small optimization idea

spacy51 · Post by **spacy51** » Mon Aug 17, 2009 1:16 pm

gba-thumb.cpp @l1587:

// LDR Rd, [Rs, #Imm]
static INSN_REGPARM void thumb68(u32 opcode)

{

if (busPrefetchCount == 0)
Code: Select all
busPrefetch = busPrefetchEnable;
u32 address = reg[(opcode>>3)&7].I + b<<2)[/b];

reg[opcode&7].I = CPUReadMemory(address);

clockTicks = 3 + dataTicksAccess32(address) + codeTicksAccess16(armNextPC);

}

Couldn't the highlighted block be optimized by replacing it with

code[/code]

Â

There are several cases like this around that area.

Post by **ZachBacon** » Mon Aug 17, 2009 3:46 pm

I'll give it a try and see what happens [img]<fileStore.core_Emoticons>/emoticons/smile.png[/img]/emoticons/smile@2x.png 2x" width="20" height="20" />

Post by **ZachBacon** » Mon Aug 17, 2009 3:54 pm

ok tried

u32 address = reg[(opcode>>3)&7].I + (((opcode>>4)&124)<<2); and it didn't work (I know not what you said to try)

so then I tried

u32 address = reg[(opcode>>3)&7].I + (((opcode>>4)&124); compile didn't work since it was missing a bracket (that was kind of your fault for not giving the extra bracket [img]<fileStore.core_Emoticons>/emoticons/tongue.png[/img]/emoticons/tongue@2x.png 2x" width="20" height="20" />)

once the extra bracket was added it worked again

u32 address = reg[(opcode>>3)&7].I + (((opcode>>4)&124));

and it was fairly optimized [img]<fileStore.core_Emoticons>/emoticons/smile.png[/img]/emoticons/smile@2x.png 2x" width="20" height="20" />

spacy51 · Post by **spacy51** » Mon Aug 17, 2009 4:08 pm

My replacement should be perfectly fine [img]<fileStore.core_Emoticons>/emoticons/huh.png[/img]/emoticons/huh@2x.png 2x" width="20" height="20" />

Â

u32 address = reg[(opcode>>3)&7].I + b)[/b];

Â

Do you seriously want to tell me there have to be two brackets???

:blah!:

Â

It would be interesting to see the actual speed differences, if all of those similar code pieces were edited like that.

Post by **ZachBacon** » Mon Aug 17, 2009 4:09 pm

well I was tired and didn't notice >_< I didn't have my coffee at all but yeah I say let's do this as an experiment first before submitting it to svn and if the others like it then submit it

spacy51 · Post by **spacy51** » Mon Aug 17, 2009 4:25 pm

It would be cool if some binary junky can tell me if the expression (((opcode>>6)&31)<<2) always gives the same results as ((opcode>>4)&124). I only had some basic binary term transformation in school ^^. I really don't want to introduce bugs like I did before I took a break [img]<fileStore.core_Emoticons>/emoticons/laugh.png[/img]/emoticons/laugh@2x.png 2x" width="20" height="20" />

jbo_85 · Post by **jbo_85** » Mon Aug 17, 2009 4:28 pm

It would be cool if some binary junky can tell me if the expression (((opcode>>6)&31)<<2) always gives the same results as ((opcode>>4)&124). I only had some basic binary term transformation in school ^^. I really don't want to introduce bugs like I did before I took a break [img]<fileStore.core_Emoticons>/emoticons/laugh.png[/img]/emoticons/laugh@2x.png 2x" width="20" height="20" />

The two expressions are the same. But it doesn't make a difference for the release build because the compiler already optimizes it.

spacy51 · Post by **spacy51** » Mon Aug 17, 2009 5:37 pm

Is there any proof that the compiler optimizes that? Hm, maybe it's possible to look at the assembler code the compiler produced.

jbo_85 · Post by **jbo_85** » Mon Aug 17, 2009 5:43 pm

I compared the executables.

spacy51 · Post by **spacy51** » Mon Aug 17, 2009 6:04 pm

OK, good to know, thanks.