Random segmentation faults on running artiq_sinara_tester #275

New Issue

mwojcik · 2022-04-26T12:14:12+08:00

mwojcik commented

2022-04-26 12:14:12 +08:00

It only happens on startup - regardless if the device is connected or not, and usually running it again helps (or running it few times causes the SIGSEGV to pop up again). I found it happens regardless of device_db file, but I still attached one just in case.

The output I get is this:

[spaqin@hera:~/m-labs/artiq_nac3]$ artiq_sinara_tester 
****** Sinara system tester ******

Segmentation fault (core dumped)

and that's it.

I attach the device_db I've been using for this and would like to also attach the core dump file but it's bigger than 4MB limit (just ask me if you need it, I'll send it over by other means)

It only happens on startup - regardless if the device is connected or not, and usually running it again helps (or running it few times causes the SIGSEGV to pop up again). I found it happens regardless of device_db file, but I still attached one just in case. The output I get is this: ``` [spaqin@hera:~/m-labs/artiq_nac3]$ artiq_sinara_tester ****** Sinara system tester ****** Segmentation fault (core dumped) ``` and that's it. I attach the device_db I've been using for this and would like to also attach the core dump file but it's bigger than 4MB limit (just ask me if you need it, I'll send it over by other means)

device_db.py

3.8 KiB

ychenfo referenced this issue

2022-04-27 04:07:50 +08:00

Phaser: "Call parameter does not match function signature!" #276

sb10q added the

high-priority

label 2022-04-27 10:56:25 +08:00

sb10q added this to the Alpha milestone 2022-04-27 10:56:27 +08:00

ychenfo was assigned by sb10q

2022-04-27 10:56:36 +08:00

ychenfo commented

2022-06-07 03:12:53 +08:00

It's a bit weird but this line of setting up the population pass seems to be the cause of the problem:

		nac3core/src/codegen/mod.rs
		Line 224 in 8e6e4d6715
	
				        pass_builder.populate_function_pass_manager(&passes);

Simply commenting it out solves the problem, at the expense of lesser optimization.

Replacing this line with the following, as written in the inkwell example

passes.add_instruction_combining_pass();
passes.add_reassociate_pass();
passes.add_gvn_pass();
passes.add_cfg_simplification_pass();
passes.add_basic_alias_analysis_pass();
passes.add_promote_memory_to_register_pass();
passes.add_instruction_combining_pass();
passes.add_reassociate_pass();
passes.initialize();

also will not cause segfault, and will produce more optimized code which seems still not as optimized as the current output, though.

It's a bit weird but this line of setting up the population pass seems to be the cause of the problem: https://git.m-labs.hk/M-Labs/nac3/src/commit/8e6e4d6715bdfa4a5639068e4a78e89b714fb8d7/nac3core/src/codegen/mod.rs#L224 Simply commenting it out solves the problem, at the expense of lesser optimization. Replacing this line with the following, as written in the [inkwell example](https://github.com/TheDan64/inkwell/blob/7253f9250d8a5e5e83125f1f6f85ba8b80d12acc/examples/kaleidoscope/main.rs#L1268-L1277) ```rust passes.add_instruction_combining_pass(); passes.add_reassociate_pass(); passes.add_gvn_pass(); passes.add_cfg_simplification_pass(); passes.add_basic_alias_analysis_pass(); passes.add_promote_memory_to_register_pass(); passes.add_instruction_combining_pass(); passes.add_reassociate_pass(); passes.initialize(); ``` also will not cause segfault, and will produce more optimized code which seems still not as optimized as the current output, though.

sb10q commented

2022-06-07 07:22:41 +08:00

Check the source code of populateFunctionPassManager in LLVM?

Check the source code of ``populateFunctionPassManager`` in LLVM?

z78078 referenced this issue

2022-07-04 16:18:02 +08:00

275-fix-random-sigsegv #302

z78078 referenced this issue

2022-07-04 16:18:04 +08:00

275-fix-random-sigsegv #303

sb10q referenced this issue from a commit

2022-07-04 18:06:36 +08:00

work around random segmentation fault (#275)

sb10q removed the

high-priority

label 2022-07-04 18:09:17 +08:00

ychenfo was unassigned by sb10q

2022-07-04 18:09:22 +08:00

z78078 was assigned by sb10q

2022-07-04 18:09:22 +08:00

sb10q removed this from the Alpha milestone 2022-07-04 18:09:29 +08:00

derppening commented

2023-10-30 14:29:27 +08:00

As of e49b760e, no segfaults have been observed when running artiq_sinata_tester. I suspect the changes to improve multithreading in LLVM have addressed this issue.

As of e49b760e, no segfaults have been observed when running `artiq_sinata_tester`. I suspect the changes to improve multithreading in LLVM have addressed this issue.

sb10q closed this issue

2023-10-30 15:11:32 +08:00

Sign in to join this conversation.