Issue 357973002: ARM64: avoid duplicate load of double constants

Keyboard Shortcuts

	File
u :	up to issue
j / k :	jump to file after / before current file
J / K :	jump to next file with a comment after / before current file
	Side-by-side diff
i :	toggle intra-line diffs
e :	expand all comments
c :	collapse all comments
s :	toggle showing all comments
n / p :	next / previous diff chunk or comment
N / P :	next / previous comment
<Up> / <Down> :	next / previous line

	Issue
u :	up to list of issues
j / k :	jump to patch after / before current patch
o / <Enter> :	open current patch in side-by-side view
i :	open current patch in unified diff view

	Issue List
j / k :	jump to issue after / before current issue
o / <Enter> :	open current issue

Issue 357973002: ARM64: avoid duplicate load of double constants (Closed)

Created:
6 years, 6 months ago by vincent.belliard

Modified:
6 years, 6 months ago

Reviewers:
Benedikt Meurer, ulan

CC:
v8-dev

Base URL:
https://v8.googlecode.com/svn/branches/bleeding_edge

Visibility:
Public.

More Reviews

Description

ARM64: avoid duplicate load of double constants R=ulan@chromium.org, bmeurer@chromium.org BUG=

Patch Set 1 #

Created: 6 years, 6 months ago

Download [raw] [tar.bz2]

		Unified diffs	Side-by-side diffs	Delta from patch set	Stats (+8 lines, -20 lines)			Patch
	M	src/arm64/lithium-arm64.cc	View		1 chunk	+7 lines, -9 lines	0 comments	Download
	M	src/arm64/lithium-codegen-arm64.cc	View		1 chunk	+1 line, -11 lines	0 comments	Download

Messages

Total messages: 2 (0 generated)

Expand Messages | Collapse Messages

vincent.belliard

Double constants are always load with HConstant. For compare-numeric-and-branch, the constant was also load. This ...

6 years, 6 months ago (2014-06-26 14:14:58 UTC) #1

Double constants are always load with HConstant. For compare-numeric-and-branch,
the constant was also load. This patch remove the second load.

Seems to add 0.3% performance to kraken and a very little performance to octane.

Before patch:
                  ;;; <@40,#358> constant-d
0x441b8580   192  fmov d2, xzr     ;; debug: position 3480
                  ;;; <@43,#46> compare-numeric-and-branch
0x441b8584   196  fcmp d0, #0.0
0x441b8588   200  b.vs #+0x8 (addr 0x441b8590)
0x441b858c   204  b.lt #+0x44 (addr 0x441b85d0)
                  ;;; <@44,#50> -------------------- B4 (unreachable/replaced)
--------------------
                  ;;; <@48,#53> -------------------- B5 --------------------
                  ;;; <@50,#54> constant-d
0x441b8590   208  movz x16, #0x2d19 ;; debug: position 3485
0x441b8594   212  movk x16, #0x5444, lsl #16
0x441b8598   216  movk x16, #0x21fb, lsl #32
0x441b859c   220  movk x16, #0x3ff9, lsl #48
0x441b85a0   224  fmov d3, x16
                  ;;; <@53,#55> compare-numeric-and-branch
0x441b85a4   228  movz x16, #0x2d19
0x441b85a8   232  movk x16, #0x5444, lsl #16
0x441b85ac   236  movk x16, #0x21fb, lsl #32
0x441b85b0   240  movk x16, #0x3ff9, lsl #48
0x441b85b4   244  fmov d30, x16
0x441b85b8   248  fcmp d0, d30
0x441b85bc   252  b.vs #+0x8 (addr 0x441b85c4)
0x441b85c0   256  b.gt #+0x10 (addr 0x441b85d0)

After patch:
                  ;;; <@40,#358> constant-d
0x5cfb8580   192  fmov d2, xzr   ;; debug: position 3480
                  ;;; <@43,#46> compare-numeric-and-branch
0x5cfb8584   196  fcmp d0, d2
0x5cfb8588   200  b.vs #+0x8 (addr 0x5cfb8590)
0x5cfb858c   204  b.lt #+0x30 (addr 0x5cfb85bc)
                  ;;; <@44,#50> -------------------- B4 (unreachable/replaced)
--------------------
                  ;;; <@48,#53> -------------------- B5 --------------------
                  ;;; <@50,#54> constant-d
0x5cfb8590   208  movz x16, #0x2d19  ;; debug: position 3485
0x5cfb8594   212  movk x16, #0x5444, lsl #16
0x5cfb8598   216  movk x16, #0x21fb, lsl #32
0x5cfb859c   220  movk x16, #0x3ff9, lsl #48
0x5cfb85a0   224  fmov d3, x16
                  ;;; <@53,#55> compare-numeric-and-branch
0x5cfb85a4   228  fcmp d0, d3
0x5cfb85a8   232  b.vs #+0x8 (addr 0x5cfb85b0)
0x5cfb85ac   236  b.gt #+0x10 (addr 0x5cfb85bc)

lgtm

Expand Messages | Collapse Messages