608355 - JM is slow on ai-astar on 64-bit

Status: RESOLVED INVALID

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

Attachment: Testcase

On Mac, I see times around 1s for -j and -m on the attached testcase in 32-bit shell, and for -j in 64-bit shell.  The 64-bit -m time is 2.5s.

Comment 1

[Boris Zbarsky [:bzbarsky]]

My profiler tells me "all your time are belong to mjit-generated gunk I know nothing about"

Comment 2

[Boris Zbarsky [:bzbarsky]]

Er, make that 1.5s for -m; I presumably had a somewhat old shell build...  That would match billm's times on Linux too.

Comment 3

[Boris Zbarsky [:bzbarsky]]

Though if I believe shark, then the majority of the time is spent on exactly three instructions:

0x1003f0c97	add      byte ptr [r8], al
0x1003f0cb1	add      byte ptr [r8], r8b
0x1003f1717	sldt     qword ptr [rax]

Those add up to 75% of the time, is the claim, with another 3% at:

0x1003f0ca9	add      byte ptr [r8], r8b

and then a long tail.

Comment 4

[Boris Zbarsky [:bzbarsky]]

On 32-bit, there are only two hot instructions:

0x3ff8a1	mov      eax, dword ptr [eax]
0x3ff88c	add      dword ptr [eax], dword 301026508

Comment 5

[Boris Zbarsky [:bzbarsky]]

Running this in 64-bit TM, I get these as the two hot instructions:

0x1003f6ee0	add      byte ptr [rax], r8b
0x1003f6ed0	add      byte ptr [rax], al

So the same thing, but with rax instead of r8, and without sldt....

Comment 6

[Boris Zbarsky [:bzbarsky]]

And the 32-bit TM, for completeness:

0x5d2345	add      dword ptr [eax], dword 266174464
0x5d2333	add      dword ptr [eax], dword 163164979

I dunno how much I trust this data at this point.... ;)

Comment 7

[Till Schneidereit [:till]]

JM isn't slow at anything, anymore.

Comment 8

[Boris Zbarsky [:bzbarsky]]

Indeed.  For what it's worth, we're 2x faster than V8 on the attached testcase.  ;)