[Date Prev][Date Next][Thread Prev][Thread Next]
[Date Index]
[Thread Index]
- Subject: Re: [ANN] llvm-lua 1.0
- From: Mike Pall <mikelu-0906@...>
- Date: Fri, 5 Jun 2009 15:40:56 +0200
David Given wrote:
>> In terms of Lua 5.1 bytecode this would replace GETGLOBAL+GETTABLE*
>> with a single LOADK. This ought to be *really* fast, both in the
>> interpreter and when compiled.
>
> For simple numerical stuff like trig functions, would the JIT manage to
> then convert them to inline code sequences where appropriate? For
> example, math.sin could theoretically become a single inline fsin
> instruction in the right circumstances. Or will it still bounce it via
> an external C function?
Both LJ1 and LJ2 inline math.sin and other functions -- even
without the proposed option. But they still have to do the two
table lookups. LJ2 is able to hoist these out of the loop, but
it's of course still done in the loop header.
The modification would allow both compilers to drop the table
lookups and the function dispatch check. This would go mostly
unnoticed for LJ2, but it would help to reduce the I-cache
footprint of the generated code.
Trivial example:
local x=2; for i=1,100 do x=math.sqrt(x) end
LJ1:
[... start of loop before this ...]
--0006-- GETGLOBAL 5 -4 ; "math"
f7d217ef BAE018EB09 mov edx,0x09eb18e0
f7d217f4 83C350 add ebx,byte +0x50
f7d217f7 E804FCFFFF call 0xf7d21400 ->GETGLOBAL
--0007-- GETTABLE 5 5 -5 ; "sqrt"
f7d217fc 8D7B50 lea edi,[ebx+0x50]
f7d217ff BA4023EB09 mov edx,0x09eb2340
f7d21804 83C350 add ebx,byte +0x50
f7d21807 E804FCFFFF call 0xf7d21410 ->GETTABLE_KSTR
[...]
--0009-- CALL 5 2 2
f7d2181b 837B5806 cmp dword [ebx+0x58],byte +0x06
f7d2181f 7562 jnz 0xf7d21883
f7d21821 817B502023EB09 cmp dword [ebx+0x50],0x09eb2320
f7d21828 7559 jnz 0xf7d21883
f7d2182a 837B6803 cmp dword [ebx+0x68],byte +0x03
f7d2182e 7553 jnz 0xf7d21883
f7d21830 DD4360 fld qword [ebx+0x60]
f7d21833 D9FA fsqrt <-------------
f7d21835 C7435803000000 mov dword [ebx+0x58],0x3
f7d2183c DD5B50 fstp qword [ebx+0x50]
[... end of loop after this ...]
LJ2:
---- TRACE 0 start test.lua:1
0006 GGET 5 0 ; "math"
0007 TGETS 5 5 1 ; "sqrt"
0008 MOV 6 0
0009 CALL 5 2 2 ; math.sqrt
0010 MOV 0 5
0011 FORL 1 => 0006
---- TRACE 0 IR
0001 > int SLOAD #2 notc
0002 > fun SLOAD #0 notc
0003 > fun FRAME 0002 test.lua:1
0004 tab FLOAD test.lua:1 func.env
0005 int FLOAD 0004 tab.hmask
0006 > int EQ 0005 +63
0007 int FLOAD 0004 tab.node
0008 > int HREFK 0007 "math" [slot 44]
0009 > tab HLOAD 0008
0010 int FLOAD 0009 tab.hmask
0011 > int EQ 0010 +31
0012 int FLOAD 0009 tab.node
0013 > int HREFK 0012 "sqrt" [slot 21]
0014 > fun HLOAD 0013
0015 > num SLOAD #1
0016 > fun FRAME 0014 math.sqrt
0017 + num FPMATH 0015 sqrt
0018 + int ADD 0001 +1
0019 > int LE 0018 +100
0020 ------ LOOP -----------
0021 + num FPMATH 0017 sqrt
0022 + int ADD 0018 +1
0023 > int LE 0022 +100
0024 int PHI 0018 0022
0025 num PHI 0017 0021
---- TRACE 0 mcode 241
f7f58f0f C705580107080000. mov dword [0x08070158], 0x0 // only for profiler
f7f58f19 C744240C00000000 mov dword [esp+0xc], 0x0
f7f58f21 89542410 mov [esp+0x10], edx // BASE of Lua frame
f7f58f25 F20F107A08 movsd xmm7, [edx+0x8] // narrowing of i
f7f58f2a F20F2CF7 cvttsd2si esi, xmm7
f7f58f2e F20F2AF6 cvtsi2sd xmm6, esi
f7f58f32 660F2EFE ucomisd xmm7, xmm6
f7f58f36 0F85484E0F10 jnz ->EXIT_2
f7f58f3c 0F8A424E0F10 jpe ->EXIT_2
f7f58f42 817AF800FCD7F7 cmp dword [edx-0x8], 0xf7d7fc00
f7f58f49 0F85354E0F10 jnz ->EXIT_2
f7f58f4f 8B0508FCD7F7 mov eax, [0xf7d7fc08] // globals
f7f58f55 83781C3F cmp dword [eax+0x1c], +0x3f // lookup "math"
f7f58f59 0F85254E0F10 jnz ->EXIT_2
f7f58f5f 8B4014 mov eax, [eax+0x14]
f7f58f62 83B82C040000FB cmp dword [eax+0x42c], -0x05
f7f58f69 750A jnz 0xf7f58f75
f7f58f6b 81B8280400006827. cmp dword [eax+0x428], 0xf7db2768
f7f58f75 0F85094E0F10 jnz ->EXIT_2
f7f58f7b 83B824040000F5 cmp dword [eax+0x424], -0x0b
f7f58f82 0F85FC4D0F10 jnz ->EXIT_2
f7f58f88 8B8020040000 mov eax, [eax+0x420]
f7f58f8e 83781C1F cmp dword [eax+0x1c], +0x1f // lookup "sqrt"
f7f58f92 0F85EC4D0F10 jnz ->EXIT_2
f7f58f98 8B4014 mov eax, [eax+0x14]
f7f58f9b 83B804020000FB cmp dword [eax+0x204], -0x05
f7f58fa2 750A jnz 0xf7f58fae
f7f58fa4 81B800020000D02B. cmp dword [eax+0x200], 0xf7db2bd0
f7f58fae 0F85D04D0F10 jnz ->EXIT_2
f7f58fb4 83B8FC010000F7 cmp dword [eax+0x1fc], -0x09
f7f58fbb 0F85C34D0F10 jnz ->EXIT_2
f7f58fc1 837A04F3 cmp dword [edx+0x4], -0x0d // type check of x
f7f58fc5 0F87B94D0F10 ja ->EXIT_2
f7f58fcb 81B8F8010000B02B. cmp dword [eax+0x1f8], 0xf7db2bb0 // check func ID
f7f58fd5 0F85A94D0F10 jnz ->EXIT_2
f7f58fdb F20F513A sqrtsd xmm7, [edx] // sqrt() once in pre-roll
f7f58fdf 83C601 add esi, +0x01
f7f58fe2 83FE64 cmp esi, +0x64
f7f58fe5 0F8F9D4D0F10 jg ->EXIT_1
->LOOP:
f7f58feb F20F51FF sqrtsd xmm7, xmm7 // sqrt() in loop
f7f58fef 83C601 add esi, +0x01
f7f58ff2 83FE64 cmp esi, +0x64
f7f58ff5 0F8EF0FFFFFF jle ->LOOP
f7f58ffb E9904D0F10 jmp ->EXIT_0
---- TRACE 0 end
--Mike
