462522 - Excessive register shuffling after call sites

Status: VERIFIED FIXED

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect)

Description

[Edwin Smith]

number2 = fcall number ( #core doubleToAtom_sse2_3 )
              push eax            eax(doubleToAtom_sse2_3) edi(ldc1)
              mov ecx,14000168    edi(ldc1)
              call number         edi(ldc1)
              mov esi,edi         edi(ldc1)
->            mov edi,-12(ebp)
->            mov ebx,esi         esi(ldc1) edi(param1)
->            mov esi,-8(ebp)     edi(param1)
->            mov ecx,ebx         ebx(ldc1) esi(param2) edi(param1)
->            mov ebx,-4(ebp)     esi(param2) edi(param1)
              mov eax,4           ecx(ldc1) ebx(param3) esi(param2) edi(param1)

in this case, ecx (ldc1) is the highest priority live var, so we want to spill one of ebx, esi, edi.  We do that by spilling ebx, however this makes ebx now the highest priority register, which then spills to esi, etc.  domino effect.

Comment 1

[Edwin Smith]

Attachment: Fixed the problem, dont be silly with the cascading spills

I'm seeing 1-3% speedup on macos10.5 (x86) from this fix.  curious to see how TM fares.

Comment 2

[Edwin Smith]

changeset:   1043:3e22c99054a9

Comment 3

[Andreas Gal :gal]

grayson, could you give this a try I am on the road

Comment 4

[Graydon Hoare :graydon]

I did. No perceived difference on shell-level sunspider. You want me to check browser-level?

Comment 5

[Andreas Gal :gal]

Nah I doubt it makes a difference there. But we should apply the patch anyway