please disable multi-threaded linking or port upstream patch for LLVM wasm-ld threadpool deadlock issue

[godmar]

A brief follow-up on my note in #123 - eosio.cdt 1.6.1 runs with parallel linking enabled, but there's a deadlock in LLVM's threadpool that doesn't allow the parallel writing of sections and custom sections in the WASM backend. See here for the upstream bug that was fixed in LLVM 8.0.1.

Other projects affected by it have disabled parallel linking, e.g. ziglang disabled parallel linking, see here for their bugfix. You can do the same, or possibly backport their fix - which is just a few lines of code.

I've copied and pasted more details, including a backtrace of when eosio-cpp deadlocks, here

[larryk85]

This is on the radar, we will be moving to LLVM 9 as the backbone next release. This will remove this issue, and hopefully not create new ones.

[godmar]

This is a really unfortunate issue. I'm right now hitting this reproducibly on a t2.medium EC2 instance (which does have 2 virtual CPUs). Of course, the deadlock occurs if these cores aren't scheduled often enough in the underlying multi-tenant environment, which makes the entire enterprise unpredictable and pretty annoying.

[godmar]

Here's a work-around that doesn't require patching and rebuilding the cdt:

cd /usr/opt/eosio.cdt/1.6.1/bin
sudo mv wasm-ld wasm-ld.real

then create a file /usr/opt/eosio.cdt/1.6.1/bin/wasm-ld like so:

#!/bin/bash

echo "Running " `dirname $0`/wasm-ld.real --no-threads "$@"
`dirname $0`/wasm-ld.real --no-threads "$@"

Then sudo chmod +x /usr/opt/eosio.cdt/1.6.1/bin/wasm-ld and then use eosio-cpp as before.

[jeffreyssmith2nd]

This should be resolved in develop with #896