flintlib
Valgrind reports unhandled instruction
When running Valgrind on an example for the FUEL library (https://gitlab.com/feynmanintegrals/fuel), it encounters an unhandled instruction:
==903142== Memcheck, a memory error detector
==903142== Copyright (C) 2002-2022, and GNU GPL'd, by Julian Seward et al.
==903142== Using Valgrind-3.21.0 and LibVEX; rerun with -h for copyright info
==903142== Command: ./test --calc flint --file example_coeffs/huge_but_short.data
==903142==
vex amd64->IR: unhandled instruction bytes: 0x62 0x73 0xB5 0x48 0x3A 0xD4 0x1 0xC4 0x43 0xD
vex amd64->IR: REX=0 REX.W=0 REX.R=0 REX.X=0 REX.B=0
vex amd64->IR: VEX=0 VEX.L=0 VEX.nVVVV=0x0 ESC=NONE
vex amd64->IR: PFX.66=0 PFX.F2=0 PFX.F3=0
==903142== valgrind: Unrecognised instruction at address 0x513fd46.
==903142== at 0x513FD46: mpoly_void_ring_init_fmpz_mpoly_ctx (void_ring.c:113)
==903142== by 0x5123EEA: fmpz_mpoly_set_str_pretty (set_str_pretty.c:24)
==903142== by 0x14B359: rational_function (ratfunc_flint.cpp:1212)
...
This did not happen with v3.0.1. However, to fix #1652, FUEL was updated to commit 3ef8253157dd22b1480f7669dbb3684008637ded. I checked that the problem persists with the most recent commit 2318e1bebd5df5723db7e048494ce31bc4baff61 in trunk.
I noticed this in the context of looking into a memory leak in FUEL, see https://gitlab.com/feynmanintegrals/fuel/-/issues/1 and #1676.
I'm bad at valgrind -- do I understand it correctly that the unrecognized instruction is in fmpz_mpoly/void_ring.c at line 113 in FLINT?
What CFLAGS were used when you compiled it with 3.0.1, and what CFLAGS are being used now? Are they the same or not?
I'm bad at valgrind -- do I understand it correctly that the unrecognized instruction is in
fmpz_mpoly/void_ring.cat line 113 in FLINT?
Yes, it pritnt a backtrace of calls, the top line is the inner call in the stack
Unless @langefa changed something, it is should be by default, and the flags should be the same. We only pass paths to gmp and mpfr
Okay, so the default CFLAGS where changes some time after that. We increased the optimization level from -O2 to -O3 and added -march=native (unless cross-compiling). It may be the case that Valgrind simply does not recognize the instruction that were generated (see https://stackoverflow.com/questions/37032339/valgrind-unrecognised-instruction for example).
Is there any recomended way to force specific CFLAGS? I will force their usage in FUEL, so that @langefa can retry (I have not seen the error message)
Assuming you are using Make to build FLINT, yes. For instance,
$ ./configure CFLAGS="-O2"
will set CFLAGS to -O2, and there may be some other default options that may be put in there too (for instance, --enable-debug is the default, so -g will also be added to the CFLAGS unless you specify --disable-debug).
@albinahlback , thank ypu @langefa , I pushed a commit so that fuel will configure flint with -O2, please give it a try
The CFLAGS are indeed the culprit. -O3 has no impact though, -march=native is solely responsible. It looks like there is some architecture dependent optimization on my machine that is not compatible with Valgrind.
What is the resolution then? Nothing wrong here in flint, the issue can be closed? I am probably then returning default opitons and you can change them if needed
This is a common problem with valgrind. Before closing the issue, we should add a section to the FLINT docs on using valgrind including a mention of this problem.
The system you built this for, does it support AVX512? I seem to disassemble the instruction to vinserti64x4, which is an AVX512 instruction which seems to be unsupported by Valgrind to-this-day.