Changeset 287236 in webkit for trunk/Source/JavaScriptCore/jit/JIT.cpp

Timestamp:

Dec 19, 2021, 12:33:06 AM (3 years ago)

Author:

Ross Kirsling

Message:

[JSC] OpPow should have a "small int exponent" fast path at lower tiers
​https://bugs.webkit.org/show_bug.cgi?id=234408

Reviewed by Yusuke Suzuki.

JSTests:

• microbenchmarks/pow-double-int.js: Added.
• microbenchmarks/pow-int-int.js: Added.

Source/JavaScriptCore:

DFG has an ArithPow fast path which just multiplies in a loop when the exponent is an int between 0 and 1000;
this can be done at lower tiers too.

Implementing this at LLInt gives the following speedup with JIT disabled:

Before After

pow-int-int 193.7180+-0.4897 ^{100.3569+-1.9804} definitely 1.9303x faster
pow-double-int 194.0744+-0.7998 ^{100.0346+-0.8655} definitely 1.9401x faster

<geometric> 193.8824+-0.4667 ^{100.0964+-0.9922} definitely 1.9370x faster

Implementing this at Baseline gives similar results with DFG disabled:

Before After

pow-int-int 195.6251+-0.9577 ^{99.9627+-0.3307} definitely 1.9570x faster
pow-double-int 196.1975+-0.9307 ^{101.0056+-0.3124} definitely 1.9424x faster

<geometric> 195.8786+-0.5883 ^{100.4767+-0.2333} definitely 1.9495x faster

Results are neutral otherwise.

• jit/JIT.cpp:

(JSC::JIT::privateCompileMainPass):
(JSC::JIT::privateCompileSlowCases):

• jit/JIT.h:
• jit/JITArithmetic.cpp:

(JSC::JIT::emit_op_pow):
(JSC::JIT::emitSlow_op_pow):

• llint/LowLevelInterpreter.asm:
• llint/LowLevelInterpreter32_64.asm:
• llint/LowLevelInterpreter64.asm:

File:

• trunk/Source/JavaScriptCore/jit/JIT.cpp (modified) (3 diffs)

trunk/Source/JavaScriptCore/jit/JIT.cpp

@@ -344 +344 @@
         DEFINE_SLOW_OP(create_async_generator)
         DEFINE_SLOW_OP(new_generator)
-        DEFINE_SLOW_OP(pow)
 
         DEFINE_OP(op_add)
@@ -436 +435 @@
         DEFINE_OP(op_lshift)
         DEFINE_OP(op_mod)
+        DEFINE_OP(op_pow)
         DEFINE_OP(op_mov)
         DEFINE_OP(op_mul)
@@ -617 +617 @@
         DEFINE_SLOWCASE_OP(op_check_traps)
         DEFINE_SLOWCASE_OP(op_mod)
+        DEFINE_SLOWCASE_OP(op_pow)
         DEFINE_SLOWCASE_OP(op_mul)
         DEFINE_SLOWCASE_OP(op_negate)