Changeset 206392 in webkit for trunk/Source/JavaScriptCore/ChangeLog

Timestamp:

Sep 26, 2016, 2:11:31 PM (9 years ago)

Author:

[email protected]

Message:

[JSC] Shrink the Math inline caches some more
​https://bugs.webkit.org/show_bug.cgi?id=162485

Reviewed by Saam Barati.

Source/JavaScriptCore:

This patch applies some lessons learnt from op_negate
to shrink the generated asm of the previous 3 inline
caches.

In order of importance:
-We do not need to pass the pointer to ArithProfile

on the slow path. We can just get the profile out
of the Math IC.
This saves us from materializing a 64bits value
in a register before the call on the slow path.

-We can remove a bunch of mov by setting up the registers

in the way the slow path needs them.
The slow path makes a function calls with the input
as second and third arguments, and return the result in
the "return register". By using those as target when
loading/storing from the stack, we remove 3 mov per slow path.

-When performing integer add, we can set the result directly in

the output register if that does not trashes one of the input
register. This removes one mov per integer add.

The inline cache average sizes on Sunspider change as follow:
-Adds: 147.573099->131.555556 (~10%)
-Muls: 186.882353->170.991597 (~8%)
-Subs: 139.127907->121.523256 (~12%)

• jit/JIT.h:
• jit/JITAddGenerator.cpp:

(JSC::JITAddGenerator::generateInline):
(JSC::JITAddGenerator::generateFastPath):

• jit/JITArithmetic.cpp:

(JSC::JIT::emitMathICFast):
(JSC::JIT::emitMathICSlow):

• jit/JITInlines.h:

(JSC::JIT::callOperation): Deleted.

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

Source/WTF:

• wtf/Bag.h:

Don't copy the arguments before initializing the nodes.

File:

• trunk/Source/JavaScriptCore/ChangeLog (modified) (1 diff)

trunk/Source/JavaScriptCore/ChangeLog

@@ +1 @@
+2016-09-26  Benjamin Poulain  <[email protected]>
+
+        [JSC] Shrink the Math inline caches some more
+        https://bugs.webkit.org/show_bug.cgi?id=162485
+
+        Reviewed by Saam Barati.
+
+        This patch applies some lessons learnt from op_negate
+        to shrink the generated asm of the previous 3 inline
+        caches.
+
+        In order of importance:
+        -We do not need to pass the pointer to ArithProfile
+         on the slow path. We can just get the profile out
+         of the Math IC.
+         This saves us from materializing a 64bits value
+         in a register before the call on the slow path.
+        -We can remove a bunch of mov by setting up the registers
+         in the way the slow path needs them.
+         The slow path makes a function calls with the input
+         as second and third arguments, and return the result in
+         the "return register". By using those as target when
+         loading/storing from the stack, we remove 3 mov per slow path.
+        -When performing integer add, we can set the result directly in
+         the output register if that does not trashes one of the input
+         register. This removes one mov per integer add.
+
+        The inline cache average sizes on Sunspider change as follow:
+        -Adds: 147.573099->131.555556 (~10%)
+        -Muls: 186.882353->170.991597 (~8%)
+        -Subs: 139.127907->121.523256 (~12%)
+
+        * jit/JIT.h:
+        * jit/JITAddGenerator.cpp:
+        (JSC::JITAddGenerator::generateInline):
+        (JSC::JITAddGenerator::generateFastPath):
+        * jit/JITArithmetic.cpp:
+        (JSC::JIT::emitMathICFast):
+        (JSC::JIT::emitMathICSlow):
+        * jit/JITInlines.h:
+        (JSC::JIT::callOperation): Deleted.
+        * jit/JITOperations.cpp:
+        * jit/JITOperations.h:
+
 2016-09-26  Mark Lam  <[email protected]>