438881 - Significant cache misses for string conversion routines

Status: RESOLVED WONTFIX

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Mingqiu Sun]

User-Agent:       Mozilla/5.0 (Windows; U; Windows NT 6.0; en-US; rv:1.8.1.14) Gecko/20080404 Firefox/2.0.0.14
Build Identifier: tamarin-central-228c19321257

We noticed significant cache misses in one of the UTF string conversion routines. By injecting the following prefetch loop, we are able to reduce the cache misses by about 40%:

+			int i;
+			uint8 *temp = out;
+			for (i=0; i< outMax ; i=i+256)  
+				_asm prefetchnta temp+i;

As a result, sunspider/string-unpack-code performance is improved by about 10%. This technique could be applied elsewhere too, where cache misses are a problem.

Reproducible: Always

Steps to Reproduce:
1. run sunspider
2. measure cache read and write misses with vtune
3. identify hot methods wrt cache misses
Actual Results:  
test                                                   avm    avm2     %sp

sunspider\access-binary-trees.as                      47.0    47.0     0.0
sunspider\access-fannkuch.as                         125.0   125.0     0.0
sunspider\access-nbody.as                            171.0   171.0     0.0
sunspider\access-nsieve.as                            62.0    62.0     0.0
sunspider\bitops-3bit-bits-in-byte.as                 15.0    15.0     0.0
sunspider\bitops-bits-in-byte.as                      16.0    16.0     0.0
sunspider\bitops-bitwise-and.as                      203.0   203.0     0.0
sunspider\bitops-nsieve-bits.as                       46.0    46.0     0.0
sunspider\controlflow-recursive.as                    46.0    46.0     0.0
sunspider\crypto-aes.as                               62.0    62.0     0.0
sunspider\crypto-md5.as                               31.0    31.0     0.0
sunspider\crypto-sha1.as                              31.0    31.0     0.0
sunspider\math-cordic.as                              93.0    93.0     0.0
sunspider\math-partial-sums.as                       203.0   203.0     0.0
sunspider\math-spectral-norm.as                      156.0   156.0     0.0
sunspider\s3d-cube.as                                109.0   109.0     0.0
sunspider\s3d-morph.as                                93.0    93.0     0.0
sunspider\s3d-raytrace.as                            125.0   125.0     0.0
sunspider\string-fasta.as                             93.0    93.0     0.0
sunspider\string-unpack-code.as                     7609.0  6906.0     9.2
sunspider\string-validate-input.as                    93.0    93.0     0.0


patch file:

diff -Naur ../tamarin-central-copy/core/UnicodeUtils.cpp core/UnicodeUtils.cpp
--- ../tamarin-central-copy/core/UnicodeUtils.cpp	2008-03-05 13:54:47.298446000 -0800
+++ core/UnicodeUtils.cpp	2008-05-22 11:36:56.626858500 -0700
@@ -73,6 +73,13 @@
 		int outLen = 0;
 		if (out)
 		{
+			#ifdef WIN32
+			int i;
+			uint8 *temp = out;
+			for (i=0; i< outMax ; i=i+256)  
+				_asm prefetchnta temp+i;
+			#endif
+
 			// Output buffer passed in; actually encode data.
 			while (inLen > 0)
 			{

Comment 1

[Mingqiu Sun]

Attachment: prefetch loop

Comment 2

[Michael Daumling]

I'll take it for now - I'll see if I can apply the same patch to the new strings.

Comment 3

[Steven Johnson]

Comment on attachment 324826 [details] [diff] [review]
prefetch loop

rejecting only because this assumes WIN32 == Intel, which isn't true in all cases (eg various WinCE builds). You probably want something that specifically asks for Visual-C-on-Intel:

#if defined(_MSC_VER) && (defined(AVMPLUS_IA32) || defined(AVMPLUS_AMD64))

Comment 4

[Mingqiu Sun]

Steven,

That is a very good point. Revised patch attached.

Comment 5

[Mingqiu Sun]

Attachment: revised pref patch based on comment

Comment 6

[Lars T Hansen]

Ready to land, I'll take it.

Comment 7

[Lars T Hansen]

Turns out we can't land this at present in its current form.  The Flash Player's minimum x86-32 requirement is MMX, not SSE1, yet SSE1 is required for the PREFETCHNTA instruction.  Since they do not build multiple static binaries but detect architectures at run-time this change would create a lot of complication for - at this time - relatively modest gain.

Comment 8

[Lars T Hansen]

Note however that the optimization may apply to various platforms and builds:

- mac intel systems are always sse2 or better
- x86-64 systems ditto

Comment 9

[Sylvestre Ledru [:Sylvestre]]

Tamarin is a dead project now. Mass WONTFIX.

Comment 10

[Sylvestre Ledru [:Sylvestre]]

Tamarin isn't maintained anymore. WONTFIX remaining bugs.