@@ -205,6 +205,21 @@ ALL_F: .octa 0xffffffffffffffffffffffffffffffff
205
205
# GCM_INIT initializes a gcm_context struct to prepare for encoding/decoding.
206
206
# Clobbers rax, r10-r13 and xmm0-xmm6, %xmm13
207
207
.macro GCM_INIT
208
+
209
+ mov arg9, %r11
210
+ mov %r11 , AadLen(%arg2) # ctx_data.aad_length = aad_length
211
+ xor %r11 , %r11
212
+ mov %r11 , InLen(%arg2) # ctx_data.in_length = 0
213
+ mov %r11 , PBlockLen(%arg2) # ctx_data.partial_block_length = 0
214
+ mov %r11 , PBlockEncKey(%arg2) # ctx_data.partial_block_enc_key = 0
215
+ mov %arg6, %rax
216
+ movdqu (%rax ), %xmm0
217
+ movdqu %xmm0 , OrigIV(%arg2) # ctx_data.orig_IV = iv
218
+
219
+ movdqa SHUF_MASK(%rip ), %xmm2
220
+ PSHUFB_XMM %xmm2 , %xmm0
221
+ movdqu %xmm0 , CurCount(%arg2) # ctx_data.current_counter = iv
222
+
208
223
mov arg7, %r12
209
224
movdqu (%r12 ), %xmm13
210
225
movdqa SHUF_MASK(%rip ), %xmm2
@@ -227,20 +242,22 @@ ALL_F: .octa 0xffffffffffffffffffffffffffffffff
227
242
pand POLY(%rip ), %xmm2
228
243
pxor %xmm2 , %xmm13
229
244
movdqa %xmm13 , HashKey(%rsp )
230
- mov %arg5, %r13 # %xmm13 holds HashKey<<1 (mod poly)
231
- and $-16 , %r13
232
- mov %r13 , %r12
233
245
234
246
CALC_AAD_HASH %xmm13 %xmm0 %xmm1 %xmm2 %xmm3 %xmm4 \
235
247
%xmm5 %xmm6
236
- mov %r13 , %r12
237
248
.endm
238
249
239
250
# GCM_ENC_DEC Encodes/Decodes given data. Assumes that the passed gcm_context
240
251
# struct has been initialized by GCM_INIT.
241
252
# Requires the input data be at least 1 byte long because of READ_PARTIAL_BLOCK
242
253
# Clobbers rax, r10-r13, and xmm0-xmm15
243
254
.macro GCM_ENC_DEC operation
255
+ movdqu AadHash(%arg2), %xmm8
256
+ movdqu HashKey(%rsp ), %xmm13
257
+ add %arg5, InLen(%arg2)
258
+ mov %arg5, %r13 # save the number of bytes
259
+ and $-16 , %r13 # %r13 = %r13 - (%r13 mod 16)
260
+ mov %r13 , %r12
244
261
# Encrypt/Decrypt first few blocks
245
262
246
263
and $(3 <<4 ), %r12
@@ -285,16 +302,23 @@ _four_cipher_left_\@:
285
302
GHASH_LAST_4 %xmm9 , %xmm10 , %xmm11 , %xmm12 , %xmm13 , %xmm14 , \
286
303
%xmm15 , %xmm1 , %xmm2 , %xmm3 , %xmm4 , %xmm8
287
304
_zero_cipher_left_\@:
305
+ movdqu %xmm8 , AadHash(%arg2)
306
+ movdqu %xmm0 , CurCount(%arg2)
307
+
288
308
mov %arg5, %r13
289
309
and $15 , %r13 # %r13 = arg5 (mod 16)
290
310
je _multiple_of_16_bytes_\@
291
311
312
+ mov %r13 , PBlockLen(%arg2)
313
+
292
314
# Handle the last <16 Byte block separately
293
315
paddd ONE(%rip ), %xmm0 # INCR CNT to get Yn
316
+ movdqu %xmm0 , CurCount(%arg2)
294
317
movdqa SHUF_MASK(%rip ), %xmm10
295
318
PSHUFB_XMM %xmm10 , %xmm0
296
319
297
320
ENCRYPT_SINGLE_BLOCK %xmm0 , %xmm1 # Encrypt(K, Yn)
321
+ movdqu %xmm0 , PBlockEncKey(%arg2)
298
322
299
323
lea (%arg4,%r11 ,1 ), %r10
300
324
mov %r13 , %r12
@@ -323,6 +347,7 @@ _zero_cipher_left_\@:
323
347
.endif
324
348
325
349
GHASH_MUL %xmm8 , %xmm13 , %xmm9 , %xmm10 , %xmm11 , %xmm5 , %xmm6
350
+ movdqu %xmm8 , AadHash(%arg2)
326
351
.ifc \operation, enc
327
352
# GHASH computation for the last <16 byte block
328
353
movdqa SHUF_MASK(%rip ), %xmm10
@@ -352,11 +377,15 @@ _multiple_of_16_bytes_\@:
352
377
# Output: Authorization Tag (AUTH_TAG)
353
378
# Clobbers rax, r10-r12, and xmm0, xmm1, xmm5-xmm15
354
379
.macro GCM_COMPLETE
355
- mov arg9, %r12 # %r13 = aadLen (number of bytes)
380
+ movdqu AadHash(%arg2), %xmm8
381
+ movdqu HashKey(%rsp ), %xmm13
382
+ mov AadLen(%arg2), %r12 # %r13 = aadLen (number of bytes)
356
383
shl $3 , %r12 # convert into number of bits
357
384
movd %r12d , %xmm15 # len(A) in %xmm15
358
- shl $3 , %arg5 # len(C) in bits (*128)
359
- MOVQ_R64_XMM %arg5, %xmm1
385
+ mov InLen(%arg2), %r12
386
+ shl $3 , %r12 # len(C) in bits (*128)
387
+ MOVQ_R64_XMM %r12 , %xmm1
388
+
360
389
pslldq $8 , %xmm15 # %xmm15 = len(A)||0x0000000000000000
361
390
pxor %xmm1 , %xmm15 # %xmm15 = len(A)||len(C)
362
391
pxor %xmm15 , %xmm8
@@ -365,8 +394,7 @@ _multiple_of_16_bytes_\@:
365
394
movdqa SHUF_MASK(%rip ), %xmm10
366
395
PSHUFB_XMM %xmm10 , %xmm8
367
396
368
- mov %arg6, %rax # %rax = *Y0
369
- movdqu (%rax ), %xmm0 # %xmm0 = Y0
397
+ movdqu OrigIV(%arg2), %xmm0 # %xmm0 = Y0
370
398
ENCRYPT_SINGLE_BLOCK %xmm0 , %xmm1 # E(K, Y0)
371
399
pxor %xmm8 , %xmm0
372
400
_return_T_\@:
@@ -554,15 +582,14 @@ _get_AAD_done\@:
554
582
555
583
.macro INITIAL_BLOCKS_ENC_DEC TMP1 TMP2 TMP3 TMP4 TMP5 XMM0 XMM1 \
556
584
XMM2 XMM3 XMM4 XMMDst TMP6 TMP7 i i_seq operation
585
+ MOVADQ SHUF_MASK(%rip ), %xmm14
557
586
558
587
movdqu AadHash(%arg2), %xmm\i # XMM0 = Y0
559
588
560
589
xor %r11 , %r11 # initialise the data pointer offset as zero
561
590
# start AES for num_initial_blocks blocks
562
591
563
- mov %arg6, %rax # %rax = *Y0
564
- movdqu (%rax ), \XMM0 # XMM0 = Y0
565
- PSHUFB_XMM %xmm14 , \XMM0
592
+ movdqu CurCount(%arg2), \XMM0 # XMM0 = Y0
566
593
567
594
.if (\i == 5 ) || (\i == 6 ) || (\i == 7 )
568
595
0 commit comments