[aarch64]: redefine aes patterns
This first patch removes aarch64 usage of the aese/aesmc and aesd/aesimc
fusions (i.e. aes fusion) implemented in the scheduler due to unpredictable
behaviour observed in cases such as:
- when register allocation goes bad (e.g. extra movs)
- aes operations with xor and zeroed keys among interleaved operations
A more stable version should be provided by instead doing the aes fusion
during the combine pass. Since the aese and aesd patterns have been
rewritten as encapsulating a xor operation, the existing combine fusion
patterns have also been updated. The purpose is to simplify the need of
having additional combine patterns for cases like the ones below:
For AESE (though it also applies to AESD as both have a xor operation):
data = data ^ key;
data = vaeseq_u8(data, zero);
---
eor v1.16b, v0.16b, v1.16b
aese v1.16b, v2.16b
Should mean and generate the same as:
data = vaeseq_u8(data, key);
---
aese v1.16b, v0.16b
2019-07-09 Sylvia Taylor <sylvia.taylor@arm.com>
* config/aarch64/aarch64-simd.md
(aarch64_crypto_aes<aes_op>v16qi): Redefine pattern with xor.
(aarch64_crypto_aes<aesmc_op>v16qi): Remove attribute enabled.
(*aarch64_crypto_aes<aes_op>v16qi_xor_combine): Remove both.
(*aarch64_crypto_aese_fused,
*aarch64_crypto_aesd_fused): Update to new definition.
* config/aarch64/aarch64.c
(aarch_macro_fusion_pair_p): Remove aese/aesmc fusion check.
* gcc.target/aarch64/crypto-fuse-1.c: Remove.
* gcc.target/aarch64/crypto-fuse-2.c: Remove.
* gcc.target/aarch64/aes-fuse-1.c: New testcase.
* gcc.target/aarch64/aes-fuse-2.c: New testcase.
From-SVN: r273304