Changeset 252825 in webkit for trunk/Source/JavaScriptCore/jit/JITInlines.h

Timestamp:

Nov 22, 2019, 8:45:14 PM (6 years ago)

Author:

[email protected]

Message:

Use LLInt profiling to rule out generating an IC for get_by_val
​https://bugs.webkit.org/show_bug.cgi?id=204536

Reviewed by Yusuke Suzuki.

When I landed the get_by_val polymorphic inline caching patch, the prepack
benchmark in JetStream2 slowed down by 10%. Through some analysis, I found
out that we were slowing down because of the time we spent in the JITs
actually generating inline caches. This patch skips generating an inline
cache when it seems like it won't be profitable. The heuristic for doing this
is simple:

• If we see more than 4 identifiers in the LLInt, we won't generate an IC

in the upper tiers.

• If we see a non-identifier JSString in the LLInt, we won't generate an IC

in the upper tiers.

This patch recovers the regression on prepack.

• bytecode/BytecodeList.rb:
• dfg/DFGByteCodeParser.cpp:

(JSC::DFG::ByteCodeParser::parseBlock):

• dfg/DFGOperations.cpp:
• dfg/DFGOperations.h:
• generator/main.rb:
• heap/TinyBloomFilter.h:

(JSC::TinyBloomFilter::bits const):
(JSC::TinyBloomFilter::TinyBloomFilter):

• jit/JIT.h:
• jit/JITOperations.cpp:
• jit/JITOperations.h:
• jit/JITPropertyAccess.cpp:

(JSC::JIT::emit_op_get_by_val):
(JSC::JIT::emitSlow_op_get_by_val):

• llint/LLIntSlowPaths.cpp:

(JSC::LLInt::LLINT_SLOW_PATH_DECL):

• runtime/Operations.h:

(JSC::getByValWithIndex):

• runtime/OptionsList.h:

File:

• trunk/Source/JavaScriptCore/jit/JITInlines.h (modified) (1 diff)

trunk/Source/JavaScriptCore/jit/JITInlines.h

@@ -180 +180 @@
 }
 
+ALWAYS_INLINE bool JIT::hasAnySlowCases(Vector<SlowCaseEntry>& slowCases, Vector<SlowCaseEntry>::iterator& iter, BytecodeIndex bytecodeIndex)
+{
+    if (iter != slowCases.end() && iter->to == bytecodeIndex)
+        return true;
+    return false;
+}
+
 ALWAYS_INLINE void JIT::addSlowCase(Jump jump)
 {