Fix checking for Wasm SIMD support

The defined(WEBP_ENABLE_WASM_SIMD) in dsp.h meant that cpu.cmake will
never successfully compile with Wasm SIMD, since it doesn't set any
compile flags. We remove that requirement, and instead determine in
dec.c if we want to be using Wasm SIMD intrinsics.

Bug: v8:12371
Change-Id: I5d313f6d95301e6153254229a9ea234fec990cd6

CMakeLists.txt

@@ -22,6 +22,9 @@ if(NOT EMSCRIPTEN)
 endif()
 option(WEBP_ENABLE_SIMD "Enable any SIMD optimization."
        ${WEBP_ENABLE_SIMD_DEFAULT})
+# Emscripten supports SSE builds using its compatibility headers, by default it
+# will use SSE4 if WEBP_ENABLE_WASM_SIMD is OFF and WEBP_ENABLE_SIMD is ON.
+option(WEBP_ENABLE_WASM_SIMD "Enable WebAssembly SIMD optimizations" OFF)
 option(WEBP_BUILD_ANIM_UTILS "Build animation utilities." ON)
 option(WEBP_BUILD_CWEBP "Build the cwebp command line tool." ON)
 option(WEBP_BUILD_DWEBP "Build the dwebp command line tool." ON)
@@ -48,7 +51,6 @@ endif()
 if(WEBP_BUILD_WEBP_JS)
   set(WEBP_BUILD_ANIM_UTILS OFF)
   set(WEBP_BUILD_CWEBP OFF)
-  set(WEBP_BUILD_DWEBP OFF)
   set(WEBP_BUILD_GIF2WEBP OFF)
   set(WEBP_BUILD_IMG2WEBP OFF)
   set(WEBP_BUILD_VWEBP OFF)
@@ -80,6 +82,9 @@ include(GNUInstallDirs)
 if(WEBP_ENABLE_SWAP_16BIT_CSP)
   add_definitions(-DWEBP_SWAP_16BIT_CSP=1)
 endif()
+if(WEBP_ENABLE_WASM_SIMD)
+  add_definitions(-DWEBP_ENABLE_WASM_SIMD_INTRINSICS)
+endif()
 
 if(NOT WEBP_BITTRACE STREQUAL "0")
   add_definitions(-DBITTRACE=${WEBP_BITTRACE})

cmake/config.h.in

@@ -108,6 +108,9 @@
 /* Set to 1 if JPEG library is installed */
 #cmakedefine WEBP_HAVE_JPEG 1
 
+/* Set to 1 if Wasm SIMD is supported */
+#cmakedefine WEBP_HAVE_WASM_SIMD
+
 /* Set to 1 if NEON is supported */
 #cmakedefine WEBP_HAVE_NEON
 

cmake/cpu.cmake

@@ -36,9 +36,9 @@ function(webp_check_compiler_flag WEBP_SIMD_FLAG ENABLE_SIMD)
 endfunction()
 
 # those are included in the names of WEBP_USE_* in c++ code.
-set(WEBP_SIMD_FLAGS "SSE41;SSE2;MIPS32;MIPS_DSP_R2;NEON;MSA")
+set(WEBP_SIMD_FLAGS "WASM_SIMD;SSE41;SSE2;MIPS32;MIPS_DSP_R2;NEON;MSA")
 set(WEBP_SIMD_FILE_EXTENSIONS
-    "_sse41.c;_sse2.c;_mips32.c;_mips_dsp_r2.c;_neon.c;_msa.c")
+    "_wasm.c;_sse41.c;_sse2.c;_mips32.c;_mips_dsp_r2.c;_neon.c;_msa.c")
 if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   # With at least Visual Studio 12 (2013)+ /arch is not necessary to build SSE2
   # or SSE4 code unless a lesser /arch is forced. MSVC does not have a SSE4
@@ -53,9 +53,9 @@ if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   set(SIMD_DISABLE_FLAGS)
 else()
   set(SIMD_ENABLE_FLAGS
-      "-msse4.1;-msse2;-mips32;-mdspr2;-mfpu=neon;-mmsa")
+      "-msimd128;-msse4.1;-msse2;-mips32;-mdspr2;-mfpu=neon;-mmsa")
   set(SIMD_DISABLE_FLAGS
-      "-mno-sse4.1;-mno-sse2;;-mno-dspr2;;-mno-msa")
+      "-mno-simd128;-mno-sse4.1;-mno-sse2;;-mno-dspr2;;-mno-msa")
 endif()
 
 set(WEBP_SIMD_FILES_TO_NOT_INCLUDE)
@@ -77,7 +77,12 @@ math(EXPR WEBP_SIMD_FLAGS_RANGE "${WEBP_SIMD_FLAGS_LENGTH} - 1")
 foreach(I_SIMD RANGE ${WEBP_SIMD_FLAGS_RANGE})
   # With Emscripten 2.0.9 -msimd128 -mfpu=neon will enable NEON, but the
   # source will fail to compile.
-  if(EMSCRIPTEN AND ${I_SIMD} GREATER_EQUAL 2)
+  if(EMSCRIPTEN AND ${I_SIMD} GREATER_EQUAL 5)
+    break()
+  endif()
+  # Emscripten supports SSE via compat headers, if WEBP_ENABLED_WASM_SIMD is
+  # specified skip testing those (because it will succeed).
+  if (EMSCRIPTEN AND ${I_SIMD} GREATER_EQUAL 1 AND ${WEBP_ENABLE_WASM_SIMD})
     break()
   endif()
 
@@ -91,6 +96,7 @@ foreach(I_SIMD RANGE ${WEBP_SIMD_FLAGS_RANGE})
   webp_check_compiler_flag(${WEBP_SIMD_FLAG} ${WEBP_ENABLE_SIMD})
   if(NOT WEBP_HAVE_${WEBP_SIMD_FLAG})
     list(GET SIMD_ENABLE_FLAGS ${I_SIMD} SIMD_COMPILE_FLAG)
+    # This enables using Emscripten's SSE compatibility headers.
     if(EMSCRIPTEN)
       set(SIMD_COMPILE_FLAG "-msimd128 ${SIMD_COMPILE_FLAG}")
     endif()

src/dsp/Makefile.am

@@ -13,6 +13,7 @@ noinst_LTLIBRARIES += libwebpdsp_mips32.la
 noinst_LTLIBRARIES += libwebpdspdecode_mips32.la
 noinst_LTLIBRARIES += libwebpdsp_mips_dsp_r2.la
 noinst_LTLIBRARIES += libwebpdspdecode_mips_dsp_r2.la
+noinst_LTLIBRARIES += libwebpdspdecode_wasm.la
 
 if BUILD_LIBWEBPDECODER
   noinst_LTLIBRARIES += libwebpdspdecode.la
@@ -106,6 +107,8 @@ libwebpdspdecode_mips_dsp_r2_la_SOURCES += yuv_mips_dsp_r2.c
 libwebpdspdecode_mips_dsp_r2_la_CPPFLAGS = $(libwebpdsp_mips_dsp_r2_la_CPPFLAGS)
 libwebpdspdecode_mips_dsp_r2_la_CFLAGS = $(libwebpdsp_mips_dsp_r2_la_CFLAGS)
 
+libwebpdspdecode_wasm_la_SOURCES = dec_wasm.c
+
 libwebpdsp_sse2_la_SOURCES =
 libwebpdsp_sse2_la_SOURCES += cost_sse2.c
 libwebpdsp_sse2_la_SOURCES += enc_sse2.c

src/dsp/cpu.c

@@ -26,8 +26,10 @@
 // SSE2 detection.
 //
 
+// Skip SSE detection if using Wasm SIMD build.
+#if defined(WEBP_USE_WASM_SIMD)
 // apple/darwin gcc-4.0.1 defines __PIC__, but not __pic__ with -fPIC.
-#if (defined(__pic__) || defined(__PIC__)) && defined(__i386__)
+#elif (defined(__pic__) || defined(__PIC__)) && defined(__i386__)
 static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
   __asm__ volatile (
     "mov %%ebx, %%edi\n"
@@ -69,8 +71,10 @@ static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
 
 #endif
 
+// Skip xgetbv definition if using Wasm SIMD build.
+#if defined(WEBP_USE_WASM_SIMD)
 // NaCl has no support for xgetbv or the raw opcode.
-#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
+#elif !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
 static WEBP_INLINE uint64_t xgetbv(void) {
   const uint32_t ecx = 0;
   uint32_t eax, edx;
@@ -100,7 +104,13 @@ static WEBP_INLINE uint64_t xgetbv(void) {
 #define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
 #endif
 
-#if defined(__i386__) || defined(__x86_64__) || defined(WEBP_HAVE_MSC_CPUID)
+#if defined(WEBP_USE_WASM_SIMD)
+static int wasmCPUInfo(CPUFeature feature) {
+  if (feature != kWasmSIMD) return 0;
+  return 1;
+}
+VP8CPUInfo VP8GetCPUInfo = wasmCPUInfo;
+#elif defined(__i386__) || defined(__x86_64__) || defined(WEBP_HAVE_MSC_CPUID)
 
 // helper function for run-time detection of slow SSSE3 platforms
 static int CheckSlowModel(int info) {

src/dsp/dec.c

@@ -740,6 +740,7 @@ extern void VP8DspInitNEON(void);
 extern void VP8DspInitMIPS32(void);
 extern void VP8DspInitMIPSdspR2(void);
 extern void VP8DspInitMSA(void);
+extern void VP8DspInitWasmSIMD(void);
 
 WEBP_DSP_INIT_FUNC(VP8DspInit) {
   VP8InitClipTables();
@@ -831,6 +832,12 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) {
     if (VP8GetCPUInfo(kMSA)) {
       VP8DspInitMSA();
     }
+#endif
+#if defined(WEBP_USE_WASM_SIMD) && defined(WEBP_ENABLE_WASM_SIMD_INTRINSICS)
+    // Check that SIMD is supported and that we want to use Wasm intrinsics.
+    if (VP8GetCPUInfo(kWasmSIMD)) {
+      VP8DspInitWasmSIMD();
+    }
 #endif
   }
 

src/dsp/dec_wasm.c

@@ -0,0 +1,32 @@
+// Copyright 2021 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// WebAssembly (Wasm) version of some decoding functions.
+//
+// This will contain Wasm implementation of some decoding functions.
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_WASM_SIMD)
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitWasmSIMD(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitWasmSIMD(void) {
+  // TODO(crbug.com/v8/12371): No special implementation for Wasm yet, will be
+  // added later.
+}
+
+#else
+
+WEBP_DSP_INIT_STUB(VP8DspInitWasmSIMD)
+
+#endif  // WEBP_USE_WASM_SIMD

src/dsp/dsp.h

@@ -80,6 +80,11 @@ extern "C" {
 #endif
 #endif
 
+#if (defined(EMSCRIPTEN) || defined(WEBP_HAVE_WASM_SIMD)) && \
+    defined(__wasm_simd128__)
+#define WEBP_USE_WASM_SIMD
+#endif
+
 // WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
 // files without intrinsics, allowing the corresponding Init() to be called.
 // Files containing intrinsics will need to be built targeting the instruction
@@ -257,7 +262,8 @@ typedef enum {
   kNEON,
   kMIPS32,
   kMIPSdspR2,
-  kMSA
+  kMSA,
+  kWasmSIMD
 } CPUFeature;
 // returns true if the CPU supports the feature.
 typedef int (*VP8CPUInfo)(CPUFeature feature);