clamp re-queried counts in terminator_EnumeratePhysicalDeviceGroups

[aizu-m]

Tracing the device-group terminator while chasing a stack-smash report:

loader.c:7829  for (indiv_gpu = 0; indiv_gpu < count_this_time; ++indiv_gpu)
                   local_phys_dev_groups[indiv_gpu + cur_icd_group_count] = ...

local_phys_dev_groups is sized to total_count, which is summed in the first pass that asks each ICD for its group count. The fill pass then re-queries every ICD into count_this_time (EnumeratePhysicalDevices on the no-device-group path, EnumeratePhysicalDeviceGroups otherwise, both with a NULL property pointer) and indexes the array by that value. The re-queried count is never bounded to the space left in the array.

So a driver that reports more devices on the second query than it did on the first walks off the end of the stack array. Both fill branches re-query, so both can overrun. Same count-drift behaviour the *2KHR emulators already guard against, here on the two-pass group path.

Clamp count_this_time to the remaining reserved space (total_count - cur_icd_group_count) right after each re-query. No change for a well-behaved driver where the two passes agree; the existing device-group tests stay green.

[ci-tester-lunarg]

Author aizu-m not on autobuild list. Waiting for curator authorization before starting CI build.

[ci-tester-lunarg]

Author aizu-m not on autobuild list. Waiting for curator authorization before starting CI build.