From 09d4f8ec9ccf9c415a97dc835c379f475c5db7ee Mon Sep 17 00:00:00 2001
From: Yuriy O'Donnell <36282596+yuriy-odonnell-epic@users.noreply.github.com>
Date: Mon, 25 Jul 2022 23:54:59 +0200
Subject: [PATCH] [blake3] Add assembly implementations (#25921)

* [blake3] Extend build script to support assembly implementations

* [blake3] Add AVX512 assembly implementation
---
 ports/blake3/CMakeLists.txt | 36 ++++++++++++++++++++++++++++++------
 ports/blake3/vcpkg.json     |  1 +
 versions/b-/blake3.json     |  5 +++++
 versions/baseline.json      |  2 +-
 4 files changed, 37 insertions(+), 7 deletions(-)

diff --git a/ports/blake3/CMakeLists.txt b/ports/blake3/CMakeLists.txt
index cc5f37e1fc..1ef6c3893f 100644
--- a/ports/blake3/CMakeLists.txt
+++ b/ports/blake3/CMakeLists.txt
@@ -10,22 +10,46 @@ target_sources(blake3 PRIVATE
     c/blake3_portable.c
 )
 
-# This is a bit too exotic to be worth supporting right now
-target_compile_definitions(blake3 PRIVATE BLAKE3_NO_AVX512=1)
-
-if(VCPKG_TARGET_ARCHITECTURE STREQUAL "x86" OR VCPKG_TARGET_ARCHITECTURE STREQUAL "x64")
+if(VCPKG_TARGET_ARCHITECTURE STREQUAL "x86")
+    target_compile_definitions(blake3 PRIVATE BLAKE3_NO_AVX512=1)
     target_sources(blake3 PRIVATE
         c/blake3_avx2.c
         c/blake3_sse2.c
         c/blake3_sse41.c
     )
-    if (NOT MSVC)
+    if(NOT MSVC)
         set_source_files_properties(c/blake3_avx2.c COMPILE_FLAGS -mavx2)
         set_source_files_properties(c/blake3_sse2.c COMPILE_FLAGS -msse2)
         set_source_files_properties(c/blake3_sse41.c COMPILE_FLAGS -msse4.1)
     endif()
+elseif(VCPKG_TARGET_ARCHITECTURE STREQUAL "x64")
+    if(WIN32 AND MSVC)
+        enable_language(ASM_MASM)
+        target_sources(blake3 PRIVATE
+            c/blake3_avx2_x86-64_windows_msvc.asm
+            c/blake3_sse2_x86-64_windows_msvc.asm
+            c/blake3_sse41_x86-64_windows_msvc.asm
+            c/blake3_avx512_x86-64_windows_msvc.asm
+        )
+    elseif(WIN32 AND NOT MSVC)
+        enable_language(ASM)
+        target_sources(blake3 PRIVATE
+            c/blake3_avx2_x86-64_windows_gnu.S
+            c/blake3_sse2_x86-64_windows_gnu.S
+            c/blake3_sse41_x86-64_windows_gnu.S
+            c/blake3_avx512_x86-64_windows_gnu.S
+        )
+    else()
+        enable_language(ASM)
+        target_sources(blake3 PRIVATE
+            c/blake3_avx2_x86-64_unix.S
+            c/blake3_sse2_x86-64_unix.S
+            c/blake3_sse41_x86-64_unix.S
+            c/blake3_avx512_x86-64_unix.S
+        )        
+    endif()
 else()
-    target_compile_definitions(blake3 PRIVATE BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1)
+    target_compile_definitions(blake3 PRIVATE BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1 BLAKE3_NO_AVX512=1)
 endif()
 
 if(VCPKG_TARGET_ARCHITECTURE STREQUAL "arm" OR VCPKG_TARGET_ARCHITECTURE STREQUAL "arm64")
diff --git a/ports/blake3/vcpkg.json b/ports/blake3/vcpkg.json
index 44c8020227..1c2d1b03ce 100644
--- a/ports/blake3/vcpkg.json
+++ b/ports/blake3/vcpkg.json
@@ -1,6 +1,7 @@
 {
   "name": "blake3",
   "version": "1.3.1",
+  "port-version": 1,
   "description": "BLAKE3 cryptographic hash function.",
   "homepage": "https://github.com/BLAKE3-team/BLAKE3",
   "license": "CC0-1.0 OR Apache-2.0",
diff --git a/versions/b-/blake3.json b/versions/b-/blake3.json
index 1a9effce0a..71f6cb45e6 100644
--- a/versions/b-/blake3.json
+++ b/versions/b-/blake3.json
@@ -1,5 +1,10 @@
 {
   "versions": [
+    {
+      "git-tree": "78cffdc59cdb9f1c75dffa671578203d0a79fa42",
+      "version": "1.3.1",
+      "port-version": 1
+    },
     {
       "git-tree": "62a5201ca424389c823b9b4be3d588e1cbb88a58",
       "version": "1.3.1",
diff --git a/versions/baseline.json b/versions/baseline.json
index d4628e1a42..2e39787bd5 100644
--- a/versions/baseline.json
+++ b/versions/baseline.json
@@ -502,7 +502,7 @@
     },
     "blake3": {
       "baseline": "1.3.1",
-      "port-version": 0
+      "port-version": 1
     },
     "blas": {
       "baseline": "1",