[Bug c++/2892] array addresses corrupted when using template classes

[Bug c++/2892] array addresses corrupted when using template classes

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-02-16 05:12 -------
Note the tree-ssa the simple testcase this code:
pr2892.cc: In function `int main()':
pr2892.cc:46: error: conversion from `int*' to non-scalar type `int[2]' requested
pr2892.cc:50: error: conversion from `int*' to non-scalar type `int[2]' requested


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |rejects-valid
   Last reconfirmed|2003-11-16 00:25:41         |2004-02-16 05:12:48
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] array addresses corrupted when using template classes

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-04-28 01:02 -------
On the mainline we get a warning:
pr2892.cc: In member function `void* X<T, F>::Foo(F) [with T = int[2], F = int[2]]':
pr2892.cc:46:   instantiated from here
pr2892.cc:8: warning: address of local variable `arg' returned

So I think this is invalid.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|rejects-valid               |
   Last reconfirmed|2004-02-16 05:12:48         |2004-04-28 01:02:53
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] array addresses corrupted when using template classes

[bangerth at dealii dot org]

------- Additional Comments From bangerth at dealii dot org  2004-04-28 06:48 -------
That is actually a pretty good point -- the function that fails is the 
only one where the type of the argument is not deducted but explicitly 
specified as a template argument, namely int[2]. It would seem to be 
that it is indeed passed by value, which is what the compiler warns about. 
Running the program subsequently fails. 
 
In all the other cases, the deduced type for the argument should be a 
reference, and taking the address should and does yield the correct result. 
Likewise, if I change the argument of the failing function to take a  
reference F&, then the program succeeds. 
 
Nathan, what do you think? Was this just as oversight, or am I missing 
something? I believe that Andrew is actually right here. 
 
W. 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] array addresses corrupted when using template classes

[nathan at gcc dot gnu dot org]

------- Additional Comments From nathan at gcc dot gnu dot org  2004-04-28 07:19 -------
I was unable to get 3.4 to give the warning that Andrew showed.  I agree with
Wolfgang, that it looks like the int[2] parameter of X::Foo is being
passed by value, rather than pointer -- which is what I said in attachment 6.



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] array addresses corrupted when using template classes

[bangerth at dealii dot org]

------- Additional Comments From bangerth at dealii dot org  2004-04-28 18:57 -------
Alright: in Nathan's attachment to this PR, if I comment test 3 out, we 
fail in test 4, so these are equivalent. 
 
The question thus boils down to: in this snippet 
--------------- 
extern "C" void abort(); 
 
template <typename T>  
void *Foo (T arg) { return &arg[0]; } 
 
int main () { 
  int bry[2]; 
  if (Foo<int[2]>(bry) != bry) 
    abort(); 
} 
------------------ 
gcc aborts because it passes the arg to Foo by value, rather than by 
reference, and the returned value is to a temporary. gcc 3.5 even 
warns about this, as has been noted before. 
 
On the other hand, icc accepts the code. The question then is: is gcc 
wrong to pass by value? I believe this is sensible, but icc seems to 
disagree. Nathan, you seem to have a much better understanding of this: 
why do you believe we are wrong here? 
 
W. 
 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] array addresses corrupted when using template classes

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-04-29 00:25 -------
(In reply to comment #12)

> The question then is: is gcc wrong to pass by value? 

Of course. After instantiation, the function signature is "void* Foo(int arg
[2])". There is nothing like "passing an array by value", and the automatic 
array-to-pointer decay should apply (so that the semantic of the call 
matches "void* Foo(int* arg)").

Notice that the instantion is still "T=int[2]", and thus different 
from "T=int*", but the call semantic would be same.

This is what Nathan says in comment #6: we don't do this decay for template 
functions, and this is our bug.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] No array-to-pointer decay happens for template functions

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  BugsThisDependsOn|                            |20208


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] No array-to-pointer decay happens for template functions

[mmitchel at gcc dot gnu dot org]

-- 
Bug 2892 depends on bug 20208, which changed state.

Bug 20208 Summary: [4.0/4.1 Regression] No array-to-pointer decay happens for template functions
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20208

           What    |Old Value                   |New Value
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
             Status|NEW                         |ASSIGNED
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892

[Bug c++/2892] No array-to-pointer decay happens for template functions

[reichelt at gcc dot gnu dot org]

------- Additional Comments From reichelt at gcc dot gnu dot org  2005-03-16 13:16 -------
Fixed in gcc 4.0.0 - probably by Mark's patch for PR 20208
which also includes a suitable testcase.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2892