From f026d93d15e62bed8c2ae419a91ad9ee40ee7401 Mon Sep 17 00:00:00 2001
From: Elliott Hughes <enh@google.com>
Date: Fri, 12 Jul 2019 16:32:50 -0700
Subject: Enable the various chrome zlib optimizations.

This is for arm/arm64 only, because there are test failures for
x86/x86-64 that we haven't understood yet.

Bug: https://issuetracker.google.com/115695768
Bug: https://issuetracker.google.com/135889247
Bug: http://b/135038583
Test: boots
Test: zlib_bench scores 2x or more faster on arm/arm64
Change-Id: I1f991cb71338bc99885edbebc964f2ad3b14234a
---
 Android.bp | 87 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 78 insertions(+), 9 deletions(-)

diff --git a/Android.bp b/Android.bp
index 2937ca3..3e060b0 100644
--- a/Android.bp
+++ b/Android.bp
@@ -1,9 +1,62 @@
+srcs_opt = [
+    "adler32_simd.c",
+    // See https://chromium-review.googlesource.com/749732.
+    "contrib/optimizations/inffast_chunk.c",
+    "contrib/optimizations/inflate.c",
+    "crc32_simd.c",
+]
+
+cflags_arm = [
+    // Since we're building for the platform, we claim to be Linux rather than
+    // Android so we use getauxval() directly instead of the NDK
+    // android_getCpuFeatures which isn't available to us anyway.
+    "-DARMV8_OS_LINUX",
+    // We no longer support non-Neon platform builds.
+    "-DADLER32_SIMD_NEON",
+    "-DINFLATE_CHUNK_SIMD_NEON",
+    // HWCAP_CRC32 is checked at runtime, so it's okay to turn crc32
+    // acceleration on for both 32- and 64-bit.
+    "-DCRC32_ARMV8_CRC32",
+    // Testing with zlib_bench shows -O3 is a win for ARM but a bit of a wash
+    // for x86, so match the BUILD file in only enabling this for ARM.
+    "-O3",
+]
+srcs_arm = [
+    "arm_features.c",
+] + srcs_opt
+
+cflags_x86 = []
+srcs_x86 = ["inflate.c"]
+
+/* TODO: this should work, but fails some tests on cuttlefish.
+cflags_x86 = [
+    // Android's x86/x86-64 ABI includes SSE2 and SSSE3.
+    "-DADLER32_SIMD_SSSE3",
+    "-DINFLATE_CHUNK_SIMD_SSE2",
+    // TODO: ...but the host build system defaults don't match our official ABI.
+    "-mssse3",
+    // PCLMUL isn't in the ABI, but it won't actually be used unless CPUID
+    // reports that the processor really does have the instruction.
+    "-mpclmul",
+    "-DCRC32_SIMD_SSE42_PCLMUL",
+]
+srcs_x86 = [
+    "crc_folding.c",
+    "fill_window_sse.c",
+    "x86.c",
+] + srcs_opt
+*/
+
+// This optimization is applicable to arm64 and x86-64.
+cflags_64 = ["-DINFLATE_CHUNK_READ_64LE"]
+
 cc_defaults {
     name: "libz_defaults",
 
     cflags: [
-        "-O3",
+        // We do support hidden visibility, so turn that on.
         "-DHAVE_HIDDEN",
+        // We do support const, so turn that on.
         "-DZLIB_CONST",
         "-Wall",
         "-Werror",
@@ -22,27 +75,35 @@ cc_defaults {
         "gzread.c",
         "gzwrite.c",
         "infback.c",
-        "inflate.c",
         "inftrees.c",
         "inffast.c",
+        "simd_stub.c",
         "trees.c",
         "uncompr.c",
         "zutil.c",
-
-        "simd_stub.c",
     ],
 
     arch: {
         arm: {
-            // measurements show that the ARM version of ZLib is about x1.17 faster
-            // than the thumb one...
-            // TODO: re-test with zlib_bench after SIMD is enabled.
-            instruction_set: "arm",
-
             // TODO: This is to work around b/24465209. Remove after root cause
             // is fixed.
             pack_relocations: false,
             ldflags: ["-Wl,--hash-style=both"],
+
+            cflags: cflags_arm,
+            srcs: srcs_arm,
+        },
+        arm64: {
+            cflags: cflags_arm + cflags_64,
+            srcs: srcs_arm,
+        },
+        x86: {
+            cflags: cflags_x86,
+            srcs: srcs_x86,
+        },
+        x86_64: {
+            cflags: cflags_x86 + cflags_64,
+            srcs: srcs_x86,
         },
     },
 }
@@ -73,6 +134,8 @@ cc_library {
     },
 }
 
+// libz_current allows modules building against the NDK to have access to zlib
+// API that's not available from the NDK libz.
 cc_library_static {
     name: "libz_current",
     defaults: ["libz_defaults"],
@@ -93,6 +156,12 @@ cc_binary {
     cflags: ["-Wall", "-Werror"],
     host_supported: true,
     shared_libs: ["libz"],
+    // We build zlib_bench32 and zlib_bench64 so it's easy to test LP32.
+    compile_multilib: "both",
+    multilib: {
+        lib32: { suffix: "32", },
+        lib64: { suffix: "64", },
+    },
 }
 
 // This module is defined in development/ndk/Android.bp. Updating these headers
-- 
cgit v1.2.3