https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116977
Bug ID: 116977
Summary: Analyzer: track OpenACC "host" vs. "device" pointers
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: openacc
Severity: normal
Priority: P3
Component: analyzer
Assignee: vedanttewari2000 at gmail dot com
Reporter: tschwinge at gcc dot gnu.org
CC: dmalcolm at gcc dot gnu.org
Target Milestone: ---
Vedant is working on adding to the GCC Static Analyzer checks for programs
using OpenACC, mentored by me. This idea goes back to my 2019 email
<https://inbox.sourceware.org/yxfpy2wfv9hq....@hertz.schwinge.homeip.net> "Re:
[PATCH 00/49] RFC: Add a static analysis framework to GCC", see "OpenACC in
there.
As a first step, we would like advice on the general implementation approach
for tracking OpenACC "host" vs. "device" pointers. For example, we'd like
'-fanalyzer' to diagnose:
#include <stdlib.h>
// Via '#include <openacc.h>':
#include <stddef.h>
# define __GOACC_NOTHROW __attribute__ ((__nothrow__))
void *acc_malloc (size_t) __GOACC_NOTHROW;
int main()
{
float *a = acc_malloc(sizeof *a);
free(a);
// { dg-error {'a' is a device pointer} TODO { xfail *-*-* } .-1 }
return 0;
}
Currently, this doesn’t emit any diagnostic (thus, "TODO", XFAILed); the goal
is to get a diagnostic for such code, as indicated (or similar).
This code will execute fine in a standard GCC configuration without offloading
enabled (because 'acc_malloc' in that case just does 'return malloc(n);'), but
running this code in a GCC configuration with offloading enabled and a GPU
available, you get:
$ ./a.out
Segmentation fault (core dumped)
('acc_free' needs to be used with 'acc_malloc'.)
Relevant OpenACC specification sections:
- OpenACC 2.6, 3.2. "Runtime Library Routines" and specifically 3.2.18.
"acc_malloc"
We're of course happy to supply any further information.
