LuaJit Trace Compiler剖析

Posted 2022-11-30 帝王铠

简要过程如下：

• 分析阶段
  概要分析的目的是识别热循环。循环计数超过某个阈值后，该循环被认为是热循环，并进入跟踪模式。
• 追踪阶段
  在跟踪阶段，循环的执行正常进行，记录的操作通常以中间表示（中间表示（IR）是由编译器或虚拟机内部使用，表示源代码中的数据结构或代码）形式存储。
• 优化和代码生成阶段
  典型的优化包括常量子表达式消除，死代码消除，寄存器分配，不变代码运动，常量折叠和转义分析。优化之后，跟踪将转换为机器代码。

开始剖析：

首先写一个lua的demo程序

local x = 0
for out=1,100 do
  for n=1,200 do
     x = x+1
  end
  print(x)
end

这种固定分支必然会触发热点优化的

$ ./luajit -jv x.lua 
[TRACE   1 x.lua:3 loop]
200
400
600
800
1000
1200
1400
1600
1800
[TRACE --- (1/3) x.lua:6 -- NYI: FFFFFFFFFFFFastFunc print]
2000
[TRACE --- (1/3) x.lua:6 -- NYI: FFFFFFFFFFFFastFunc print]
2200
[TRACE --- (1/3) x.lua:6 -- NYI: FFFFFFFFFFFFastFunc print]
2400
[TRACE --- (1/3) x.lua:6 -- NYI: FFFFFFFFFFFFastFunc print]
2600
[TRACE   2 (1/3) x.lua:6 -- fallback to interpreter]

$ ./luajit -jdump x2.lua
---- TRACE 1 start x2.lua:3
0010  ADDVN    0   0   0  ; 1
0011  FORL     5 => 0010
---- TRACE 1 IR
0001    int SLOAD  #6    CI
0002 >  num SLOAD  #1    T
0003  + num ADD    0002  +1  
0004  + int ADD    0001  +1  
0005 >  int LE     0004  +200
0006 ------ LOOP ------------
0007  + num ADD    0003  +1  
0008  + int ADD    0004  +1  
0009 >  int LE     0008  +200
0010    int PHI    0004  0008
0011    num PHI    0003  0007
---- TRACE 1 mcode 81
2a4fffa3  mov dword [0x4092d4a0], 0x1
2a4fffae  movsd xmm0, [0x4093dae0]
2a4fffb7  cvtsd2si ebp, [rdx+0x28]
2a4fffbc  cmp dword [rdx+0x4], 0xfffeffff
2a4fffc3  jnb 0x2a4f0010        ->0
2a4fffc9  movsd xmm7, [rdx]
2a4fffcd  addsd xmm7, xmm0
2a4fffd1  add ebp, +0x01
2a4fffd4  cmp ebp, 0xc8
2a4fffda  jg 0x2a4f0014 ->1
->LOOP:
2a4fffe0  addsd xmm7, xmm0
2a4fffe4  add ebp, +0x01
2a4fffe7  cmp ebp, 0xc8
2a4fffed  jle 0x2a4fffe0        ->LOOP
2a4fffef  jmp 0x2a4f001c        ->3
---- TRACE 1 stop -> loop

200
400
600
800
1000
1200
1400
1600
1800
---- TRACE 2 start 1/3 x2.lua:6
0012  GGET     5   0      ; "print"
0013  MOV      6   0
0014  CALL     5   1   2
0000  . FUNCC               ; print
---- TRACE 2 abort x2.lua:6 -- NYI: FFFFFFFFFFFFastFunc print

2000
---- TRACE 2 start 1/3 x2.lua:6
0012  GGET     5   0      ; "print"
0013  MOV      6   0
0014  CALL     5   1   2
0000  . FUNCC               ; print
---- TRACE 2 abort x2.lua:6 -- NYI: FFFFFFFFFFFFastFunc print

2200
---- TRACE 2 start 1/3 x2.lua:6
0012  GGET     5   0      ; "print"
0013  MOV      6   0
0014  CALL     5   1   2
0000  . FUNCC               ; print
---- TRACE 2 abort x2.lua:6 -- NYI: FFFFFFFFFFFFastFunc print

通过浏览源码 初步定位
LuaJIT用一张hash表，维护了相关指令（跳转和调用）的热度

  HotCount hotcount[HOTCOUNT_SIZE];	/* Hot counters. */

(gdb) r
Starting program: /home/x/luajit-2.0/src/./luajit -jv x.lua

Temporary breakpoint 5, lj_trace_hot (J=0x40000550, pc=0x4000b498) at lj_trace.c:662
662     
(gdb) bt
#0  lj_trace_hot (J=0x40000550, pc=0x4000b498) at lj_trace.c:662
#1  0x000000000041d5a4 in lj_vm_hotloop ()
#2  0x000000000040b9f0 in lua_pcall (L=L@entry=0x40000378, nargs=nargs@entry=0, nresults=<optimized out>, errfunc=errfunc@entry=2) at lj_api.c:1052
#3  0x000000000040450c in docall (L=0x40000378, narg=0, clear=0) at luajit.c:122
#4  0x000000000040529e in handle_script (n=<optimized out>, argv=<optimized out>, L=<optimized out>) at luajit.c:290
#5  pmain (L=0x40000378) at luajit.c:540
#6  0x000000000041bb86 in lj_BC_FUNCC ()
#7  0x000000000040ba6a in lua_cpcall (L=L@entry=0x40000378, func=func@entry=0x404ac0 <pmain>, ud=ud@entry=0x0) at lj_api.c:1074
#8  0x0000000000404054 in main (argc=3, argv=0x7fffffffdff8) at luajit.c:568

继续深入分析lj_vm_hotloop ()

lj_vm_hotloop:
        .byte 139,106,248,139,109,16,15,182,69,199,141,4,194,139,108,36
        .byte 24,137,85,16,137,69,24,137,222,65,141,190,224,245,255,255
        .byte 73,137,174,64,246,255,255,137,92,36,28,232
        .long lj_trace_hot-.-4
        .byte 235,171

        .globl lj_vm_callhook
        .hidden lj_vm_callhook
        .type lj_vm_callhook, @function
        .size lj_vm_callhook, 6

找他的实现代码，换个方向。

源码中的vm_xxx.dasc文件，本文以vm_x86.dasc为例。

vm_86.dasc由dynasm.lua解析，然后跟buildvm.c一起链接。

分析阶段
源码lj_dispatch.h

/* Type of hot counter. Must match the code in the assembler VM. */
/* 16 bits are sufficient. Only 0.0015% overhead with maximum slot penalty. */
typedef uint16_t HotCount;

/* Number of hot counter hash table entries (must be a power of two). */
#define HOTCOUNT_SIZE		64
#define HOTCOUNT_PCMASK		((HOTCOUNT_SIZE-1)*sizeof(HotCount))

/* Hotcount decrements. */
#define HOTCOUNT_LOOP		2
#define HOTCOUNT_CALL		1


#define hotcount_get(gg, pc) \\
  (gg)->hotcount[(u32ptr(pc)>>2) & (HOTCOUNT_SIZE-1)]
#define hotcount_set(gg, pc, val) \\
  (hotcount_get((gg), (pc)) = (HotCount)(val))

以及lj_jit.h

  _(\\007, hotloop,	56)	/* # of iter. to detect a hot loop/call. */ \\

|// Decrement hashed hotcount and trigger trace recorder if zero.
|.macro hotloop, reg
|  mov reg, PC
|  shr reg, 1
|  and reg, HOTCOUNT_PCMASK
|  sub word [DISPATCH+reg+GG_DISP2HOT], HOTCOUNT_LOOP
|  jb ->vm_hotloop
|.endmacro

翻译器每次执行pcall会调用hotloop和hotcall宏来更新并检查上面的热点计数器。
汇编前面的竖线是直接放在汇编里的一种自定义解析方式。
最终会跳转到vm_hotloop，如下

|->vm_hotloop:			// Hot loop counter underflow.
  |.if JIT
  |  mov LFUNC:RB, [BASE-8]		// Same as curr_topL(L).
  |  mov RB, LFUNC:RB->pc
  |  movzx RD, byte [RB+PC2PROTO(framesize)]
  |  lea RD, [BASE+RD*8]
  |  mov L:RB, SAVE_L
  |  mov L:RB->base, BASE
  |  mov L:RB->top, RD
  |  mov FCARG2, PC
  |  lea FCARG1, [DISPATCH+GG_DISP2J]
  |  mov aword [DISPATCH+DISPATCH_J(L)], L:RBa
  |  mov SAVE_PC, PC
  |  call extern lj_trace_hot@8		// (jit_State *J, const BCIns *pc)
  |  jmp <3
  |.endif

最后调用lj_trace_hot，进入追踪阶段。
综上，探测热点阶段粗略的理解就是根据函数调用次数进行计数，启发式判断热点是否达到阈值，进行追踪

---

追踪阶段
从lj_trace_hot开始，tracing subsystem有五个状态

trace_state
  LJ_TRACE_IDLE,	/* Trace compiler idle. */
  LJ_TRACE_ACTIVE = 0x10,
  LJ_TRACE_RECORD,	/* Bytecode recording active. */
  LJ_TRACE_START,	/* New trace started. */
  LJ_TRACE_END,		/* End of trace. */
  LJ_TRACE_ASM,		/* Assemble trace. */
  LJ_TRACE_ERR		/* Trace aborted with error. */

把lua翻译成特定代码路径的字节码线性序列，字节码流被译作中间码形式，然后再优化和编译成汇编。