Re: [GHC] #14753: Assembly for stg_enter_info starts with a dead 'mov'

2 Feb 2018

      #14753: Assembly for stg_enter_info starts with a dead 'mov'
-------------------------------------+-------------------------------------
        Reporter:  alpmestan         |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  low               |            Milestone:
       Component:  Compiler          |              Version:  8.4.1-alpha1
  (CodeGen)                          |
      Resolution:                    |             Keywords:
Operating System:  Linux             |         Architecture:  x86_64
                                     |  (amd64)
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Description changed by alpmestan:

Old description:
...
In the course of my investigation in #14675, I happened to look at the
x86_64 assembly for `stg_enter_info`, through gdb:
{{{#!asm
Breakpoint 2, stg_enter_info () at rts/HeapStackCheck.cmm:166
166     {
(gdb) disassemble
Dump of assembler code for function stg_enter_info:
=> 0x0000000002cf5900 <+0>:     mov    0x0(%rbp),%rax
   0x0000000002cf5904 <+4>:     mov    0x8(%rbp),%rax
   0x0000000002cf5908 <+8>:     test   $0x7,%al
   0x0000000002cf590a <+10>:    jne    0x2cf593d 
   0x0000000002cf590c <+12>:    mov    (%rax),%rbx
   0x0000000002cf590f <+15>:    cmpl   $0x1a,-0x8(%rbx)
   0x0000000002cf5913 <+19>:    jb     0x2cf5947 
   0x0000000002cf5915 <+21>:    cmpl   $0x1c,-0x8(%rbx)
   0x0000000002cf5919 <+25>:    jb     0x2cf592d 
   0x0000000002cf591b <+27>:    cmpl   $0x1d,-0x8(%rbx)
   0x0000000002cf591f <+31>:    jb     0x2cf5933 
   0x0000000002cf5921 <+33>:    mov    %rbx,%rcx
   0x0000000002cf5924 <+36>:    mov    %rax,%rbx
   0x0000000002cf5927 <+39>:    add    $0x10,%rbp
   0x0000000002cf592b <+43>:    jmpq   *%rcx
   0x0000000002cf592d <+45>:    cmpl   $0x1b,-0x8(%rbx)
   0x0000000002cf5931 <+49>:    jb     0x2cf5921 
   0x0000000002cf5933 <+51>:    mov    0x8(%rax),%rax
   0x0000000002cf5937 <+55>:    mov    %rax,0x8(%rbp)
   0x0000000002cf593b <+59>:    jmp    0x2cf5908 
   0x0000000002cf593d <+61>:    mov    %rax,%rbx
   0x0000000002cf5940 <+64>:    add    $0x10,%rbp
   0x0000000002cf5944 <+68>:    jmpq   *0x0(%rbp)
   0x0000000002cf5947 <+71>:    cmpl   $0xf,-0x8(%rbx)
   0x0000000002cf594b <+75>:    jb     0x2cf5969 
   0x0000000002cf594d <+77>:    cmpl   $0x19,-0x8(%rbx)
   0x0000000002cf5951 <+81>:    jb     0x2cf595d 
   0x0000000002cf5953 <+83>:    mov    %rax,%rbx
   0x0000000002cf5956 <+86>:    add    $0x10,%rbp
   0x0000000002cf595a <+90>:    jmpq   *0x0(%rbp)
   0x0000000002cf595d <+93>:    movslq -0x8(%rbx),%rcx
   0x0000000002cf5961 <+97>:    cmp    $0x17,%rcx
   0x0000000002cf5965 <+101>:   jne    0x2cf5921 
   0x0000000002cf5967 <+103>:   jmp    0x2cf5953 
   0x0000000002cf5969 <+105>:   cmpl   $0x8,-0x8(%rbx)
   0x0000000002cf596d <+109>:   jb     0x2cf5921 
   0x0000000002cf596f <+111>:   movslq -0x8(%rbx),%rbx
   0x0000000002cf5973 <+115>:   add    $0xfffffffffffffff8,%rbx
   0x0000000002cf5977 <+119>:   jmpq   *0x2e36d90(,%rbx,8)
}}}
Notice the first two instructions:
{{{#!asm
mov    0x0(%rbp),%rax
mov    0x8(%rbp),%rax
}}}
We successively write two different things to `rax`, making the first
`mov` effectively useless. I'm not quite sure the impact is of any
significance but it still seems worth writing this down and addressing it
at some point in the future, as there's just no way that first
instruction will ever be useful.
I think the assembly is generated from
[https://github.com/ghc/ghc/blob/master/rts/HeapStackCheck.cmm#L164 this
code], which in turns call
[https://github.com/ghc/ghc/blob/master/includes/Cmm.h#L301 `ENTER`],
which in turns calls
[https://github.com/ghc/ghc/blob/master/includes/Cmm.h#L272 `LOAD_INFO`],
which appears to rcorrespond to the first few instructions in the
assembly I pasted, if we ignore the extra `mov`.
(Assembly generated by ghc 8.4.1 built from the same commit as the alpha1
for 8.4.1 from hvr's PPA, in a 64bits Ubuntu VM. Not sure we generate a
dead `mov` with the other codegens, haven't tested.)
New description:

 In the course of my investigation in #14675, I happened to look at the
 x86_64 assembly for `stg_enter_info`, through gdb:

 {{{#!asm
 Breakpoint 2, stg_enter_info () at rts/HeapStackCheck.cmm:166
 166     {
 (gdb) disassemble
 Dump of assembler code for function stg_enter_info:
 => 0x0000000002cf5900 <+0>:     mov    0x0(%rbp),%rax
    0x0000000002cf5904 <+4>:     mov    0x8(%rbp),%rax
    0x0000000002cf5908 <+8>:     test   $0x7,%al
    0x0000000002cf590a <+10>:    jne    0x2cf593d 
    0x0000000002cf590c <+12>:    mov    (%rax),%rbx
    0x0000000002cf590f <+15>:    cmpl   $0x1a,-0x8(%rbx)
    0x0000000002cf5913 <+19>:    jb     0x2cf5947 
    0x0000000002cf5915 <+21>:    cmpl   $0x1c,-0x8(%rbx)
    0x0000000002cf5919 <+25>:    jb     0x2cf592d 
    0x0000000002cf591b <+27>:    cmpl   $0x1d,-0x8(%rbx)
    0x0000000002cf591f <+31>:    jb     0x2cf5933 
    0x0000000002cf5921 <+33>:    mov    %rbx,%rcx
    0x0000000002cf5924 <+36>:    mov    %rax,%rbx
    0x0000000002cf5927 <+39>:    add    $0x10,%rbp
    0x0000000002cf592b <+43>:    jmpq   *%rcx
    0x0000000002cf592d <+45>:    cmpl   $0x1b,-0x8(%rbx)
    0x0000000002cf5931 <+49>:    jb     0x2cf5921 
    0x0000000002cf5933 <+51>:    mov    0x8(%rax),%rax
    0x0000000002cf5937 <+55>:    mov    %rax,0x8(%rbp)
    0x0000000002cf593b <+59>:    jmp    0x2cf5908 
    0x0000000002cf593d <+61>:    mov    %rax,%rbx
    0x0000000002cf5940 <+64>:    add    $0x10,%rbp
    0x0000000002cf5944 <+68>:    jmpq   *0x0(%rbp)
    0x0000000002cf5947 <+71>:    cmpl   $0xf,-0x8(%rbx)
    0x0000000002cf594b <+75>:    jb     0x2cf5969 
    0x0000000002cf594d <+77>:    cmpl   $0x19,-0x8(%rbx)
    0x0000000002cf5951 <+81>:    jb     0x2cf595d 
    0x0000000002cf5953 <+83>:    mov    %rax,%rbx
    0x0000000002cf5956 <+86>:    add    $0x10,%rbp
    0x0000000002cf595a <+90>:    jmpq   *0x0(%rbp)
    0x0000000002cf595d <+93>:    movslq -0x8(%rbx),%rcx
    0x0000000002cf5961 <+97>:    cmp    $0x17,%rcx
    0x0000000002cf5965 <+101>:   jne    0x2cf5921 
    0x0000000002cf5967 <+103>:   jmp    0x2cf5953 
    0x0000000002cf5969 <+105>:   cmpl   $0x8,-0x8(%rbx)
    0x0000000002cf596d <+109>:   jb     0x2cf5921 
    0x0000000002cf596f <+111>:   movslq -0x8(%rbx),%rbx
    0x0000000002cf5973 <+115>:   add    $0xfffffffffffffff8,%rbx
    0x0000000002cf5977 <+119>:   jmpq   *0x2e36d90(,%rbx,8)
 }}}

 Notice the first two instructions:

 {{{#!asm
 mov    0x0(%rbp),%rax
 mov    0x8(%rbp),%rax
 }}}

 We successively write two different things to `rax`, making the first
 `mov` effectively useless. I'm not quite sure the impact is of any
 significance but it still seems worth writing this down and addressing it
 at some point in the future, as there's just no way that first instruction
 will ever be useful.

 I think the assembly is generated from
 [https://github.com/ghc/ghc/blob/master/rts/HeapStackCheck.cmm#L164 this
 code], which in turns call
 [https://github.com/ghc/ghc/blob/master/includes/Cmm.h#L301 ENTER], which
 in turns calls [https://github.com/ghc/ghc/blob/master/includes/Cmm.h#L272
 LOAD_INFO], which appears to rcorrespond to the first few instructions in
 the assembly I pasted, if we ignore the extra `mov`.

 (Assembly generated by ghc 8.4.1 built from the same commit as the alpha1
 for 8.4.1 from hvr's PPA, in a 64bits Ubuntu VM. Not sure we generate a
 dead `mov` with the other codegens, haven't tested.)

--

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14753#comment:1
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler