[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 --- Comment #7 from Tom de Vries --- Author: vries Date: Mon Sep 25 10:36:23 2017 New Revision: 253145 URL: https://gcc.gnu.org/viewcvs?rev=253145&root=gcc&view=rev Log: [nvptx] Add exit after call to noreturn function 2017-09-25 Tom de Vries PR target/80035 PR target/81069 * config/nvptx/nvptx.c (nvptx_output_call_insn): Add exit after call to noreturn function. Modified: trunk/gcc/ChangeLog trunk/gcc/config/nvptx/nvptx.c
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 Tom de Vries changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #8 from Tom de Vries --- patch committed, marking resolved-fixed.
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 --- Comment #6 from Tom de Vries --- (In reply to Tom de Vries from comment #2) > Replacing trap with exit or ret (or adding it after trap), makes the sigsegv > go away. Nvidia suggested a workaround of adding exit after trap for this (PR81069 comment 3) bug report.
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 --- Comment #1 from Tom de Vries --- Created attachment 40965 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40965&action=edit pr58640.s
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035
Tom de Vries changed:
What|Removed |Added
CC||tschwinge at gcc dot gnu.org
--- Comment #2 from Tom de Vries ---
> {
> .param.u32 %out_arg1;
> st.param.u32 [%out_arg1], %r31;
> call exit, (%out_arg1);
> trap; // (noreturn)
> }
At nvptx_output_call_insn, we see:
...
if (find_reg_note (insn, REG_NORETURN, NULL))
/* No return functions confuse the PTX JIT, as it doesn't realize
the flow control barrier they imply. It can seg fault if it
encounters what looks like an unexitable loop. Emit a trailing
trap, which it does grok. */
fprintf (asm_out_file, "\t\ttrap; // (noreturn)\n");
...
Replacing trap with exit or ret (or adding it after trap), makes the sigsegv go
away.
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035
Alexander Monakov changed:
What|Removed |Added
CC||amonakov at gcc dot gnu.org
--- Comment #3 from Alexander Monakov ---
Noreturn is a bit of a distraction. On the one hand, yes, on this and related
testcases it may be possible to workaround things more successfully on the gcc
side, when gcc knows that the callee _never_ returns. On the other hand, the
real issue is that ptxas gets confused around loops lacking exit edges (and
perhaps other "unusual" CFG structures). They can easily appear even in
original source and do not require presence of noreturn functions:
extern __device__ void g();
extern __device__ void h();
__device__ void f(int v)
{
if (v) goto L;
for (;;) {
g();
L:
h();
}
}
This minimized testcase leads to ptxas segfault with nvcc, so they are not
immune either. And here a compiler can't do anything to make code more
digestible to ptxas, afaict, so it really should be fixed in ptxas.
FWIW, it seems to work with -arch sm_50, so it might be fixed for newer
architectures (I suspect CUDA driver and toolkit carry separate code generators
for sm_3x and sm_5x).
Can you share if you (Mentor Embeddded) reported such issues to NVIDIA in the
past?
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 --- Comment #4 from Tom de Vries --- (In reply to Tom de Vries from comment #2) > Replacing trap with exit or ret (or adding it after trap), makes the sigsegv > go away. This problem goes away for ptxas -ori starting cuda 7.0.
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035
--- Comment #5 from Tom de Vries ---
(In reply to Alexander Monakov from comment #3)
> ptxas gets confused around loops lacking
> exit edges (and perhaps other "unusual" CFG structures). They can easily
> appear even in original source and do not require presence of noreturn
> functions:
>
> extern __device__ void g();
> extern __device__ void h();
> __device__ void f(int v)
> {
> if (v) goto L;
> for (;;) {
> g();
> L:
> h();
> }
> }
>
This problem goes away for ptxas -ori in cuda 6.5 and later (haven't tried
earlier versions).
[Bug target/80035] [nvptx] non-returning function call causes ptxas sigsegv
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80035 Thomas Schwinge changed: What|Removed |Added Status|UNCONFIRMED |ASSIGNED Last reconfirmed||2017-03-15 Assignee|unassigned at gcc dot gnu.org |vries at gcc dot gnu.org Ever confirmed|0 |1
