Bug 1499682 - SIMD-accelerate the data state in the HTML tokenizer. r=smaug,sergesanspaille

Other tokenizer states and the serializer are potential follow-ups.

The code movement from nsHtml5Tokenizer.cpp to nsHtml5Tokeniner.h is
for enabling the eventual non-unified build of nsHtml5TokenizerSIMD.cpp
once the LLVM bug has been fixed.

Differential Revision: https://phabricator.services.mozilla.com/D227317

Author: hsivonen

build/moz.configure/toolchain.configure

@@ -3914,6 +3914,35 @@ set_config(
     ),
 )
 
+
+@depends(target, c_compiler)
+def htmlaccel_config(target, c_compiler):
+    # Keep this is sync with the mozilla::htmlaccel::htmlaccelEnabled function.
+    #
+    # The code compiles on SSSE3, but AVX+BMI generates better code
+    # and has been available for 12 years at the time of landing this,
+    # so let's give the best code to users with reasonably recent hardware.
+    #
+    # Not enabled on 32-bit x86, due to lack of insight into what hardware is
+    # representative at this point in time and due to lack of such hardware
+    # for testing to see what config would actually be an optimization.
+    #
+    # aarch64 does not need extra flags.
+    #
+    # clang-cl doesn't tolerate -flax-vector-conversions but GCC requires it.
+    #
+    # -mavx2 doesn't change codegen vs. -mavx. AVX2 and BMI always co-occur
+    # in Intel CPUs, but there are AMD CPUs that have AVX and BMI without
+    # AVX2.
+    if target.cpu != "x86_64":
+        return []
+    if c_compiler.type == "gcc":
+        return ["-mavx", "-mbmi", "-flax-vector-conversions"]
+    return ["-mavx", "-mbmi"]
+
+
+set_config("HTML_ACCEL_FLAGS", htmlaccel_config)
+
 # dtrace support
 ##
 option("--enable-dtrace", help="Build with dtrace support")

parser/html/javasrc/Tokenizer.java

@@ -85,6 +85,42 @@ UNIFIED_SOURCES += [
     "nsParserUtils.cpp",
 ]
 
+# Each target needs to compile:
+# (nsHtml5TokenizerALU.cpp XOR nsHtml5TokenizerALUStubs.cpp)
+# AND
+# (nsHtml5TokenizerSIMD.cpp XOR nsHtml5TokenizerSIMDStubs.cpp)
+# AND
+# (nsHtml5TokenizerALU.cpp OR nsHtml5TokenizerSIMD.cpp)
+#
+# Make sure the result is consistent with mozilla::htmlaccel::htmlaccelEnabled().
+#
+# Due to https://github.com/llvm/llvm-project/issues/160886, none of the
+# code here actually ends up with SIMD instructions, and SIMD stays in
+# htmlaccelNotInline.cpp instead. Once the LLVM bug is fixed, the functions
+# in htmlaccelNotInline.cpp should becomed always inlined and
+# nsHtml5TokenizerSIMD.cpp should be built with HTML_ACCEL_FLAGS.
+
+if (CONFIG["TARGET_CPU"] == "x86_64") and (
+    CONFIG["CC_TYPE"] != "gcc" or int(CONFIG["CC_VERSION"].split(".")[0]) >= 12
+):
+    UNIFIED_SOURCES += [
+        "nsHtml5TokenizerALU.cpp",
+        "nsHtml5TokenizerSIMD.cpp",
+    ]
+elif (
+    CONFIG["TARGET_CPU"] == "aarch64" and CONFIG["TARGET_ENDIANNESS"] == "little"
+) and (CONFIG["CC_TYPE"] != "gcc" or int(CONFIG["CC_VERSION"].split(".")[0]) >= 12):
+    # aarch64 doesn't need special flags for SIMD.
+    UNIFIED_SOURCES += [
+        "nsHtml5TokenizerALUStubs.cpp",
+        "nsHtml5TokenizerSIMD.cpp",
+    ]
+else:
+    UNIFIED_SOURCES += [
+        "nsHtml5TokenizerALU.cpp",
+        "nsHtml5TokenizerSIMDStubs.cpp",
+    ]
+
 FINAL_LIBRARY = "xul"
 
 LOCAL_INCLUDES += [