Extract the AES-NI encrypt -> decrypt assembly conversion
aes_hw_set_decrypt_key calls aes_hw_set_encrypt_key and then does a
conversion, all in assembly. On x86(_64), aes_hw_set_encrypt_key
internally checks OPENSSL_ia32cap_P to call one of two variants.
In preparation for splitting those variants into separate functions, get
the in-asm function call out o f the day by extracting an
aes_hw_encrypt_key_to_decrypt_key function.
Bug: 673
Change-Id: I23eefc00bdc8cb1f20e17fb6716974e91f1c32c4
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/68689
Reviewed-by: Bob Beck <bbe@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
diff --git a/gen/bcm/aesni-x86_64-apple.S b/gen/bcm/aesni-x86_64-apple.S
index 48d3cfc..ccf9f8f 100644
--- a/gen/bcm/aesni-x86_64-apple.S
+++ b/gen/bcm/aesni-x86_64-apple.S
@@ -1905,61 +1905,51 @@
ret
-.globl _aes_hw_set_decrypt_key
-.private_extern _aes_hw_set_decrypt_key
+.globl _aes_hw_encrypt_key_to_decrypt_key
+.private_extern _aes_hw_encrypt_key_to_decrypt_key
.p2align 4
-_aes_hw_set_decrypt_key:
-
+_aes_hw_encrypt_key_to_decrypt_key:
_CET_ENDBR
- subq $8,%rsp
-
-
- call __aesni_set_encrypt_key
+ movl 240(%rdi),%esi
shll $4,%esi
- testl %eax,%eax
- jnz L$dec_key_ret
- leaq 16(%rdx,%rsi,1),%rdi
- movups (%rdx),%xmm0
- movups (%rdi),%xmm1
- movups %xmm0,(%rdi)
- movups %xmm1,(%rdx)
- leaq 16(%rdx),%rdx
- leaq -16(%rdi),%rdi
+ leaq 16(%rdi,%rsi,1),%rdx
+
+ movups (%rdi),%xmm0
+ movups (%rdx),%xmm1
+ movups %xmm0,(%rdx)
+ movups %xmm1,(%rdi)
+ leaq 16(%rdi),%rdi
+ leaq -16(%rdx),%rdx
L$dec_key_inverse:
- movups (%rdx),%xmm0
- movups (%rdi),%xmm1
+ movups (%rdi),%xmm0
+ movups (%rdx),%xmm1
.byte 102,15,56,219,192
.byte 102,15,56,219,201
- leaq 16(%rdx),%rdx
- leaq -16(%rdi),%rdi
- movups %xmm0,16(%rdi)
- movups %xmm1,-16(%rdx)
- cmpq %rdx,%rdi
+ leaq 16(%rdi),%rdi
+ leaq -16(%rdx),%rdx
+ movups %xmm0,16(%rdx)
+ movups %xmm1,-16(%rdi)
+ cmpq %rdi,%rdx
ja L$dec_key_inverse
- movups (%rdx),%xmm0
+ movups (%rdi),%xmm0
.byte 102,15,56,219,192
pxor %xmm1,%xmm1
- movups %xmm0,(%rdi)
+ movups %xmm0,(%rdx)
pxor %xmm0,%xmm0
-L$dec_key_ret:
- addq $8,%rsp
-
ret
-
.globl _aes_hw_set_encrypt_key
.private_extern _aes_hw_set_encrypt_key
.p2align 4
_aes_hw_set_encrypt_key:
-__aesni_set_encrypt_key:
_CET_ENDBR
@@ -2331,7 +2321,6 @@
xorps %xmm1,%xmm2
ret
-
.section __DATA,__const
.p2align 6
L$bswap_mask:
