[ was: Re: [nvptx, committed, PR81069] Insert diverging jump alap in
nvptx_single ]
On 07/17/2017 10:41 AM, Tom de Vries wrote:
Hi,
Consider nvptx_single:
...
/* Single neutering according to MASK. FROM is the incoming block and
TO is the outgoing block. These may be the same block. Insert at
start of FROM:
if (tid.<axis>) goto end.
and insert before ending branch of TO (if there is such an insn):
end:
<possibly-broadcast-cond>
<branch>
We currently only use differnt FROM and TO when skipping an entire
loop. We could do more if we detected superblocks. */
static void
nvptx_single (unsigned mask, basic_block from, basic_block to)
...
When compiling libgomp.oacc-fortran/nested-function-1.f90 at -O1, we
observed the following pattern:
...
<bb2>:
goto bb3;
<bb3>: (with single predecessor)
<some code>
<cond-branch>
...
which was translated by nvptx_single into:
...
<bb2>
if (tid.<axis>) goto end.
goto bb3;
<bb3>:
<some code>
end:
<broadcast-cond>
<cond-branch>
...
There is no benefit to be gained from doing the goto bb3 in neutered
mode, and there is no need to, so we might as well insert the neutering
branch as late as possible:
...
<bb2>
goto bb3;
<bb3>:
if (tid.<axis>) goto end.
<some code>
end:
<broadcast-cond>
<cond-branch>
...
This patch implements inserting the neutering branch as late as possible.
[ As it happens, the actual code for
libgomp.oacc-fortran/nested-function-1.f90 at -O1 was more complicated:
there were other bbs inbetween bb2 and bb3. While this doesn't change
anything from a control flow graph point of view, it did trigger a bug
in the ptx JIT compiler where it inserts the synchronization point for
the diverging branch later than the immediate post-dominator point at
the end label. Consequently, the condition broadcast was executed in
divergent mode (which is known to give undefined results), resulting in
a hang.
This patch also works around this ptx JIT compiler bug, for this
test-case. ]
Build and tested on x86_64 with nvptx accelerator.
Committed.
Jakub,
I'd like to backport this nvptx patch to the gcc-7-branch.
The patch doesn't trivially fit into the category of regression or
documentation fix.
Without this patch, when building an nvptx offloading compiler and
running the libgomp testsuite for the gcc-7-branch, the GPU hangs, and
I've had a report from a colleague who experienced system crashes
because of it.
However, in principle gcc is not doing anything wrong: the generated
code is according to the ptx spec. It's just that the patch makes it
less likely to run into a ptx JIT bug.
Then again, it's an nvptx patch, neither a primary nor secondary target.
I'll commit the backport some time this week, unless there are objections.
Thanks,
- Tom
0001-Insert-diverging-jump-alap-in-nvptx_single.patch
Insert diverging jump alap in nvptx_single
2017-07-17 Tom de Vries <t...@codesourcery.com>
PR target/81069
* config/nvptx/nvptx.c (nvptx_single): Insert diverging branch as late
as possible.
---
gcc/config/nvptx/nvptx.c | 22 +++++++++++++++++++---
1 file changed, 19 insertions(+), 3 deletions(-)
diff --git a/gcc/config/nvptx/nvptx.c b/gcc/config/nvptx/nvptx.c
index daeec27..cb11686 100644
--- a/gcc/config/nvptx/nvptx.c
+++ b/gcc/config/nvptx/nvptx.c
@@ -3866,9 +3866,25 @@ nvptx_single (unsigned mask, basic_block from,
basic_block to)
rtx_insn *tail = BB_END (to);
unsigned skip_mask = mask;
- /* Find first insn of from block */
- while (head != BB_END (from) && !INSN_P (head))
- head = NEXT_INSN (head);
+ while (true)
+ {
+ /* Find first insn of from block. */
+ while (head != BB_END (from) && !INSN_P (head))
+ head = NEXT_INSN (head);
+
+ if (from == to)
+ break;
+
+ if (!(JUMP_P (head) && single_succ_p (from)))
+ break;
+
+ basic_block jump_target = single_succ (from);
+ if (!single_pred_p (jump_target))
+ break;
+
+ from = jump_target;
+ head = BB_HEAD (from);
+ }
/* Find last insn of to block */
rtx_insn *limit = from == to ? head : BB_HEAD (to);