From f52721b2b4e4949d26fa51572e3ad9000e1ed3be Mon Sep 17 00:00:00 2001
From: Jun He <jun.he@arm.com>
Date: Fri, 6 Aug 2021 14:46:53 +0800
Subject: [PATCH] decompression: optimize ExtractOffset for Arm

Inspired by kExtractMasksCombined, this patch uses shift
to replace table lookup. On Arm the codegen is 2 shift ops
(lsl+lsr). Comparing to previous ldr which requires 4 cycles
latency, the lsl+lsr only need 2 cycles.
Slight (~0.3%) uplift observed on N1, and ~3% on A72.

Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: I5b53632d22d9e5cf1a49d0c5cdd16265a15de23b
---
 snappy.cc | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/snappy.cc b/snappy.cc
index 3f446c6..4931104 100644
--- a/snappy.cc
+++ b/snappy.cc
@@ -1081,6 +1081,9 @@ inline uint32_t ExtractOffset(uint32_t val, size_t tag_type) {
          reinterpret_cast<const char*>(&kExtractMasksCombined) + 2 * tag_type,
          sizeof(result));
   return val & result;
+#elif defined(__aarch64__)
+  constexpr uint64_t kExtractMasksCombined = 0x0000FFFF00FF0000ull;
+  return val & (uint32_t)((kExtractMasksCombined >> (tag_type * 16)) & 0xFFFF);
 #else
   static constexpr uint32_t kExtractMasks[4] = {0, 0xFF, 0xFFFF, 0};
   return val & kExtractMasks[tag_type];