diff options
Diffstat (limited to 'arch/x86/crypto/aes-i586-asm_32.S')
| -rw-r--r-- | arch/x86/crypto/aes-i586-asm_32.S | 89 |
1 files changed, 43 insertions, 46 deletions
diff --git a/arch/x86/crypto/aes-i586-asm_32.S b/arch/x86/crypto/aes-i586-asm_32.S index f942f0c..1093bed 100644 --- a/arch/x86/crypto/aes-i586-asm_32.S +++ b/arch/x86/crypto/aes-i586-asm_32.S @@ -46,9 +46,9 @@ #define in_blk 16 /* offsets in crypto_tfm structure */ -#define ekey (crypto_tfm_ctx_offset + 0) -#define nrnd (crypto_tfm_ctx_offset + 256) -#define dkey (crypto_tfm_ctx_offset + 260) +#define klen (crypto_tfm_ctx_offset + 0) +#define ekey (crypto_tfm_ctx_offset + 4) +#define dkey (crypto_tfm_ctx_offset + 244) // register mapping for encrypt and decrypt subroutines @@ -221,8 +221,8 @@ .global aes_enc_blk -.extern ft_tab -.extern fl_tab +.extern crypto_ft_tab +.extern crypto_fl_tab .align 4 @@ -236,7 +236,7 @@ aes_enc_blk: 1: push %ebx mov in_blk+4(%esp),%r2 push %esi - mov nrnd(%ebp),%r3 // number of rounds + mov klen(%ebp),%r3 // key size push %edi #if ekey != 0 lea ekey(%ebp),%ebp // key pointer @@ -255,26 +255,26 @@ aes_enc_blk: sub $8,%esp // space for register saves on stack add $16,%ebp // increment to next round key - cmp $12,%r3 + cmp $24,%r3 jb 4f // 10 rounds for 128-bit key lea 32(%ebp),%ebp je 3f // 12 rounds for 192-bit key lea 32(%ebp),%ebp -2: fwd_rnd1( -64(%ebp) ,ft_tab) // 14 rounds for 256-bit key - fwd_rnd2( -48(%ebp) ,ft_tab) -3: fwd_rnd1( -32(%ebp) ,ft_tab) // 12 rounds for 192-bit key - fwd_rnd2( -16(%ebp) ,ft_tab) -4: fwd_rnd1( (%ebp) ,ft_tab) // 10 rounds for 128-bit key - fwd_rnd2( +16(%ebp) ,ft_tab) - fwd_rnd1( +32(%ebp) ,ft_tab) - fwd_rnd2( +48(%ebp) ,ft_tab) - fwd_rnd1( +64(%ebp) ,ft_tab) - fwd_rnd2( +80(%ebp) ,ft_tab) - fwd_rnd1( +96(%ebp) ,ft_tab) - fwd_rnd2(+112(%ebp) ,ft_tab) - fwd_rnd1(+128(%ebp) ,ft_tab) - fwd_rnd2(+144(%ebp) ,fl_tab) // last round uses a different table +2: fwd_rnd1( -64(%ebp), crypto_ft_tab) // 14 rounds for 256-bit key + fwd_rnd2( -48(%ebp), crypto_ft_tab) +3: fwd_rnd1( -32(%ebp), crypto_ft_tab) // 12 rounds for 192-bit key + fwd_rnd2( -16(%ebp), crypto_ft_tab) +4: fwd_rnd1( (%ebp), crypto_ft_tab) // 10 rounds for 128-bit key + fwd_rnd2( +16(%ebp), crypto_ft_tab) + fwd_rnd1( +32(%ebp), crypto_ft_tab) + fwd_rnd2( +48(%ebp), crypto_ft_tab) + fwd_rnd1( +64(%ebp), crypto_ft_tab) + fwd_rnd2( +80(%ebp), crypto_ft_tab) + fwd_rnd1( +96(%ebp), crypto_ft_tab) + fwd_rnd2(+112(%ebp), crypto_ft_tab) + fwd_rnd1(+128(%ebp), crypto_ft_tab) + fwd_rnd2(+144(%ebp), crypto_fl_tab) // last round uses a different table // move final values to the output array. CAUTION: the // order of these assigns rely on the register mappings @@ -297,8 +297,8 @@ aes_enc_blk: .global aes_dec_blk -.extern it_tab -.extern il_tab +.extern crypto_it_tab +.extern crypto_il_tab .align 4 @@ -312,14 +312,11 @@ aes_dec_blk: 1: push %ebx mov in_blk+4(%esp),%r2 push %esi - mov nrnd(%ebp),%r3 // number of rounds + mov klen(%ebp),%r3 // key size push %edi #if dkey != 0 lea dkey(%ebp),%ebp // key pointer #endif - mov %r3,%r0 - shl $4,%r0 - add %r0,%ebp // input four columns and xor in first round key @@ -333,27 +330,27 @@ aes_dec_blk: xor 12(%ebp),%r5 sub $8,%esp // space for register saves on stack - sub $16,%ebp // increment to next round key - cmp $12,%r3 + add $16,%ebp // increment to next round key + cmp $24,%r3 jb 4f // 10 rounds for 128-bit key - lea -32(%ebp),%ebp + lea 32(%ebp),%ebp je 3f // 12 rounds for 192-bit key - lea -32(%ebp),%ebp - -2: inv_rnd1( +64(%ebp), it_tab) // 14 rounds for 256-bit key - inv_rnd2( +48(%ebp), it_tab) -3: inv_rnd1( +32(%ebp), it_tab) // 12 rounds for 192-bit key - inv_rnd2( +16(%ebp), it_tab) -4: inv_rnd1( (%ebp), it_tab) // 10 rounds for 128-bit key - inv_rnd2( -16(%ebp), it_tab) - inv_rnd1( -32(%ebp), it_tab) - inv_rnd2( -48(%ebp), it_tab) - inv_rnd1( -64(%ebp), it_tab) - inv_rnd2( -80(%ebp), it_tab) - inv_rnd1( -96(%ebp), it_tab) - inv_rnd2(-112(%ebp), it_tab) - inv_rnd1(-128(%ebp), it_tab) - inv_rnd2(-144(%ebp), il_tab) // last round uses a different table + lea 32(%ebp),%ebp + +2: inv_rnd1( -64(%ebp), crypto_it_tab) // 14 rounds for 256-bit key + inv_rnd2( -48(%ebp), crypto_it_tab) +3: inv_rnd1( -32(%ebp), crypto_it_tab) // 12 rounds for 192-bit key + inv_rnd2( -16(%ebp), crypto_it_tab) +4: inv_rnd1( (%ebp), crypto_it_tab) // 10 rounds for 128-bit key + inv_rnd2( +16(%ebp), crypto_it_tab) + inv_rnd1( +32(%ebp), crypto_it_tab) + inv_rnd2( +48(%ebp), crypto_it_tab) + inv_rnd1( +64(%ebp), crypto_it_tab) + inv_rnd2( +80(%ebp), crypto_it_tab) + inv_rnd1( +96(%ebp), crypto_it_tab) + inv_rnd2(+112(%ebp), crypto_it_tab) + inv_rnd1(+128(%ebp), crypto_it_tab) + inv_rnd2(+144(%ebp), crypto_il_tab) // last round uses a different table // move final values to the output array. CAUTION: the // order of these assigns rely on the register mappings |