Re: (too much) allocated memory
From: TenThumbs (tenthumbs_at_cybernex.net)
Date: 05/28/04
- Next message: Scott Moore: "Re: Interesting article by Randall Hyde"
- Previous message: Robert Redelmeier: "Re: About ELF format"
- In reply to: Tim Roberts: "Re: (too much) allocated memory"
- Next in thread: Grumble: "Re: (too much) allocated memory"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Fri, 28 May 2004 17:50:48 +0000 (UTC)
Tim Roberts wrote:
> [omitted]
>
> Because the behavior is seemingly non-deterministic. I can't trust a
> compiler if I can't predict how it will translate a given piece of code.
I see what upsets you but I think I can explain part of it. Consider this
code:
extern int /* __attribute__ ((regparm(3))) */
bar (char *, char *, char *);
int foo (char *s)
{
char f1[8];
char f2[8];
char f3[8];
f1[0] = s[0];
f2[0] = s[1];
f3[0] = s[2];
return bar (f1, f2, f3);
}
int goo (char *s)
{
char f4[8];
char f5[16];
f4[0] = s[0];
f5[0] = s[1];
return bar (f4, f5, (char *)0);
}
You have to use the variables or gcc will optimize them away. Using
gcc-3.4.0 -S -fverbose-asm -O2
I get
.text
.p2align 4,,15
..globl foo
.type foo, @function
foo:
pushl %ebp #
movl %esp, %ebp #,
subl $40, %esp #,
movl 8(%ebp), %edx # s, s
movzbl (%edx), %eax #* s, tmp66
movb %al, -24(%ebp) # tmp66, f1
movzbl 1(%edx), %eax #, tmp67
movb %al, -16(%ebp) # tmp67, f2
movzbl 2(%edx), %eax #, tmp68
movb %al, -8(%ebp) # tmp68, f3
leal -8(%ebp), %eax #, tmp70
movl %eax, 8(%esp) # tmp70,
leal -16(%ebp), %eax #, tmp71
movl %eax, 4(%esp) # tmp71,
leal -24(%ebp), %eax #, tmp72
movl %eax, (%esp) # tmp72,
call bar #
leave
ret
.size foo, .-foo
.p2align 4,,15
..globl goo
.type goo, @function
goo:
pushl %ebp #
movl %esp, %ebp #,
subl $56, %esp #,
movl 8(%ebp), %edx # s, s
movzbl (%edx), %eax #* s, tmp62
movb %al, -32(%ebp) # tmp62, f4
movzbl 1(%edx), %eax #, tmp63
movb %al, -24(%ebp) # tmp63, f5
xorl %eax, %eax #
movl %eax, 8(%esp) #,
leal -24(%ebp), %eax #, tmp65
movl %eax, 4(%esp) # tmp65,
leal -32(%ebp), %eax #, tmp66
movl %eax, (%esp) # tmp66,
call bar #
leave
ret
.size goo, .-goo
In the first function {f1, f2, f3} = ebp - {24,16,8} but in the second
{f4, f5} = ebp - {32,24}. Putting the arrays 8 bytes lower triggers the
additional 16 byte alignment. Using the -d? options I see that something
happens during global register allocation but I have no idea exactly what.
Could be a bug or a feature. You'd have to ask the gcc folks. You can also use
-Os.
--
Mercury
2004-05-28 08:53:32.057 UTC (JD 2453153.870510)
X = 0.349565508, Y = -0.145894088, Z = -0.113696023
X' = 0.007336451, Y' = 0.023564389, Z' = 0.011826050
- Next message: Scott Moore: "Re: Interesting article by Randall Hyde"
- Previous message: Robert Redelmeier: "Re: About ELF format"
- In reply to: Tim Roberts: "Re: (too much) allocated memory"
- Next in thread: Grumble: "Re: (too much) allocated memory"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
|
