[Bug gas/22871] Encode instructions of 64-bit operand without the REX_W bit

[address@hidden]

https://sourceware.org/bugzilla/show_bug.cgi?id=22871

--- Comment #19 from Linus Torvalds <address@hidden> ---
(In reply to Linus Torvalds from comment #18)
> 
> Very interesting. I can confirm that testb seems slower on Skylake too.

Oh no, I take that back.

There's something else going on.

Sometimes I get results like this:

./test
testb  : 98279608
testw  : 987334300
testl  : 101367392
testq  : 100247139

and then those are stable for a while.

I think there may be some random instruction queue issues going on too. Doing
another loop around the benchmark, I get:

Round 0
testb  : 24604514
testw  : 24822594
testl  : 25159723
testq  : 25236629
Round 1
testb  : 24582167
testw  : 26043817
testl  : 24716638
testq  : 23855638
Round 2
testb  : 26776404
testw  : 25030268
testl  : 24753978
testq  : 24595795
Round 3
testb  : 23413229
testw  : 26248700
testl  : 25172790
testq  : 24017507

so the numbers simply fluctuate a lot.

Using -O2 and making the LOOP count 10x bigger, I get

Round 0
testb  : 149632125
testw  : 149115753
testl  : 148144016
testq  : 148172064
Round 1
testb  : 147595062
testw  : 145581017
testl  : 146435807
testq  : 147235969
Round 2
testb  : 146678329
testw  : 147896478
testl  : 145252248
testq  : 147045806
Round 3
testb  : 147764220
testw  : 150151744
testl  : 151424862
testq  : 149017052

so it's all really just odd noise. Don't ask me what's going on, but I don't
think it's that 'testb' is slower.

-- 
You are receiving this mail because:
You are on the CC list for the bug.