Standard compliance vs. real world code

[tmeissner]

Hi @nickg

I updated our docker image from a september version of nvc to the latest one. Now I get analyzing errors with code I use since years without problems (with nvc, ghdl, Questa and synthesis):

** Error: a choice that is not locally static is allowed only if the array aggregate contains a single element association
    > ../../src/CoreMmc/FifoDualPortE.vhd:56
    |
 56 |   constant C_FIFO_EMPTY : unsigned(FIFO_AWIDTH downto 0) := (FIFO_AWIDTH => '1', FIFO_AWIDTH-1 downto 0 => '0');
    |                                                                   ^^^^^^^^^^^

FIFO_AWIDTH is a generic in this context.

This code was compiled by the older version of nvc (and all other tools I use, even synthesis). I can't workaround this error with the relaxed option, so I have to change to code.

The other error I get is with a generic function in a generic package:

nvc --std=19 --work=simulation -L ./ -a ../../../FpgaUtils.repo/Simulation/QueueP.vhd
** Error: no visible subprogram TO_STRING matches signature [G_DATA_TYPE return STRING]
    > ../../../FpgaUtils.repo/Simulation/QueueP.vhd:31
    |
 31 |            function G_TO_STRING(data : G_DATA_TYPE) return string is to_string
    |                                                                      ^^^^^^^^^

This error is more clear to me as nvc obvously can't find a to_string() function for the not yet defined G_DATA_TYPE type. However, this code was also compiled by the older version of nvc (and all other simulation tools I use).

So, would it be possible to add these things to work with the relaxed option? The QueueP package is use since years in a lot of projects. I fear that I have to change a lot of simulation code when I remove the is to_string function predefine.

[nickg]

I updated our docker image from a september version of nvc to the latest one. Now I get analyzing errors with code I use since years without problems (with nvc, ghdl, Questa and synthesis):

** Error: a choice that is not locally static is allowed only if the array aggregate contains a single element association
    > ../../src/CoreMmc/FifoDualPortE.vhd:56
    |
 56 |   constant C_FIFO_EMPTY : unsigned(FIFO_AWIDTH downto 0) := (FIFO_AWIDTH => '1', FIFO_AWIDTH-1 downto 0 => '0');
    |                                                                   ^^^^^^^^^^^

FIFO_AWIDTH is a generic in this context.

This code was compiled by the older version of nvc (and all other tools I use, even synthesis). I can't workaround this error with the relaxed option, so I have to change to code.

Does this really work with GHDL? I tried the following:

entity ent is
    generic (g : integer);
end entity;

architecture test of ent is
    constant empty : bit_vector(g downto 0) := (g => '1', g - 1 downto 0 => '0');
begin
end architecture;

$ ghdl --version
GHDL 5.0.0-dev (4.1.0.r385.g221ee97334a6.dirty) [Dunoon edition]
...
$ ghdl -a -frelaxed test.vhd
test.vhd:6:48:error: not static choice exclude others choice
    constant empty : bit_vector(g downto 0) := (g => '1', g - 1 downto 0 => '0');
                                               ^
test.vhd:6:48:error: non-locally static choice for an aggregate is allowed only if only choice
    constant empty : bit_vector(g downto 0) := (g => '1', g - 1 downto 0 => '0');
                                               ^

Questa also gives an error but looks like it can be suppressed. The problem with non-locally-static choices is that the compiler has to generate code to calculate the bounds of the aggregate at runtime as well as check that every element in that range has been given a value. If the aggregate only contains a single association then this is relatively simple which is why the LRM makes an exception for it. In this case you can use g & (g - 1 downto 0 => '0') instead?

There was a bug I fixed a few weeks ago where generics were treated as locally static in some cases in 2008 mode which is why this was accepted before. But it's just chance that it worked: changing the choice expression slightly would have caused crashes.

The other error I get is with a generic function in a generic package:

nvc --std=19 --work=simulation -L ./ -a ../../../FpgaUtils.repo/Simulation/QueueP.vhd
** Error: no visible subprogram TO_STRING matches signature [G_DATA_TYPE return STRING]
    > ../../../FpgaUtils.repo/Simulation/QueueP.vhd:31
    |
 31 |            function G_TO_STRING(data : G_DATA_TYPE) return string is to_string
    |                                                                      ^^^^^^^^^

This error is more clear to me as nvc obvously can't find a to_string() function for the not yet defined G_DATA_TYPE type. However, this code was also compiled by the older version of nvc (and all other simulation tools I use).

I'm pretty sure this error is correct. See 6.5.4 "Interface subprogram declarations" in the 2008 LRM:

If an interface subprogram declaration contains an interface subprogram default in the form of a subprogram name, the subprogram name shall denote a subprogram, and the denoted subprogram and the interface subprogram shall have conforming profiles (see 4.10).

So the default has to denote a subprogram at the time the interface declaration is analysed, not when it is instantiated. The requirement for conforming profiles is actually stricter than NVC implements at the moment (i.e. it requires the parameter names to be the same, not just the types).

You can make this work by using <> to look up the correct to_string overload at the time the package is instantiated:

    generic (
        type G_DATA_TYPE;
        function to_string (x : G_DATA_TYPE) return string is <>;
        function G_TO_STRING(data : G_DATA_TYPE) return string is to_string );