[X86] Atom scheduler generating incorrect throughputs and latencies

[topperc]

Bugzilla Link	36895
Version	trunk
OS	All
Blocks	#31672
CC	@legrosbuffle,@gchatelet,@RKSimon

Extended Description

There's lots of incorrect information

For example, ADDPS should have a latency of 5 and throughput of 1, but the scheduler test shows llvm reporting latency 5 and throughput 5.

llvm reports latency 125 and throughput 62.5 for sqrtpd, but I think it really should be a throughput of 124 or 125.

I think there are a lot more issues like this.

[RKSimon]

*** Bug llvm/llvm-bugzilla-archive#36792 has been marked as a duplicate of this bug. ***

[RKSimon]

@​gchatelet I think you mentioned you have access to some Bonnell era atom hardware - have you been able to get llvm-exegesis running on it?

[gchatelet]

I have a DiamondVille which is first generation Bonnell according to wikipedia.
I'll report here when we have it running.

[gchatelet]

Bonnel inconsistencies

[gchatelet]

I finally got it to run. I attached the report to the bug.

[gchatelet]

Results are weird. It may be that the we're frontend bound on this CPU.

[RKSimon]

Thanks Guillaume - do you have the raw latency/uops yaml data please?

[gchatelet]

Latency YAML file

[RKSimon]

D525 - llvm-exegesis report
Attached is a recent exegesis run of latency, uops and inverse_throughput on a 64-bit D525 'Pineview' atom cpu

[RKSimon]

Results are weird. It may be that the we're frontend bound on this CPU.

Part of the problem seems to be due to Atom (and Silvermont etc) struggling to efficiently decode the longer SSE instruction encodings that reference the upper xmm8-xmm15 registers.