diff options
author | Jun He <jun.he@arm.com> | 2021-08-17 16:36:43 +0800 |
---|---|---|
committer | Jun He <jun.he@arm.com> | 2021-08-30 09:51:37 +0800 |
commit | aeb5de55a9c646ca334e4f6252af28536ca22349 (patch) | |
tree | 84902cec7dfeea4ab7be23b73c950462521cd2fd | |
parent | 7062d7f1d8a5ca7ade075579f36681ff48616dc3 (diff) | |
download | snappy-git-aeb5de55a9c646ca334e4f6252af28536ca22349.tar.gz |
decompress: refine data depdency
The final ip advance value doesn't have to wait for
the result of offset to load *tag. It can be computed
along with the offset, so the codegen will use one
csinc in parallel with ldrb. This will improve the
throughput.
With this change it is observed ~4.2% uplift in UFlat/10
and ~3.7% in UFlatMedley
Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: I20ab211235bbf578c6c978f2bbd9160a49e920da
-rw-r--r-- | snappy.cc | 12 |
1 files changed, 8 insertions, 4 deletions
@@ -1015,12 +1015,16 @@ size_t AdvanceToNextTagARMOptimized(const uint8_t** ip_p, size_t* tag) { // delta2 = ((c >> 2) + 1) ip++ // This is different from X86 optimizations because ARM has conditional add // instruction (csinc) and it removes several register moves. - const size_t literal_tag_offset = (*tag >> 2) + 1; const size_t tag_type = *tag & 3; const bool is_literal = (tag_type == 0); - *tag = is_literal ? ip[literal_tag_offset] : ip[tag_type]; - ip += is_literal ? literal_tag_offset : tag_type; - ip++; + if (is_literal) { + size_t next_literal_tag = (*tag >> 2) + 1; + *tag = ip[next_literal_tag]; + ip += next_literal_tag + 1; + } else { + *tag = ip[tag_type]; + ip += tag_type + 1; + } return tag_type; } |