Program does not terminate when compiled with `-ssa-passes minimal`

[nanikamado]

$ cat test.sml
print "ok\n"

$ mlton -verbose 1 test.sml
MLton 20251030.021344-g61baac710 starting
   Compile SML starting
   Compile SML finished in 1.59 + 0.57 (26% GC)
   Compile and Assemble starting
      cc -c -I/home/mado/mlton_latest/lib/mlton/targets/self/include \
          -std=gnu11 -fno-common -O1 -fno-strict-aliasing \
          -foptimize-sibling-calls -w \
          -I/home/mado/mlton_latest/lib/mlton/include -m64 -o \
          /tmp/fileNXmxXy.o /tmp/fileEUR4bT.1.c
      cc -c -m64 -o /tmp/file6XGccd.o /tmp/file5JDfWA.0.s
   Compile and Assemble finished in 0.07 + 0.00 (0% GC)
   Link starting
      cc -o test /tmp/fileNXmxXy.o /tmp/file6XGccd.o \
          -L/home/mado/mlton_latest/lib/mlton/targets/self -lmlton -lgdtoa \
          -lm -lgmp -m64 -Wl,-znoexecstack
   Link finished in 0.02 + 0.00 (0% GC)
MLton 20251030.021344-g61baac710 finished in 1.68 + 0.57 (25% GC)
$ ./test
ok

$ mlton -verbose 1 -ssa-passes minimal test.sml
MLton 20251030.021344-g61baac710 starting
   Compile SML starting
   Compile SML finished in 2.34 + 0.74 (24% GC)
   Compile and Assemble starting
      cc -c -I/home/mado/mlton_latest/lib/mlton/targets/self/include \
          -std=gnu11 -fno-common -O1 -fno-strict-aliasing \
          -foptimize-sibling-calls -w \
          -I/home/mado/mlton_latest/lib/mlton/include -m64 -o \
          /tmp/fileTnDEsr.o /tmp/fileSHlP0N.3.c
      cc -c -m64 -o /tmp/filewa6FBv.o /tmp/fileF5ohaH.2.s
      cc -c -m64 -o /tmp/file29IFo3.o /tmp/file1tXAZL.1.s
      cc -c -m64 -o /tmp/fileWJ1NMO.o /tmp/fileYeOlFF.0.s
   Compile and Assemble finished in 0.14 + 0.00 (0% GC)
   Link starting
      cc -o test /tmp/fileTnDEsr.o /tmp/filewa6FBv.o /tmp/file29IFo3.o \
          /tmp/fileWJ1NMO.o -L/home/mado/mlton_latest/lib/mlton/targets/self \
          -lmlton -lgdtoa -lm -lgmp -m64 -Wl,-znoexecstack
   Link finished in 0.03 + 0.00 (0% GC)
MLton 20251030.021344-g61baac710 finished in 2.52 + 0.74 (23% GC)
$ ./test
(does not terminate)

Environment:

System:
  Kernel: 6.17.8-200.fc42.x86_64 arch: x86_64 bits: 64
  Desktop: GNOME v: 48.6 Distro: Fedora Linux 42 (Workstation Edition)
CPU:
  Info: 6-core AMD Ryzen 5 5625U with Radeon Graphics [MT MCP] speed (MHz):
    avg: 1111 min/max: 423/4390

Is this a bug, or am I using the compiler incorrectly?

[MatthewFluet]

I don't think that it is necessarily a bug. It's just a consequences of -ssa-passes minimal disabling almost all SSA passes, including inlining, dead-code elimination, common-subexpression elimination, constant propagation, globalization, and self tail-call to loop optimizations. The resulting program could just be very slow. Remember that implicit in that simple program is executing the entirety of the Basis Library; while there is little meaningful computation, there is some initialization of data structures and creation of many many closures (which remain, because the corresponding functions aren't inlined).

That said, there might be some strange interaction between the -ssa-passes minimal and later stages of the compiler that trigger an infinite loop. For example, compiling with -ssa-passes "inlineLeaf(true,true,20):constantPropagation:polyEqual" yields an executable that runs quickly. And dropping either of inlineLeaf(true,true,20) or constantPropagation leads to an executable that may be non-terminating. And, I observe that compiling with -ssa-passes "inlineLeaf(true,true,20):constantPropagation:polyEqual" -disable-pass ssaOrderFunctions triggers an internal compiler error with the SSA RefFlatten pass (refFlatten raised: Fail: RefFlatten.Value.unify: unequal Grounds in call tail st_0 (global_0, x_0) in L_0 in x_1). (Compiling with -ssa-passes minimal is essentially the same as -ssa-passes "polyEqual:ssaOrderFunctions".)

What is the goal for compiling with -ssa-passes minimal?

[nanikamado]

What is the goal for compiling with -ssa-passes minimal?

I found a benchmark program that MLton optimizes extremely well, and I got curious about what kind of optimizations were critically contributing to that result. So I was experimenting by turning some optimization passes on and off.

I'm not a serious user, so if this isn't a bug, it's ok for me. Maybe I just shouldn't have been messing with this option.

[MatthewFluet]

You might find using -verbose 2 (to see the names of all passes) and -disable-pass <regex> (can be given multiple times) a more useful way to experiment.

[nanikamado]

Oh, thank you!