[nvptx] Handle nop in prevent_branch_around_nothing
authorTom de Vries <tdevries@suse.de>
Sun, 23 Jan 2022 05:42:24 +0000 (06:42 +0100)
committerTom de Vries <tdevries@suse.de>
Tue, 1 Feb 2022 18:28:39 +0000 (19:28 +0100)
commit456de10c549379b74d4858f00d4b8817035a73fc
tree4dafac323caa9bcde432d0530d4cc579097702ff
parente0451f93d9faa13495132f4e246e9bef30b51417
[nvptx] Handle nop in prevent_branch_around_nothing

When running libgomp test-case reduction-7.c on an nvptx accelerator
(T400, driver version 470.86) and GOMP_NVPTX_JIT=-O0, I run into:
...
reduction-7.exe:reduction-7.c:312: v_p_2: \
  Assertion `out[j * 32 + i] == (i + j) * 2' failed.
FAIL: libgomp.oacc-c/../libgomp.oacc-c-c++-common/reduction-7.c \
  -DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none \
  -O0  execution test
...

During investigation I found ptx code like this:
...
@ %r163 bra $L262;
$L262:
...

There's a known problem with executing this type of code, and a workaround is
in place to address this: prevent_branch_around_nothing.  The workaround does
not trigger though because it doesn't handle the nop insn.

Fix this by handling the nop insn in prevent_branch_around_nothing.

Tested libgomp on x86_64 with nvptx accelerator.

gcc/ChangeLog:

2022-01-27  Tom de Vries  <tdevries@suse.de>

PR target/100428
* config/nvptx/nvptx.cc (prevent_branch_around_nothing): Handle nop
insn.
gcc/config/nvptx/nvptx.cc