Native compilation - specific optimisation surely possible?

[Alan Mackenzie <acm@muc.de>]

Hello, Emacs.

The following very short function:


;; -*- lexical-binding: t -*-
(defun comp-test-55 (x)
  (unless (integerp x)
    x))


byte compiles to:


byte code for comp-test-55:
  doc:   ...
    args: (arg1)
    0       dup
    1       integerp
    2       not
    3       goto-if-nil-else-pop 1
    6       dup
    7:1     return


, then on an amd-64 machine, native compiles to (annotation added by
me):



00000000000012c0 <F636f6d702d746573742d3535_comp_test_55_0>:
Setup of the function:
    12c0:	55                   	push   %rbp
    12c1:	53                   	push   %rbx
    12c2:	48 89 fb             	mov    %rdi,%rbx
    12c5:	48 83 ec 08          	sub    $0x8,%rsp
    12c9:	48 8b 05 18 2d 00 00 	mov    0x2d18(%rip),%rax        # 3fe8 <freloc_link_table@@Base-0x240>
    12d0:	48 8b 28             	mov    (%rax),%rbp
fixnump:
    12d3:	8d 47 fe             	lea    -0x2(%rdi),%eax
    12d6:	a8 03                	test   $0x3,%al
    12d8:	75 26                	jne    1300 <F636f6d702d746573742d3535_comp_test_55_0+0x40>

    12da:	48 8b 05 ff 2c 00 00 	mov    0x2cff(%rip),%rax        # 3fe0 <d_reloc@@Base-0x220>
    12e1:	48 8b 78 10          	mov    0x10(%rax),%rdi
Nil in %rdi?:
    12e5:	31 f6                	xor    %esi,%esi
    12e7:	ff 95 c0 27 00 00    	call   *0x27c0(%rbp)      `eq' <========================
    12ed:	48 85 c0             	test   %rax,%rax
    12f0:	48 0f 45 c3          	cmovne %rbx,%rax
Tear down of the function:
    12f4:	48 83 c4 08          	add    $0x8,%rsp
    12f8:	5b                   	pop    %rbx
    12f9:	5d                   	pop    %rbp
    12fa:	c3                   	ret    
    12fb:	0f 1f 44 00 00       	nopl   0x0(%rax,%rax,1)
bignump:
    1300:	8d 47 fb             	lea    -0x5(%rdi),%eax
    1303:	a8 07                	test   $0x7,%al
    1305:	74 09                	je     1310 <F636f6d702d746573742d3535_comp_test_55_0+0x50>

    1307:	31 ff                	xor    %edi,%edi
    1309:	eb da                	jmp    12e5 <F636f6d702d746573742d3535_comp_test_55_0+0x25>
    130b:	0f 1f 44 00 00       	nopl   0x0(%rax,%rax,1)
pseudovectorp:
    1310:	be 02 00 00 00       	mov    $0x2,%esi
    1315:	ff 55 08             	call   *0x8(%rbp)
    1318:	84 c0                	test   %al,%al
    131a:	75 be                	jne    12da <F636f6d702d746573742d3535_comp_test_55_0+0x1a>
    131c:	31 ff                	xor    %edi,%edi
    131e:	eb c5                	jmp    12e5 <F636f6d702d746573742d3535_comp_test_55_0+0x25>

..  The input parameter x (or arg1) is passed into the function in the
register %rdi.  integerp is coded successively as fixnump followed (if
necessary) by bignump.  The fixnump is coded beautifully in three
instructions.

I don't understand what's happening at 12da.  It seems that the address
of a stack pointer is being loaded into %rax, from which the result of
`fixnump' (which was already in %rax) is loaded into %rdi.  

But my main point is the compilation of the `not' instruction at 12e5.
The operand to `not' is in %rdi.  It is coded up as (eq %rdi nil) by
loading 0 (nil) into %rsi at 12e5, then making a function call to `eq'
at 12e7.

Surely the overhead of the function call for `eq' makes this a candidate
for optimisation?  `not' could be coded up in two instructions (test
%rdi,%rdi followed by a conditional jump or (faster) the cmovne which is
%already there).

`not' is presumably a common opcode in byte compiled functions.  `eq'
surely more so.  So why are we coding these up as function calls?

Andrea?

-- 
Alan Mackenzie (Nuremberg, Germany).