Scott Moore wrote:
{******************************************************************************
GPC standard ISO 7185 test failures
The following failures were found:
=============================================================================== C:\GPC\usr\test>gpc --standard-pascal -Wno-warnings -o fails fails.pas fails.pas:115: error: ISO 7185 Pascal allows only simple constants fails.pas: In main program: fails.pas:136: error: this use of packed array components is an extension of fails.pas:136: error: Borland Pascal, traditional Macintosh Pascal fails.pas:137: error: this use of packed record fields is an extension of Borland fails.pas:137: error: Pascal, traditional Macintosh Pascal ===============================================================================
None of these are valid failures, or indeed even have a reasonable explaination in the standard. Also, all of these used to work fine in GPC:
GNU Pascal version 2.1 (20020510), based on gcc-2.95.3-6 (mingw special)
See the testing details for that compiler:
Yes, they are plain bugs (too eager error checks). The first one (string constants) should be fixed in versions starting from gpc-20060215. The patch below (to gpc-20070904) should fix the two other.
Now, with "--standard-pascal" off:
=============================================================================== C:\GPC\usr\test>gpc -Wno-warnings -o fails fails.pas
C:\GPC\usr\test>fails 1 10 100 2147483647 True False 'too much<eoln> too soon' s/b 'too much<eoln> too soon<eoln> ' ===============================================================================
Items:
- Integer field default appears to be "1".
This appears to indicate the default fields for integer is set to 1. This is technically valid, but clearly preverts the intent of the language. The "Pascal users manual and report" shows several examples of the field for integer being clearly set to the maximum size of integer. Ie., for a maximum integer of 11 digits including the sign (for +/-2147483647) the default should be 11 digits for integer. The output of numbers in collumns under the assumption that they would correctly spaced for all integers would fail under the current GPC (see again the several examples in the "report"). I have had the discussion with another implementor, and this treatment of defaults falls under "techically valid (by the ISO 7185 standard), but would break examples of the original Pascal as formulated by Wirth", which I consider a quality issue.
I tried this also with the "--standard-pascal" flag on (by commenting out the error causing statements), and got the same result.
Again, this used to pass.
Hmm, after commenting out the two offending statements in your program and using "--standard-pascal" flag I get:
1 10 100 9223372036854775807 True False 'too much<eoln> too soon<eoln> ' s/b 'too much<eoln> too soon<eoln> '
The same using the program as-is and patched gpc-20070904. In both cases gpc run on AMD64 machine under Mandrake 9.2 Linux.
- Boolean outputs nonstandard.
There is no explaination for the values of boolean. Since they are treated as the strings "true" or "false", there is no possible default field that would deliver the results shown. If the default were 5 (size of 'false') the result would be:
True False
For 4 (size of 'true') it would be:
True Fals
I tried this also with the "--standard-pascal" flag on (by commenting out the error causing statements), and got the same result.
Again, this used to pass.
- Last line in file without "eoln" not corrected as required by standard.
In ISO 7185 "6.4.3.5 File types" lines without proper eoln are referred to as "partial lines", and are stated not to occur. In effect, the implementation must either make sure that an eoln is always generated at the end of the file, or that on read, an eoln is essentially inserted if missing at the end of the file. In effect, a text file must be completely empty or be structured as a series of whole lines, including an eoln at the end of each line.
I don't believe this ever passed in GPC, but I was told it would be fixed in later revisions.
IIRC the promise was that this will work with "--standard-pascal". I belive that there was no intent to change default behaviour.
Here is a -v dump on the GPC used:
<snip>
gpc version 20051116, based on gcc-3.4.4
I do not have binary gpc-20051116 at hand, but I belive that except for the first error it should behave the same as gpc-20060215 or gpc-20070904 (there were a number of ISO compilance fixes in gpc-20060215, but none should affect other problems).
The patch follows:
--- p/predef.c 4 Sep 2007 06:02:58 -0000 1.20 +++ p/predef.c 22 Nov 2007 23:37:32 -0000 @@ -422,7 +422,10 @@ actual_set_parameters (tree val, int reference) { tree domain = TYPE_DOMAIN (TREE_TYPE (val)), addr; - int addressable = mark_addressable (val); + unsigned long save_pascal_dialect = co->pascal_dialect; + int addressable = mark_addressable2 (val, 1); + co->pascal_dialect = ANY_PASCAL; + gcc_assert (addressable);
/* Callers now handle the constant empty set. */ @@ -435,6 +438,7 @@ addr = build_pascal_unary_op (ADDR_EXPR, val); else addr = build1 (ADDR_EXPR, build_pointer_type (TREE_TYPE (val)), val); + co->pascal_dialect = save_pascal_dialect; return tree_cons (NULL_TREE, addr, tree_cons (NULL_TREE, convert (pascal_integer_type_node, TYPE_MIN_VALUE (domain)), build_tree_list (NULL_TREE, convert (pascal_integer_type_node, TYPE_MAX_VALUE (domain)))));