Re: [avr-gcc-list] Success running the testsuite with avrtest

[Paulo Marques]

Quoting Paulo Marques <address@hidden>:
> [...]
> Although there are 66 tests that FAIL, only 51 are execution 
> failures, and 15 are compilation failures. These 51 failures 
> correspond to just 10 programs that fail to execute, with different 
> optimization options (-O0, -O1, etc.):

I just tried compiling these tests myself to run them one by one on 
avrtest and found out that most of them fail because of 32 bit integer 
size assumptions or other odities:

> execute/20010122-1.c

This test uses __builtin_return_address.

> execute/20060102-1.c
20060102-1.c:5: warning: right shift count >= width of type

> execute/20061101-1.c
20061101-1.c:8: warning: comparison is always true due to limited range 
of data type

> execute/float-floor.c

This test is very sensitive to floating point implementation and 
probably needs an actual 64 bit double type to work properly.

> execute/pr17377.c

This test also uses __builtin_return_address.

> execute/pr22493-1.c

This uses SIMD instructions.

> execute/pr23135.c
pr23135.c:8: warning: integer overflow in expression
pr23135.c:8: warning: integer overflow in expression

> execute/pr28651.c
pr28651.c:11: warning: large integer implicitly truncated to unsigned type

> execute/pr29797-1.c
pr29797-1.c:10: warning: comparison is always true due to limited range 
of data type

> execute/simd-1.c

This one also uses SIMD instructions.

So we're down to the following failing reasons:
- test case assumes 32 bit integers
- test case uses __builtin_return_address
- test case uses SIMD instructions

The __builtin_return_address implementation really has a bug (at least 
with gcc-4.2.2). It works if there is no prologue pushing stuff into 
the stack. Otherwise it can produce code like this:

void NOINLINE *test3 (void)
{
 void * temp;
 temp = __builtin_return_address (0);
 dummy ();
 return temp;
}

000004de <test3>:
4de:    0f 93           push    r16
4e0:    1f 93           push    r17
4e2:    ed b7           in      r30, 0x3d       ; 61
4e4:    fe b7           in      r31, 0x3e       ; 62
4e6:    01 81           ldd     r16, Z+1        ; 0x01
4e8:    12 81           ldd     r17, Z+2        ; 0x02
4ea:    0e 94 30 01     call    0x260   ; 0x260 <dummy>
4ee:    c8 01           movw    r24, r16
4f0:    1f 91           pop     r17
4f2:    0f 91           pop     r16
4f4:    08 95           ret

The "ldd r16, Z+1" and "ldd r17, Z+2" are going to fetch the 2 
registers pushed into the stack instead of the return address.

I guess I need to start testing with the gcc from SVN directly so that 
I don't run into bugs that have already been fixed.

--
Paulo Marques


----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.