Fix issue #670: Internal compiler error #919
Conversation
The issue as reported is about the use of BOZ constants, but the internal compiler error was generated by the handling of the array constructor in the parameter definition. This fix from the NVIDIA compiler addresses the issue. Added two tests as well, from the original issue, and a cut-down version used internally to identify the fix.
| @@ -13286,7 +13286,7 @@ fixup_param_vars(SST *var, SST *init) | |||
| sptr1 = get_param_alias_var(sptr, dtype); | |||
| } else if (DTY(dtype) == TY_ARRAY) { | |||
| ad = AD_DPTR(dtype); | |||
| if (AD_ASSUMSZ(ad) || AD_ADJARR(ad) || AD_DEFER(ad)) { | |||
| if (AD_ADJARR(ad) || AD_DEFER(ad)) { | |||
There was a problem hiding this comment.
Flang would have given an error here for the following code previously, with this change the error does not appear.
program test
integer, parameter :: asumed_sz(*) = 1
print *, assumed_sz(1)
end program
There was a problem hiding this comment.
If the initializer in your example was changed to (/1/), the shape should be implied to be DIMENSION(1), but Flang currently would still complain incorrectly. gfortran can compile that case successfully.
Setting an assumed-shape array to a scalar integer has unclear semantics, and Flang should complain about this case.
It seems to me that the code should differentiate these cases and produce errors (or not) accordingly.
|
There's some kind of interaction with this fix in Flang that doesn't appear in the NVIDIA compiler. I need to dig further to see if there are other changes that are also required to me merged. (I didn't see anything when I first looked, but something here can't be explained.) |
|
I have checked with the latest master branch, and the problem is no longer occurring, thanks to a fix that landed as 97527ab. Both test cases in this PR run correctly, and the compiler behaves correctly in the case mentioned by Kiran. |
|
Closing stale PR. |
The issue as reported is about the use of BOZ constants,
but the internal compiler error was generated by the handling
of the array constructor in the parameter definition.
This fix from the NVIDIA compiler addresses the issue.
Added two tests as well, from the original issue, and
a cut-down version used internally to identify the fix.