Evaluation broken on aarch64

[TravisWhitaker]

Evaluation diverges on aarch64. With the current tip of master:

$ nix-build -E 'with import ../nixpkgs ((import ./default.nix {}).nixpkgsArgs); haskell-nix.snapshots."lts-17.9".dhall.components.exes.dhall'
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
trace: WARNING: 8.8.2 is out of date, consider using 8.8.4.
trace: Checking materialization in /home/anduril/sources/haskell.nix/materialized/bootstrap/ghc882/happy-1.19.11
(repeat until OOM)
...
Segmentation fault (core dumped)

Seems to be restricted exprs requiring the 8.10 series though, evaluation proceeds if I do e.g.

$ nix-build -E 'with import ../nixpkgs ((import ./default.nix {}).nixpkgsArgs); haskell-nix.snapshots."lts-16.31".dhall.components.exes.dhall'

[TravisWhitaker]

The problem is indeed caused by any expr that touches the bootstrapping stuff for 8.10.x, even just building the compiler. Bisection shows that 6134e66 is responsible for breaking eval on aarch64, still digging in to why.

[TravisWhitaker]

CC @hamishmack

[TravisWhitaker]

This seems to cause it.

[sloane-shark]

I think I ran into this trying to build a project on Aarch64 that uses GHC 8.10.4. It surfaced as an infinite recursion instead of a segfault. It seems like i can work around it by downgrading GHC to 8.8.4.

[TravisWhitaker]

Turns out this is show-stopping for 8.8.4 on aarch64 too, pkgs.haskell-nix.roots "ghc884" also fails in the same way.

[TravisWhitaker]

Seems it's actually two separate problems, bisection shows that #1129 is exactly what breaks pkgs.haskell-nix.roots "ghc884".

[TravisWhitaker]

Seems it's actually two separate problems, bisection shows that #1129 is exactly what breaks pkgs.haskell-nix.roots "ghc884".

Well, not exactly two separate problems, #1129 just uses cabal-install and nix-tools bootstrapped with ghc 8.10.*, which triggers the same case I originally presented even if you're building things with 8.8.4.

[TravisWhitaker]

This specific eval problem should be fixed on master now.

#1207 (comment)

[peterbecich]

I also run out of memory on amd64: #1327