Fix codegen not handling invoke exprs with Codeinstances iwith jl_fptr_sparam_addr invoke types.

[gbaraldi]

fixes #56739
I didn't succeed in making a test for this. The sole trigger seems to be

using HMMGradients
T = Float32
A = T[0.0 1.0 0.0; 0.0 0.5 0.5; 1.0 0.0 0.0]
t2tr = Dict{Int,Vector{Int}}[Dict(1 => [2]),]
t2IJ= HMMGradients.t2tr2t2IJ(t2tr)

Nt = length(t2tr)+1
Ns = size(A,1)
y = rand(T,Nt,Ns)

c = rand(Float32, Nt)
beta = backward(Nt,A,c,t2IJ,y)
gamma = posterior(Nt,t2IJ,A,y)

in @oscardssmith memorynew PR

One other option is to have the builtin handle receiving a CI. That might make the code cleaner and does handle the case where we receive a dynamic CI (is that even a thing)

[vtjnash]

SGTM. I think we'll delete that other code path (make it an assertion) anyways, since the dynamic version is unused and unnecessary (now that the invoke function supports this instead)

[oscardssmith]

Turns out this is very reproducable on nightly without the memorynew PR: (it works fine on 1.11 though)

julia> f(a) where {T} = a
WARNING: method definition for f at REPL[1]:1 declares type variable T but does not use it.
f (generic function with 1 method)

julia> f(1)
1

julia> g(x) = @noinline f(x)
g (generic function with 1 method)

julia> g(1)

[1492307] signal 11 (1): Segmentation fault
in expression starting at REPL[4]:1
_jl_invoke at /home/oscardssmith/julia/src/gf.c:3340 [inlined]
ijl_invoke at /home/oscardssmith/julia/src/gf.c:3369
g at ./REPL[3]:1
unknown function (ip: 0x7ac2a408e585) at (unknown file)
jl_apply at /home/oscardssmith/julia/src/julia.h:2233 [inlined]
do_call at /home/oscardssmith/julia/src/interpreter.c:125
eval_value at /home/oscardssmith/julia/src/interpreter.c:243
eval_stmt_value at /home/oscardssmith/julia/src/interpreter.c:194 [inlined]
eval_body at /home/oscardssmith/julia/src/interpreter.c:705
jl_interpret_toplevel_thunk at /home/oscardssmith/julia/src/interpreter.c:898
jl_toplevel_eval_flex at /home/oscardssmith/julia/src/toplevel.c:1068
__repl_entry_eval_expanded_with_loc at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:341
jl_apply at /home/oscardssmith/julia/src/julia.h:2233 [inlined]
jl_f__call_latest at /home/oscardssmith/julia/src/builtins.c:883
#invokelatest#1 at ./essentials.jl:1056 [inlined]
invokelatest at ./essentials.jl:1052 [inlined]
toplevel_eval_with_hooks at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:348
toplevel_eval_with_hooks at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:352
toplevel_eval_with_hooks at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:352
toplevel_eval_with_hooks at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:352
toplevel_eval_with_hooks at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:345 [inlined]
eval_user_input at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:370
repl_backend_loop at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:482
#start_repl_backend#41 at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:467
start_repl_backend at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:464 [inlined]
#run_repl#48 at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:690
run_repl at /home/oscardssmith/julia/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:676
jfptr_run_repl_22939 at /home/oscardssmith/julia/usr/share/julia/compiled/v1.12/REPL/u0gqU_qk7sj.so (unknown line)
run_std_repl at ./client.jl:490
jfptr_run_std_repl_108302 at /home/oscardssmith/julia/usr/lib/julia/sys.so (unknown line)
jl_apply at /home/oscardssmith/julia/src/julia.h:2233 [inlined]
jl_f__call_latest at /home/oscardssmith/julia/src/builtins.c:883
#invokelatest#1 at ./essentials.jl:1056 [inlined]
invokelatest at ./essentials.jl:1052 [inlined]
run_main_repl at ./client.jl:511
repl_main at ./client.jl:593 [inlined]
_start at ./client.jl:568
jfptr__start_108359 at /home/oscardssmith/julia/usr/lib/julia/sys.so (unknown line)
jl_apply at /home/oscardssmith/julia/src/julia.h:2233 [inlined]
true_main at /home/oscardssmith/julia/src/jlapi.c:922
jl_repl_entrypoint at /home/oscardssmith/julia/src/jlapi.c:1081
main at /home/oscardssmith/julia/cli/loader_exe.c:58
unknown function (ip: 0x7ac58162814f) at /lib/x86_64-linux-gnu/libc.so.6
__libc_start_main at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
_start at ./julia (unknown line)
Allocations: 1231993 (Pool: 1231898; Big: 95); GC: 2
Segmentation fault

[giordano]

That looks like #56739

[oscardssmith]

that would be because it is :)

[topolarity]

the dynamic version is unused and unnecessary (now that the invoke function supports this instead)

Does that mean we make Expr(:invoke, ::CodeInstance) require the argument to be a constant in the expression (literally - not just an IR value with ::Const type)? Can we document that and add it to the IR verifier?

[Keno]

Does that mean we make Expr(:invoke, ::CodeInstance) require the argument to be a constant in the expression (literally - not just an IR value with ::Const t

Yes, that has always been required.

[oscardssmith]

should we run benchmarks on this to make sure it isn't deoptimizing too many of our calls?

[topolarity]

Yes, that has always been required.

Doesn't that mean we're fixing the wrong bug?

[gbaraldi]

While I added an extra check to the apparently unused dynamic case. The actual fix is that we no longer try to emit specsig calls for MIs with unbound sparams.