74 		u64 T = U64(0xe100000000000000) & (0-(V.lo&1)); \
75 		V.lo  = (V.hi<<63)|(V.lo>>1); \
79 		u32 T = 0xe1000000U & (0-(u32)(V.lo&1)); \
80 		V.lo  = (V.hi<<63)|(V.lo>>1); \
127 	Htable[0].lo = 0;  in gcm_init_8bit()
129 	V.lo = H[1];  in gcm_init_8bit()
140 			Hi[j].lo = H0.lo^Htable[j].lo;  in gcm_init_8bit()
219 		Z.lo ^= Htable[n].lo;  in gcm_gmult_8bit()
225 		rem  = (size_t)Z.lo&0xff;  in gcm_gmult_8bit()
226 		Z.lo = (Z.hi<<56)|(Z.lo>>8);  in gcm_gmult_8bit()
237 		Xi[1] = BSWAP8(Z.lo);  in gcm_gmult_8bit()
243 		v = (u32)(Z.lo>>32);	PUTU32(p+8,v);  in gcm_gmult_8bit()
244 		v = (u32)(Z.lo);	PUTU32(p+12,v);  in gcm_gmult_8bit()
249 		Xi[1] = Z.lo;  in gcm_gmult_8bit()
264 	Htable[0].lo = 0;  in gcm_init_4bit()
266 	V.lo = H[1];  in gcm_init_4bit()
279 			Hi[j].lo = V.lo^Htable[j].lo;  in gcm_init_4bit()
290 	Htable[3].hi  = V.hi^Htable[2].hi, Htable[3].lo  = V.lo^Htable[2].lo;  in gcm_init_4bit()
292 	Htable[5].hi  = V.hi^Htable[1].hi, Htable[5].lo  = V.lo^Htable[1].lo;  in gcm_init_4bit()
293 	Htable[6].hi  = V.hi^Htable[2].hi, Htable[6].lo  = V.lo^Htable[2].lo;  in gcm_init_4bit()
294 	Htable[7].hi  = V.hi^Htable[3].hi, Htable[7].lo  = V.lo^Htable[3].lo;  in gcm_init_4bit()
296 	Htable[9].hi  = V.hi^Htable[1].hi, Htable[9].lo  = V.lo^Htable[1].lo;  in gcm_init_4bit()
297 	Htable[10].hi = V.hi^Htable[2].hi, Htable[10].lo = V.lo^Htable[2].lo;  in gcm_init_4bit()
298 	Htable[11].hi = V.hi^Htable[3].hi, Htable[11].lo = V.lo^Htable[3].lo;  in gcm_init_4bit()
299 	Htable[12].hi = V.hi^Htable[4].hi, Htable[12].lo = V.lo^Htable[4].lo;  in gcm_init_4bit()
300 	Htable[13].hi = V.hi^Htable[5].hi, Htable[13].lo = V.lo^Htable[5].lo;  in gcm_init_4bit()
301 	Htable[14].hi = V.hi^Htable[6].hi, Htable[14].lo = V.lo^Htable[6].lo;  in gcm_init_4bit()
302 	Htable[15].hi = V.hi^Htable[7].hi, Htable[15].lo = V.lo^Htable[7].lo;  in gcm_init_4bit()
315 			Htable[j].hi = V.lo;  in gcm_init_4bit()
316 			Htable[j].lo = V.hi;  in gcm_init_4bit()
321 			Htable[j].hi = V.lo<<32|V.lo>>32;  in gcm_init_4bit()
322 			Htable[j].lo = V.hi<<32|V.hi>>32;  in gcm_init_4bit()
347 	Z.lo = Htable[nlo].lo;  in gcm_gmult_4bit()
350 		rem  = (size_t)Z.lo&0xf;  in gcm_gmult_4bit()
351 		Z.lo = (Z.hi<<60)|(Z.lo>>4);  in gcm_gmult_4bit()
359 		Z.lo ^= Htable[nhi].lo;  in gcm_gmult_4bit()
367 		rem  = (size_t)Z.lo&0xf;  in gcm_gmult_4bit()
368 		Z.lo = (Z.hi<<60)|(Z.lo>>4);  in gcm_gmult_4bit()
376 		Z.lo ^= Htable[nlo].lo;  in gcm_gmult_4bit()
382 		Xi[1] = BSWAP8(Z.lo);  in gcm_gmult_4bit()
388 		v = (u32)(Z.lo>>32);	PUTU32(p+8,v);  in gcm_gmult_4bit()
389 		v = (u32)(Z.lo);	PUTU32(p+12,v);  in gcm_gmult_4bit()
394 		Xi[1] = Z.lo;  in gcm_gmult_4bit()
423 	Z.lo = Htable[nlo].lo;
426 		rem  = (size_t)Z.lo&0xf;
427 		Z.lo = (Z.hi<<60)|(Z.lo>>4);
435 		Z.lo ^= Htable[nhi].lo;
444 		rem  = (size_t)Z.lo&0xf;
445 		Z.lo = (Z.hi<<60)|(Z.lo>>4);
453 		Z.lo ^= Htable[nlo].lo;
505 	Z.lo = Htable[cnt].lo;
506 	Hshr4[cnt].lo = (Z.hi<<60)|(Z.lo>>4);
508 	Hshl4[cnt]    = (u8)(Z.lo<<4);
512 	for (Z.lo=0, Z.hi=0, cnt=15; cnt; --cnt) {
519 		Z.lo ^= Htable[nlo].lo;
521 		rem = (size_t)Z.lo&0xff;
523 		Z.lo = (Z.hi<<56)|(Z.lo>>8);
527 		Z.lo ^= Hshr4[nhi].lo;
537 	Z.lo ^= Htable[nlo].lo;
539 	rem = (size_t)Z.lo&0xf;
541 	Z.lo = (Z.hi<<60)|(Z.lo>>4);
545 	Z.lo ^= Htable[nhi].lo;
552 		Xi[1] = BSWAP8(Z.lo);
558 		v = (u32)(Z.lo>>32);	PUTU32(p+8,v);
559 		v = (u32)(Z.lo);	PUTU32(p+12,v);
564 		Xi[1] = Z.lo;
594 	V.lo = H[1];
617 			Z.lo ^= V.lo&M;
626 		Xi[1] = BSWAP8(Z.lo);
632 		v = (u32)(Z.lo>>32);	PUTU32(p+8,v);
633 		v = (u32)(Z.lo);	PUTU32(p+12,v);
638 		Xi[1] = Z.lo;
703 		u64 hi,lo;
705 		lo = (u64)GETU32(p+8)<<32|GETU32(p+12);
707 		ctx->H.u[1] = lo;