Bug 141548 – Generate incq instead of addq when the immediate value is one

• Status: RESOLVED
• Resolution: FIXED
• Component: New Bugs

[Benjamin Poulain]

Generate incq instead of addq when the immediate value is one

[Benjamin Poulain]

Created attachment 246499 [details]
Patch

[Gavin Barraclough]

Comment on attachment 246499 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=246499&action=review

Have you performance tested? – inc used to be slower on some old intel chips (since it only sets 5/6 of the main flags) so add was often preferable – though that may well no longer be the case.

> Source/JavaScriptCore/assembler/X86Assembler.h:566
> +        m_formatter.oneByteOp64(OP_GROUP5_Ev, GROUP1_OP_ADD, base, offset);

Shouldn't really be using 'GROUP1_OP_ADD' here – I think this should use 'GROUP5_OP_INC' (defining it, if it doesn't exist).

[Benjamin Poulain]

(In reply to comment #2)
> Comment on attachment 246499 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=246499&action=review
> 
> Have you performance tested? – inc used to be slower on some old intel chips
> (since it only sets 5/6 of the main flags) so add was often preferable –
> though that may well no longer be the case.

I was intrigued by your comment so I looked into the perf tables all the way to pentium. For throughput and latency, Inc was equal or better than Add in all the cases I looked at.

I checked the Intel Perf guideline, it says: "INC and DEC instructions should be replaced with ADD or SUB instructions, because ADD and SUB overwrite all flags, whereas INC and DEC do not, therefore creating false dependencies on earlier instructions that set the flags."
-> exactly what you were saying.

I checked what other compilers do. GCC seems to use ADD systematically. Clang uses INC systematically. ICC uses INC systematically.

I'll have to test on our benchmarks...

[Benjamin Poulain]

Committed r180075: <http://trac.webkit.org/changeset/180075>