Bad PhiC node placement/usage corrupts gc-preserve use

[KristofferC]

Report: https://github.com/maleadt/BasePkgEvalReports/blob/0bb12607c245153c8e7215788e580d43d7d2d51a/pkgeval-9bd498d_vs_2e6715c/logs/Strided/1.4.0-DEV-082f446f52.log

julia: /workspace/srcdir/src/codegen.cpp:4342: jl_cgval_t emit_expr(jl_codectx_t&, jl_value_t*, ssize_t): Assertion `token.V->getType()->isTokenTy()' failed.

signal (6): Aborted
in expression starting at /home/pkgeval/.julia/packages/Strided/qA9SM/test/runtests.jl:228
gsignal at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
abort at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
unknown function (ip: 0x7f452283e399)
__assert_fail at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
emit_expr at /workspace/srcdir/src/codegen.cpp:4342
emit_ssaval_assign at /workspace/srcdir/src/codegen.cpp:3836
emit_stmtpos at /workspace/srcdir/src/codegen.cpp:4029 [inlined]
emit_function at /workspace/srcdir/src/codegen.cpp:6649
jl_compile_linfo at /workspace/srcdir/src/codegen.cpp:1257
jl_compile_method_internal at /workspace/srcdir/src/gf.c:1889
_jl_invoke at /workspace/srcdir/src/gf.c:2153 [inlined]
jl_invoke at /workspace/srcdir/src/gf.c:2165
jl_toplevel_eval_flex at /workspace/srcdir/src/toplevel.c:808
jl_parse_eval_all at /workspace/srcdir/src/ast.c:872
jl_load at /workspace/srcdir/src/toplevel.c:872
include at ./client.jl:439
_jl_invoke at /workspace/srcdir/src/gf.c:2144 [inlined]
jl_apply_generic at /workspace/srcdir/src/gf.c:2322
jl_apply at /workspace/srcdir/src/julia.h:1692 [inlined]
do_call at /workspace/srcdir/src/interpreter.c:369
eval_value at /workspace/srcdir/src/interpreter.c:458
eval_stmt_value at /workspace/srcdir/src/interpreter.c:409 [inlined]
eval_body at /workspace/srcdir/src/interpreter.c:817
jl_interpret_toplevel_thunk at /workspace/srcdir/src/interpreter.c:911
jl_toplevel_eval_flex at /workspace/srcdir/src/toplevel.c:814
jl_toplevel_eval_flex at /workspace/srcdir/src/toplevel.c:764
jl_toplevel_eval_in at /workspace/srcdir/src/toplevel.c:843
eval at ./boot.jl:331
_jl_invoke at /workspace/srcdir/src/gf.c:2144 [inlined]
jl_apply_generic at /workspace/srcdir/src/gf.c:2322
exec_options at ./client.jl:264
_start at ./client.jl:484
jfptr__start_2076.clone_1 at /opt/julia/lib/julia/sys.so (unknown line)
_jl_invoke at /workspace/srcdir/src/gf.c:2144 [inlined]
jl_apply_generic at /workspace/srcdir/src/gf.c:2322
jl_apply at /workspace/srcdir/ui/../src/julia.h:1692 [inlined]
true_main at /workspace/srcdir/ui/repl.c:96
main at /workspace/srcdir/ui/repl.c:217
__libc_start_main at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
unknown function (ip: 0x4015e4)
Allocations: 585069033 (Pool: 584874239; Big: 194794); GC: 445

[vtjnash]

How coincidental: this is a duplicate of #34229 that I opened the day before.

julia> function f(a)
         GC.@preserve a try
         catch
         end
       end
f (generic function with 1 method)

julia> f("")
julia: /data/vtjnash/julia/src/codegen.cpp:4358: jl_cgval_t emit_expr(jl_codectx_t&, jl_value_t*, ssize_t): Assertion `token.V->getType()->isTokenTy()' failed.

[Keno]

I'm thinking the way to fix this would be to allow ssa uses across try/catch regions, coupled with an LLVM pass that either manually spills all state to the stack or outlines the body of the try/catch block.

[vtjnash]

Hm, yeah, I think yuyichao was exploring something like that in #24573, but I don't know how feasible that is. In theory, the C standard says that LLVM is supposed to be OK with any variable that isn't modified in the try block being correct afterwards, and we seemed to be mostly OK for some years with assuming that. I still hope that someday we could move to using the lower-cost and more correct invoke instruction handling though.

[maleadt]

FWIW, bisected to cd92f60 (the upgrade to LLVM 8.0.1).