq %rbp; pushq %rbx; pushq %rcx; vmovdqu w(CTX), RK1; leaq (4*4*4)(%rdx), %rax; - inpack_blocks(%rdx, RA1, RB1, RC1, RD1, RK1, RX, RY, RK2); - inpack_blocks(%rax, RA2, RB2, RC2, RD2, RK1, RX, RY, RK2); + inpack_blocks(%rdx, RA1, RB1, RC1, RD1, RK1, RX0, RY0, RK2); + vmovq RA1, RGI1; + vpextrq $1, RA1, RGI2; + inpack_blocks(%rax, RA2, RB2, RC2, RD2, RK1, RX0, RY0, RK2); - xorq RID1, RID1; - xorq RID2, RID2; + movq %rsi, %r11; encrypt_cycle(0); encrypt_cycle(1); @@ -235,26 +286,27 @@ __twofish_enc_blk_8way: encrypt_cycle(4); encrypt_cycle(5); encrypt_cycle(6); - encrypt_cycle(7); + encrypt_cycle_last(7); vmovdqu (w+4*4)(CTX), RK1; popq %rcx; popq %rbx; + popq %rbp; - leaq (4*4*4)(%rsi), %rax; + leaq (4*4*4)(%r11), %rax; testb %cl, %cl; jnz __enc_xor8; - outunpack_blocks(%rsi, RC1, RD1, RA1, RB1, RK1, RX, RY, RK2); - outunpack_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX, RY, RK2); + outunpack_blocks(%r11, RC1, RD1, RA1, RB1, RK1, RX0, RY0, RK2); + outunpack_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX0, RY0, RK2); ret; __enc_xor8: - outunpack_xor_blocks(%rsi, RC1, RD1, RA1, RB1, RK1, RX, RY, RK2); - outunpack_xor_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX, RY, RK2); + outunpack_xor_blocks(%r11, RC1, RD1, RA1, RB1, RK1, RX0, RY0, RK2); + outunpack_xor_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX0, RY0, RK2); ret; @@ -269,16 +321,18 @@ twofish_dec_blk_8way: * %rdx: src */ + pushq %rbp; pushq %rbx; vmovdqu (w+4*4)(CTX), RK1; leaq (4*4*4)(%rdx), %rax; - inpack_blocks(%rdx, RC1, RD1, RA1, RB1, RK1, RX, RY, RK2); - inpack_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX, RY, RK2); + inpack_blocks(%rdx, RC1, RD1, RA1, RB1, RK1, RX0, RY0, RK2); + vmovq RC1, RGI1; + vpextrq $1, RC1, RGI2; + inpack_blocks(%rax, RC2, RD2, RA2, RB2, RK1, RX0, RY0, RK2); - xorq RID1, RID1; - xorq RID2, RID2; + movq %rsi, %r11; decrypt_cycle(7); decrypt_cycle(6); @@ -287,14 +341,15 @@ twofish_dec_blk_8way: decrypt_cycle(3); decrypt_cycle(2); decrypt_cycle(1); - decrypt_cycle(0); + decrypt_cycle_last(0); vmovdqu (w)(CTX), RK1; popq %rbx; + popq %rbp; - leaq (4*4*4)(%rsi), %rax; - outunpack_blocks(%rsi, RA1, RB1, RC1, RD1, RK1, RX, RY, RK2); - outunpack_blocks(%rax, RA2, RB2, RC2, RD2, RK1, RX, RY, RK2); + leaq (4*4*4)(%r11), %rax; + outunpack_blocks(%r11, RA1, RB1, RC1, RD1, RK1, RX0, RY0, RK2); + outunpack_blocks(%rax, RA2, RB2, RC2, RD2, RK1, RX0, RY0, RK2); ret; -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
