www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - comparing pointers passed to and returned from funcs

reply spir <denis.spir gmail.com> writes:
Hello,

It seems to be the kind of stupid issue that will make you laugh about me. But 
I cannot grasp and want to move forward anyway; so, let us be bold and take the 
risk ;-)

I'm modeling a little dynamic language. Elements (values, objects) are pointers 
to structs (actually tagged unions) allocated on the heap. I have a problem in 
passing and returning those pointers to and from primitive procedures.

Precisely, unlike in D, Logical (boolean) operations only accept Logical 
elements true/false (called TRUE/FALSE on implementation side for obvious
reason):
     enum TRUE = new DElement(true);
     enum FALSE = new DElement(false);

  So, I thought I could write those operations by comparing pointers directly,
eg:
     Element not (Element operand) {
         // checks type
         operand.checkLogical()
         return (operand == TRUE) ? FALSE : TRUE;
     }
     ...
     assert ( not(TRUE) == FALSE );

This fails! It even fails "doubly"...
When I call not(TRUE), TRUE (the pointer) inside the func is not equal to the 
global constant. Thus, not always returns FALSE. And in fact this is also 
wrong, because I have the same problem on the return value as well: the FALSE 
returned is not equal to the global FALSE.

But the pointed structs are ok. Each one holds a D boolean of the correct value 
(a member called 'logical'). Thus, the following succeeds:

     Element not (Element operand) {
         // checks type & returns the 'logical' member
         auto log = operand.checkLogical();
         return (log) ? FALSE : TRUE;
     }
     ...
     assert ( not(TRUE).logical == false );

Here, I operate on the structs instead of on the pointers, both to perform the 
operation and in the assert. What I understand is: all happens like if D would 
copy the pointed structs on parameter passing and on return. I thought D would 
only copy the pointers (in both directions), which would let me compare said 
pointers directly.
What do I miss?

It is not a serious problem since the workaround is easy and not very costly. 
But I wish to understand why I cannot operate on constant 'identity'. As said 
above, this must a trivial issue I simply cannot guess...

Thank you,
Denis
-- 
_________________
vita es estrany
spir.wikidot.com
Mar 01 2011
next sibling parent reply bearophile <bearophileHUGS lycos.com> writes:
spir:


 It seems to be the kind of stupid issue that will make you laugh about me. But 
 I cannot grasp and want to move forward anyway; so, let us be bold and take
the 
 risk ;-)



Be bold. Understanding things is much more important. I've being wrong hundreds
of times on D newsgroups, but I'm getting better.


 This fails! It even fails "doubly"...



Reduced program (and Walter thinks that avoiding printf-related bugs is not
important. He's wrong. I use printf often enough in D):


import core.stdc.stdio: printf;
struct Foo {
    bool b;
    this(bool b_) { this.b = b_; }
}
enum Foo* TRUE = new Foo(true);
enum Foo* FALSE = new Foo(false);
Foo* not(Foo* op) {
    return (op == TRUE) ? FALSE : TRUE;
}
void main() {
    //assert(not(TRUE) == FALSE);
    printf("%x\n", TRUE);
    printf("%x\n", FALSE);
    printf("%x\n", not(TRUE));
    printf("%x\n", not(not(TRUE)));
}



 Here, I operate on the structs instead of on the pointers, both to perform the 
 operation and in the assert. What I understand is: all happens like if D would 
 copy the pointed structs on parameter passing and on return. I thought D would 
 only copy the pointers (in both directions), which would let me compare said 
 pointers directly.
 What do I miss?



DMD is not copying the structs, just pointers.

I think it's happening something like with enum associative arrays. Pointers
are run-time things. You are asking for a new at compile-time. This is the asm
produced by that program (-O -release):


_D4test3Foo6__ctorMFNcbZS4test3Foo	comdat
		push	EAX
		mov	CL,8[ESP]
		mov	[EAX],CL
		pop	ECX
		ret	4

_D4test3notFPS4test3FooZPS4test3Foo	comdat
L0:		push	EAX
		mov	ECX,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	EAX
		push	EBX
		push	1
		push	ECX
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		push	1
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		cmp	EAX,8[ESP]
		jne	L3D
		mov	EDX,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	1
		push	EDX
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		push	0
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		jmp short	L56
L3D:		mov	EBX,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	1
		push	EBX
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		push	1
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
L56:		pop	EBX
		add	ESP,8
		ret

__Dmain	comdat
L0:		push	EAX
		mov	EAX,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	EBX
		push	ESI
		push	1
		push	1
		push	EAX
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		mov	ECX,offset FLAT:_DATA
		push	EAX
		push	ECX
		call	near ptr _printf
		mov	EDX,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	0
		push	1
		push	EDX
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		mov	EBX,offset FLAT:_DATA
		push	EAX
		push	EBX
		call	near ptr _printf
		mov	ESI,offset FLAT:_D20TypeInfo_PS4test3Foo6__initZ
		push	1
		push	1
		push	ESI
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		call	near ptr _D4test3notFPS4test3FooZPS4test3Foo
		push	EAX
		push	EBX
		call	near ptr _printf
		push	1
		push	1
		push	ESI
		call	near ptr __d_newarrayT
		add	ESP,8
		mov	EAX,EDX
		call	near ptr _D4test3Foo6__ctorMFNcbZS4test3Foo
		call	near ptr _D4test3notFPS4test3FooZPS4test3Foo
		call	near ptr _D4test3notFPS4test3FooZPS4test3Foo
		push	EAX
		push	EBX
		call	near ptr _printf
		add	ESP,020h
		xor	EAX,EAX
		pop	ESI
		pop	EBX
		pop	ECX
		ret


My suspect is confirmed, DMD is acting as with enum associative arrays. This is
a compiler bug. DMD has to act correctly and not create a struct every time, or
it has to give a compile-time error. Not this mess.

Bye,
bearophile
Mar 01 2011
next sibling parent reply bearophile <bearophileHUGS lycos.com> writes:
This seems to work:

import core.stdc.stdio: printf;
struct Foo {
    bool b;
    this(bool b_) { this.b = b_; }
}
const(Foo)* TRUE, FALSE;
static this() {
    TRUE = new const(Foo)(true);
    FALSE = new const(Foo)(false);
}
const(Foo)* not(const(Foo)* op) {
    return (op == TRUE) ? FALSE : TRUE;
}
void main() {
    assert(not(TRUE) == FALSE);
    printf("%x\n", TRUE);
    printf("%x\n", FALSE);
    printf("%x\n", not(TRUE));
    printf("%x\n", not(not(TRUE)));
}

Bye,
bearophile
Mar 01 2011
next sibling parent bearophile <bearophileHUGS lycos.com> writes:
http://d.puremagic.com/issues/show_bug.cgi?id=5678
Mar 01 2011
prev sibling parent reply Bekenn <leaveme alone.com> writes:
On 3/1/11 3:00 PM, bearophile wrote:

 const(Foo)* TRUE, FALSE;



I'd remove those parens; you don't want people modifying TRUE or FALSE.
Mar 01 2011
parent reply bearophile <bearophileHUGS lycos.com> writes:
Bekenn:


 I'd remove those parens; you don't want people modifying TRUE or FALSE.



Please, show me working code that implements your idea :-)

Bye,
bearophile
Mar 01 2011
next sibling parent reply Bekenn <leaveme alone.com> writes:
On 3/1/11 4:12 PM, bearophile wrote:

 Bekenn:


 I'd remove those parens; you don't want people modifying TRUE or FALSE.



 Please, show me working code that implements your idea :-)



Touche.  I'll have to test that out once I get back from work...
Mar 01 2011
parent reply bearophile <bearophileHUGS lycos.com> writes:
Bekenn:


 Touche.  I'll have to test that out once I get back from work...



Sorry for that answer of mine, we are here to learn and cooperate, not to fight
:-) It's just I sometimes have problems with const things in D, and sometimes
DMD has problems with const things.

Bye,
bearophile
Mar 01 2011
parent Bekenn <leaveme alone.com> writes:
On 3/1/11 5:31 PM, bearophile wrote:

 Bekenn:


 Touche.  I'll have to test that out once I get back from work...



 Sorry for that answer of mine, we are here to learn and cooperate, not to
fight :-) It's just I sometimes have problems with const things in D, and
sometimes DMD has problems with const things.



Hey, no problem; I thought it was funny...
Mar 01 2011
prev sibling parent Bekenn <leaveme alone.com> writes:
On 3/1/2011 4:12 PM, bearophile wrote:

 Bekenn:


 I'd remove those parens; you don't want people modifying TRUE or FALSE.



 Please, show me working code that implements your idea :-)

 Bye,
 bearophile



Here you go; I only changed the one line.  Compiles and works just fine 
in dmd 2.051.

	const Foo* TRUE, FALSE;
instead of
	const(Foo)* TRUE, FALSE;
Mar 01 2011
prev sibling parent spir <denis.spir gmail.com> writes:
On 03/02/2011 02:24 PM, Steven Schveighoffer wrote:

 On Tue, 01 Mar 2011 18:11:00 -0500, bearophile <bearophileHUGS lycos.com>
wrote:


 http://d.puremagic.com/issues/show_bug.cgi?id=5678



 I think there is a general bug where any time the compiler uses an enum, it
 simply replaces the expression declared for the enum.

 So basically

 enum TRUE = new DElement(true);

 void main()
 {
 auto delem1 = TRUE;
 auto delem2 = TRUE;
 assert(delem1 is delem2); // fails
 }

 gets morphed into this:

 void main()
 {
 auto delem1 = new Delement(true);
 auto delem2 = new Delement(true);
 assert(delem1 is delem2); // fails
 }

 Obviously this works great when the enum is a value type or a string literal
 (which is created at compile time). However, it is not so great for things like
 AAs, array literals, objects, or structs.

 I think there are a few of these bugs in bugzilla, and there should be at least
 a tracker, and if not, they should all be combined. This is a serious problem
 in D, and really creates havoc (both performance-wise and semantically). I
 don't anticipate there is an easy fix.

 Essentially, I'd say enum is completely useless except for builtin types and
 strings.



Thank you Steven & Bearophile. This solves my problem.
I first did not get Bearophile's answer about run/compile-time constants (I 
mean enums). I thought the time when they are created is irrelevant, isn't it?
Anyway, placing the constant defs inside a module's <static this () {...}> 
block does what I mean. It does, in fact, ensure *unicity*. But I don't really 
understand the relation with Steven's explanation above: why/how does the fact 
that a constant's def is inside static this() prevent the compiler to rewrite 
it like shown above? Also, I basically don't understand why dmd does that 
anyway: it's obviously un-ecological ;-)

Denis
-- 
_________________
vita es estrany
spir.wikidot.com
Mar 02 2011
prev sibling parent "Steven Schveighoffer" <schveiguy yahoo.com> writes:
On Tue, 01 Mar 2011 18:11:00 -0500, bearophile <bearophileHUGS lycos.com>  
wrote:


 http://d.puremagic.com/issues/show_bug.cgi?id=5678



I think there is a general bug where any time the compiler uses an enum,  
it simply replaces the expression declared for the enum.

So basically

enum TRUE = new DElement(true);

void main()
{
    auto delem1 = TRUE;
    auto delem2 = TRUE;
    assert(delem1 is delem2); // fails
}

gets morphed into this:

void main()
{
    auto delem1 = new Delement(true);
    auto delem2 = new Delement(true);
    assert(delem1 is delem2); // fails
}

Obviously this works great when the enum is a value type or a string  
literal (which is created at compile time).  However, it is not so great  
for things like AAs, array literals, objects, or structs.

I think there are a few of these bugs in bugzilla, and there should be at  
least a tracker, and if not, they should all be combined.  This is a  
serious problem in D, and really creates havoc (both performance-wise and  
semantically).  I don't anticipate there is an easy fix.

Essentially, I'd say enum is completely useless except for builtin types  
and strings.

-Steve
Mar 02 2011