size_t nr = MPW_BITS - nb;
mpw w;
- av += nw;
- w = av < avl ? *av++ : 0;
- while (av < avl) {
- mpw t;
- if (dv >= dvl) goto done;
- t = *av++;
- *dv++ = MPW((w >> nb) | (t << nr));
- w = t;
+ if (nw >= avl - av)
+ w = 0;
+ else {
+ av += nw;
+ w = *av++;
+
+ while (av < avl) {
+ mpw t;
+ if (dv >= dvl) goto done;
+ t = *av++;
+ *dv++ = MPW((w >> nb) | (t << nr));
+ w = t;
+ }
}
+
if (dv < dvl) {
*dv++ = MPW(w >> nb);
MPX_ZERO(dv, dvl);
void mpx_usubn(mpw *dv, mpw *dvl, mpw n) { MPX_USUBN(dv, dvl, n); }
-/* --- @mpx_uaddnlsl@ --- *
+/* --- @mpx_usubnlsl@ --- *
*
* Arguments: @mpw *dv, *dvl@ = destination and first argument vector
* @mpw a@ = second argument
#if CPUFAM_X86
MAYBE_UMUL4(x86_sse2)
+ MAYBE_UMUL4(x86_avx)
#endif
#if CPUFAM_AMD64
MAYBE_UMUL4(amd64_sse2)
+ MAYBE_UMUL4(amd64_avx)
#endif
static mpx_umul__functype *pick_umul(void)
{
#if CPUFAM_X86
+ DISPATCH_PICK_COND(mpx_umul, maybe_umul4_x86_avx,
+ cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpx_umul, maybe_umul4_x86_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
#endif
#if CPUFAM_AMD64
+ DISPATCH_PICK_COND(mpx_umul, maybe_umul4_amd64_avx,
+ cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpx_umul, maybe_umul4_amd64_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
#endif