596458 - Simple loops can force unnecessary spills

Status: RESOLVED WONTFIX

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

Description

[Steven Johnson]

Given a simple function like this in AS3:

function writeloop():int {
    var v:int = 2;
    for ( var i:uint=0; i < 100000 ; i++ )
        global_var = v;
    return i;
}

We are generating LIR such that the loop index variable is spilled and reloaded every time through the loop:

  8:jump 22
  12:label
  13:getlocal1
  14:getglobalscope
  15:swap
  16:setslot 4
  18:getlocal2
  19:increment
  20:convert_u
  21:setlocal2
  22:getlocal2
  23:pushint 100000
  25:iflt 12

This will turn into x86 code something like

0x5f2c62:	mov    -0x58(%ebp),%esi
0x5f2c65:	mov    -0x28(%ebp),%ebx
0x5f2c68:	mov    %ebx,0x10(%esi)
0x5f2c6b:	mov    -0x20(%ebp),%ebx
0x5f2c6e:	lea    0x1(%ebx),%ebx
0x5f2c71:	mov    %ebx,-0x20(%ebp)
0x5f2c74:	mov    -0x20(%ebp),%ebx
0x5f2c77:	mov    %ebx,-0x60(%ebp)
0x5f2c7a:	cmp    $0x186a0,%ebx
0x5f2c80:	jb     0x5f2c62

Note the spill and reload of ebx, due to the 8->22 jump at loop start; in this loop, it's totally unnecessary as nothing else needs to touch the register. 

Ideally, the register allocator in nanojit could be smart enough to detect this and do the right thing. Alternately, perhaps we could restructure the way we emit LIR for simple loops so as to minimize this possibility.