Feature request: `-keep-all`, `-keep-dir`

[pscollins]

MLIR has a set of compiler command line flags to control printing IR for intermediate passes; a useful combination is (e.g.):

• -mlir-print-ir-after-all: dumps IR after every pass
• -mlir-print-ir-tree-dir=out/: writes each pass output into a unique, numbers file, e.g. if the first 3 passes to run are passA, passB, passA, the outputs are out/0_passA.mlir, out/1_passB.mlir, out/2_passA.mlir, etc. Specifying a directory makes it easy to dump outputs for e.g. an existing Makefile rule, without having to change the directory where the compiler binary runs; including the pass index solves the problem of collisions and also makes it easy for someone who is unfamiliar with the compilation pipeline to tell the ordering.

These two flags enable a helpful debugging workflow: dump everything, find an interesting symbol/type/label/etc in one pass, and then grep through all of the per-pass output to see the passes that introduce/remove it

I hacked something in my local branch of MPL to do something similar:

• MPLLang/mpl@b9dc27d
• MPLLang/mpl@6facf8e

Would it make sense to clean these up and upstream them here?

[MatthewFluet]

The former (-keep-all) is already available as -keep-pass '.*' --- the -keep-pass (and -diag-pass) compile-time options accept a regular expression. The the indexing is reasonable, though I would ask that that be under a separate compile-time option, since unconditionally indexing could break some existing -keep workflows.

The latter (-keep-dir) is reasonable as is well. At one point, I thought that the -keep outputs were automatically put at the same directory as the -output, but maybe we changed it from that to placing them in the current directory.

[pscollins]

Great!

I would ask that that be under a separate compile-time option, since unconditionally indexing

OK -- any preference on the name? -keep-with-index {true|false}? -keep-prepend-pass-index?

Also, is there any documentation or anything that should be updated at the same time?

maybe we changed it from that to placing them in the current directory

FWIW I see -keep g follow -output but not -keep-pass:

pscollins@instance-20251211-022831:/tmp/test$ echo 'print "test"' > test.sml
pscollins@instance-20251211-022831:/tmp/test$ mkdir subdir
pscollins@instance-20251211-022831:/tmp/test$ mlton -keep-pass 'deadCode' -keep g -output subdir/test.bin test.sml
pscollins@instance-20251211-022831:/tmp/test$ ls *
test.deadCode.post.core-ml  test.deadCode.pre.core-ml  test.sml

subdir:
test.bin  test.bin.0.s  test.bin.1.c