GitLab

Build GHC via LLVM backend fails if option -keep-llvm-files is present and multiple threads are used, i.e., make -j.

Compile GHC via ./boot && ./configure && make -j using the following build.mk file:

include mk/flavours/quick-llvm.mk
GhcStage2HcOpts += -keep-llvm-files
GhcLibHcOpts += -keep-llvm-files
GhcRtsHcOpts += -keep-llvm-files

results in the following error:

...
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.dyn_o] Error 1
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.l_o] Error 1
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.thr_l_dyn_o] Error 1
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.thr_o] Error 1
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.thr_dyn_o] Error 1
make[1]: *** [rts/ghc.mk:315: rts/dist/build/AutoApply.l_dyn_o] Error 1
opt: rts/dist/build/AutoApply.ll:20589:28: error: expected value token
  %ln4ve = mul i64 %ln4vd, g, i64* noalias nocapture %Sp_Arg, i64* noalias nocapture %Hp_Arg, i64 %R1_Arg, i64 %SpLim_Arg) align 8 nounwind
                           ^
`opt' failed in phase `LLVM Optimiser'. (Exit code: 1)
...

An excerpt of file rts/dist/build/AutoApply.ll looks as follows:

20583   tail call ghccc void (i64*, i64*, i64*, i64, i64) %ln4v7( i64* %Base_Arg, i64* %ln4v8, i64* %ln4v9, i64 %ln4va, i64 %SpLim_Arg ) nounwind
20584   ret void
20585 cBT:
20586   %ln4vb = load i64*, i64** %Sp_Var
20587   %ln4vc = ptrtoint i64* %ln4vb to i64
20588   %ln4vd = sub i64 0, 1
20589   %ln4ve = mul i64 %ln4vd, 8
20590   %ln4vf = add i64 %ln4vc, %ln4ve
20591   %ln4vg = inttoptr i64 %ln4vf to i64*
20592   store i64* %ln4vg, i64** %Sp_Var
20593   %ln4vh = load i64*, i64** %Sp_Var

For example, the function signature is messed up.

It looks like as if all builds of different ways of AutoApply are using the same intermediate LLVM file resulting in a race condition.

GHC release 8.6.5 (git 92b6a023) was used as stage0 compiler and as the actual target compiler.