diff options
| -rw-r--r-- | modules/alg_amd64/asm_macr.h | 3 | ||||
| -rw-r--r-- | modules/alg_amd64/mp_muladd.S | 24 |
2 files changed, 18 insertions, 9 deletions
diff --git a/modules/alg_amd64/asm_macr.h b/modules/alg_amd64/asm_macr.h index 689bfa3cb..93b675858 100644 --- a/modules/alg_amd64/asm_macr.h +++ b/modules/alg_amd64/asm_macr.h @@ -67,6 +67,7 @@ func_name: #define IMM(VAL) $VAL #define ADD(TO, FROM) addq FROM, TO +#define ADD_LAST_CARRY(REG) adcq IMM(0), REG #define ADD_IMM(TO, NUM) ADD(TO, IMM(NUM)) #define ADD_W_CARRY(TO1, TO2, FROM) addq FROM, TO1; adcq IMM(0), TO2; #define SUB_IMM(TO, NUM) sub IMM(NUM), TO @@ -78,4 +79,6 @@ func_name: #define NOT(REG) notq REG #define ZEROIZE(REG) XOR(REG, REG) +#define RETURN_VALUE_IS(V) ASSIGN(%rax, V) + #endif diff --git a/modules/alg_amd64/mp_muladd.S b/modules/alg_amd64/mp_muladd.S index e34b855bb..71477cf2b 100644 --- a/modules/alg_amd64/mp_muladd.S +++ b/modules/alg_amd64/mp_muladd.S @@ -8,11 +8,15 @@ START_LISTING(mp_muladd.S) START_FUNCTION(bigint_mul_add_words) + #define LOOP_CTR %r11d #define X_ARR %rsi #define Z_ARR %rdi #define Y %rcx #define CARRY %r8 +#define Z_WORD %r9 +#define MUL_LO %rax +#define MUL_HI %rdx ZEROIZE(CARRY) @@ -21,14 +25,16 @@ START_FUNCTION(bigint_mul_add_words) JUMP_IF_ZERO(LOOP_CTR, .DONE) JUMP_IF_LT(LOOP_CTR, 8, .MULADD1_LOOP) -#define MULADD_OP(N) \ - ASSIGN(%rax, ARRAY8(X_ARR, N)) ; \ - ASSIGN(%r9, ARRAY8(Z_ARR, N)) ; \ - MUL(Y) ; \ - ADD_W_CARRY(%rax, %rdx, CARRY) ; \ - ADD_W_CARRY(%rax, %rdx, %r9) ; \ - ASSIGN(CARRY, %rdx) ; \ - ASSIGN(ARRAY8(Z_ARR, N), %rax) +#define MULADD_OP(N) \ + ASSIGN(MUL_LO, ARRAY8(X_ARR, N)) ; \ + ASSIGN(Z_WORD, ARRAY8(Z_ARR, N)) ; \ + MUL(Y) ; \ + ADD(Z_WORD, CARRY) ; \ + ASSIGN(CARRY, MUL_HI) ; \ + ADD_LAST_CARRY(CARRY) ; \ + ADD(Z_WORD, MUL_LO) ; \ + ADD_LAST_CARRY(CARRY) ; \ + ASSIGN(ARRAY8(Z_ARR, N), Z_WORD) START_LOOP(.MULADD8) MULADD_OP(0) @@ -56,5 +62,5 @@ START_LOOP(.MULADD1) LOOP_UNTIL_EQ(LOOP_CTR, 0, .MULADD1) .DONE: - ASSIGN(%rax, CARRY) + RETURN_VALUE_IS(CARRY) END_FUNCTION(bigint_mul_add_words) |