The current StringFromCharCode DFG intrinsic assumes that its operand charCode must be an Int32. This results in 26000+ BadType OSR exits in the LongSpider crypto-aes benchmark. With support for Untyped operands, the number of OSR exits drops to 202.
Created attachment 268828 [details] proposed patch. The patch passes the JSC tests. However, still need to run benchmarks on them.
Created attachment 268872 [details] x86_64 benchmark result.
Created attachment 268873 [details] x86 benchmark results.
The benchmarks shows that perf is neutral with this patch. All of the differences are due to noise (does not repeat on retry) except for one: ftl-polymorphic-StringFromCharCode which I just added for this patch: ftl-polymorphic-StringFromCharCode 246.9686+-17.4090 ^ 207.8919+-11.0696 ^ definitely 1.1880x faster This is only for 64 bit x86_64 though. It does not manifest on x86 because there appears to be a bug on 32-bit where the StringFromCharCode intrinsic is not recognized and therefore not deployed: https://bugs.webkit.org/show_bug.cgi?id=153066. This patch is ready for a review.
Comment on attachment 268828 [details] proposed patch. View in context: https://bugs.webkit.org/attachment.cgi?id=268828&action=review r=me > Source/JavaScriptCore/dfg/DFGOperations.cpp:1316 > +EncodedJSValue JIT_OPERATION operationStringFromValueCharCode(ExecState* exec, EncodedJSValue encodedValue) I would call this operationStringFromCharCodeUntyped. If we have to do this in more cases, we should create an operationToUInt32 instead.
Thanks for the review. Landed in r194996: <http://trac.webkit.org/r194996>.