https://bugs.llvm.org/show_bug.cgi?id=43330
Bug ID: 43330
Summary: OpenMP target: Mapping of partial arrays fails
Product: OpenMP
Version: unspecified
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Runtime Library
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected]
Created attachment 22511
--> https://bugs.llvm.org/attachment.cgi?id=22511&action=edit
Reproducer of the bug
According to the OMP Specs (Chapter 2.1.5) an array section can have the
following format (beside others):
[ lower-bound : length : stride] or
[ lower-bound : ]
In the latter case the length and the stride are determined as follows (see
specs):
"When the stride is absent it defaults to 1.
When the length is absent it defaults to (size - lower-bound)/stride, where
size is the size of the array dimension."
Given a declaration "int a[N];" the following code should be semantically the
same:
(1) #pragma omp target data map(from: a[1:N-1])
(2) #pragma omp target data map(from: a[1:])
However, if I test this with "-fopenmp-targets=x86_64-unknown-linux-gnu" I get
the following debug output from libomptarget (LIBOMPTARGEG_DEBUG=1) for both
cases (and N=1000):
(1) Libomptarget --> Moving 3996 bytes (tgt:0x0000000000623dc0) ->
(hst:0x00007ffe65c4d474)
(2) Libomptarget --> Moving 4000 bytes (tgt:0x0000000000623dc0) ->
(hst:0x00007fff2d5a45f4)
Thus, in case (2) there are too many bytes transfered. This has bad side
effects, because some stack variables on the host may be overwritten (which
actually happens to me).
I attached a reproducer. Can you confirm this bug (or even better fix it ;-)).
I am not sure if this is an issue of the compiler or the runtime.
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
