diff options
| author | Andy Polyakov <appro@openssl.org> | 2012-11-28 13:19:10 +0000 |
|---|---|---|
| committer | Andy Polyakov <appro@openssl.org> | 2012-11-28 13:19:10 +0000 |
| commit | 904732f68bcc6ebd3f8961a9272bc811dc26bcbd (patch) | |
| tree | 7ab988b23467d2545e4aa2e50bd176bf6923fc2a /crypto/bn/asm/bn-c64xplus.asm | |
| parent | cf5ecc3e1fd112dd8a544bfb26bfb96c96b604c7 (diff) | |
C64x+ assembly pack: improve EABI support.
Diffstat (limited to 'crypto/bn/asm/bn-c64xplus.asm')
| -rw-r--r-- | crypto/bn/asm/bn-c64xplus.asm | 51 |
1 files changed, 44 insertions, 7 deletions
diff --git a/crypto/bn/asm/bn-c64xplus.asm b/crypto/bn/asm/bn-c64xplus.asm index 161547c3b0..f07b09e439 100644 --- a/crypto/bn/asm/bn-c64xplus.asm +++ b/crypto/bn/asm/bn-c64xplus.asm @@ -12,6 +12,18 @@ ;; SPLOOPs spin at ... 2*n cycles [plus epilogue]. ;;==================================================================== .text + .if __TI_EABI__ + .asg bn_mul_add_words,_bn_mul_add_words + .asg bn_mul_words,_bn_mul_words + .asg bn_sqr_words,_bn_sqr_words + .asg bn_add_words,_bn_add_words + .asg bn_sub_words,_bn_sub_words + .asg bn_div_words,_bn_div_words + .asg bn_sqr_comba8,_bn_sqr_comba8 + .asg bn_mul_comba8,_bn_mul_comba8 + .asg bn_sqr_comba4,_bn_sqr_comba4 + .asg bn_mul_comba4,_bn_mul_comba4 + .endif .asg B3,RA .asg A4,ARG0 @@ -158,14 +170,39 @@ _bn_sub_words: .endasmfunc .global _bn_div_words - .global __divull _bn_div_words: .asmfunc - CALLP __divull,A3 ; jump to rts64plus.lib -|| MV ARG0,A5 -|| MV ARG1,ARG0 -|| MV ARG2,ARG1 -|| ZERO B5 + LMBD 1,A6,A0 ; leading zero bits in dv + LMBD 1,A4,A1 ; leading zero bits in hi +|| MVK 32,B0 + CMPLTU A1,A0,A2 +|| ADD A0,B0,B0 + [ A2] BNOP RA +||[ A2] MVK -1,A4 ; return overflow +||[!A2] MV A4,A3 ; reassign hi + [!A2] MV B4,A4 ; reassign lo, will be quotient +||[!A2] MVC B0,ILC + [!A2] SHL A6,A0,A6 ; normalize dv +|| MVK 1,A1 + + [!A2] CMPLTU A3,A6,A1 ; hi<dv? +||[!A2] SHL A4,1,A5:A4 ; lo<<1 + [!A1] SUB A3,A6,A3 ; hi-=dv +||[!A1] OR 1,A4,A4 + [!A2] SHRU A3,31,A1 ; upper bit +||[!A2] ADDAH A5,A3,A3 ; hi<<1|lo>>31 + + SPLOOP 3 + [!A1] CMPLTU A3,A6,A1 ; hi<dv? +||[ A1] ZERO A1 +|| SHL A4,1,A5:A4 ; lo<<1 + [!A1] SUB A3,A6,A3 ; hi-=dv +||[!A1] OR 1,A4,A4 ; quotient + SHRU A3,31,A1 ; upper bit +|| ADDAH A5,A3,A3 ; hi<<1|lo>>31 + SPKERNEL + + BNOP RA,5 .endasmfunc ;;==================================================================== @@ -256,7 +293,7 @@ _bn_mul_comba4: || LDW *A5++,B6 ; ap[0] || MV A0,A3 ; const A3=M .else - ;; This alternative is exercise in fully unrolled Comba + ;; This alternative is an exercise in fully unrolled Comba ;; algorithm implementation that operates at n*(n+1)+12, or ;; as little as 32 cycles... LDW *ARG1[0],B16 ; a[0] |