#1016 vvp incorrectly exits on termination of a begin/end

[Evan Lavelle]

git, v10-branch

The attached zip is compiler-generated V95 output, apart from the indenting and the comments. The expected output is:

Incr/decr operators...      
line 91, CHECK_RES
line 91-1, CHECK_RES
line 91-2B, CHECK_RES
line 91-B, cleared _$ID111
line 91-C, done
line 91: completed

However, vvp/a.out exits in between the final two $write calls (in other words, 'line 91: completed' does not appear).

[Evan Lavelle]

Note that this worked in earlier versions of iverilog.

[Cary R.]

I can confirm V10 produces different results then devel which produced the following:

Incr/decr operators...
line 91, CHECK_RES
line 91-1, CHECK_RES
line 91-2B, CHECK_RES
line 91-B, cleared $ID111
line 91-C, done
line 91: completed
line 92, CHECK_RES
line 93, CHECK_RES
line 94, CHECK_RES
line 100, arithmetic
line 103, CHECK_RES
in $L59...
incr passcount...
done
test complete: 30 passes, 0 fails

Which is more than you indicated above.

[Martin Whitaker]

The commit that results in this passing in devel is 2fedb4942ec7537b26a4d9ac26e4e8606fda019b. And indeed, applying that change to v10 appears to fix the problem. But I suspect that is just masking the real problem, which is likely a buffer overrun or rogue pointer. I can also get the test to pass in v10 by:

--- a/vvp/vpi_scope.cc
+++ b/vvp/vpi_scope.cc
@@ -428,7 +428,7 @@ void vpip_attach_to_scope(struct __vpiScope*scope, vpiHandle obj)

       if (scope->intern == 0)
            scope->intern = (vpiHandle*)
-                 malloc(sizeof(vpiHandle));
+                 malloc(10*sizeof(vpiHandle));
       else
            scope->intern = (vpiHandle*)
                  realloc(scope->intern, sizeof(vpiHandle)*scope->nintern);

but fail again if I increase the initial allocation to 100*sizeof(vpiHandle). I can't spot any problem with that section of code.

I guess it's time to run with valgrind...

[Martin Whitaker]

Forgot to say: if I comment out the two disable statements in task _$ID32, that also causes the test to pass.

[Martin Whitaker]

This should now be fixed in both the master and v10 branches. Please reopen this bug if you find otherwise.