<mihaid>
hello, I have a question regarding the LTO flag. I have noticed it has been disabled by default, and I was working on enabling it for PGO. Should I try enabling it as an option at translation?
<fijal>
mihaid: it happens to cause bugs, so that's why we disabled it
<arigato>
is there some interaction between LTO and PGO that must be taken care of?
<mihaid>
fijal: i know, I have seen the issue (issues/2572). However, the issue says this happens for when debugging under Ubuntu 14.04. This is why I asked if this should become an option to enable when building PyPy.
<mihaid>
arigato: no, PGO + LTO work smooth, at least on Ubuntu 16.04 and gcc 6.2.0
<arigato>
mihaid: let me rephrase, I think there is some confusion
Mateon2 has joined #pypy
<arigato>
the original reason for why we disabled LTO is that on Ubuntu 14.04 with gcc-6, where you get gcc 6.2.0 too, it builds bogus code
<arigato>
now the additional reason for disabling LTO is that "-g -flto" is not usable under gdb, for pypy
<arigato>
mjacob is experimenting for some workarounds for that (maybe compiling most files with -g except the ones you're interested in...)
<arigato>
now my question just above was, I think that LTO is unrelated to PGO, is that correct? in the sense that we can add PGO to pypy and not worry about LTO, or vice-versa
<Mateon2>
Hello, I'm trying to use pypy (2.7) for machine learning with spearmint for hyperparameter optimization. I'm having trouble with numpy and tensorflow. I am on a debian system. I installed numpy with pypy -m pip install numpy, but I get errors on importing, and I can't find tensorflow in pip at all.
buhman has quit [Ping timeout: 246 seconds]
<arigato>
Mateon2: what error do you get on importing numpy?
<mihaid>
arigato: yes, PGO has nothing to do with LTO. It works just fine without it. I was asking because they showed good improvement together, but I had no idea the issue behind LTO was this bad.
<arigato>
Mateon2: ah
<arigato>
Mateon2: upgrade to pypy 5.7.1 or the about-to-be-released pypy 5.8.0
<Mateon2>
arigato: Ah, okay. I got this version from my package manager, I'll upgrade from source
<arigato>
it's numpy that added a call to Py_EnterRecursiveCall(), which we happened not to support in older versions
<arigato>
you can also force an older numpy to be installed, but I can't tell you exactly how
<arigato>
and it's not the best of ideas, given that we're always fixing pypy-specific issues
<Mateon2>
Is pypy2-v5.8.0-linux64 not the release I want? There is no make install, so I copied everything to /usr/local, but it complains about libssl
<Mateon2>
Okay, didn't see the install instructions, sorry
<Mateon2>
But, ./bin/pypy also fails, with the same libssl.so.1.0.0 error message
cstratak has quit [Remote host closed the connection]
<Mateon2>
arigato: Well, I can't figure it out, as the packaged binaries fail to load libssl. The current version on my system is 1.1.0, which apparently conflicts with 1.0.0. I'm now building from source and enjoying mandelbrot
<arigato>
Mateon2: yes, that's explained on that link
<arigato>
along with alternative downloads like Squeak's portable pypy
<arigato>
Squeaky's
cstratak has quit [Ping timeout: 258 seconds]
marr has joined #pypy
nimaje1 has joined #pypy
nimaje1 is now known as nimaje
nimaje is now known as Guest64554
<Mateon2>
Awesome, I finished building pypy, and numpy works flawlessly.
<LarstiQ>
Mateon2: nice to hear, and tensorflow?
<Mateon2>
Now, my other problem is that `pypy -m pip install tensorflow` doesn't find anything, do I have to build from source?
<LarstiQ>
Mateon2: in the past I saw wishlist bugs for supporting pypy
<LarstiQ>
Mateon2: you might need to do some detective work
Guest1244 has quit [Ping timeout: 260 seconds]
<arigato>
note that we use the unmodified pip, so if it doesn't find anything, then probably it wouldn't on CPython either
<arigato>
are you sure about "pip install tensorflow" on CPython?