codegen: Add llvm_options parameter to code_llvm for pass instrumentation #60698

gbaraldi · 2026-01-15T12:22:26Z

Add the ability to pass LLVM pass manager options directly to code_llvm() for debugging and analysis purposes. This allows users to inspect IR at various stages of the optimization pipeline without needing to set environment variables.

Supported options include:

-print-after-all: Print IR after each pass
-print-before-all: Print IR before each pass
-print-after=<passname>: Print IR after a specific pass
-print-before=<passname>: Print IR before a specific pass
-print-module-scope: Print entire module instead of just the function
-filter-print-funcs=<name>: Filter output to specific functions

Example usage:
code_llvm(+, (Int, Int); llvm_options="-print-after=loop-vectorize")

oscardssmith · 2026-01-15T16:14:34Z

This is amazing! Thank you so much!

oscardssmith · 2026-01-15T16:17:07Z

The only UI question IMO is whether we want the string interface or to use a kwarg interface (e.g. print-after=:allocopt)

gbaraldi · 2026-01-15T17:03:32Z

I initially had something like that but to me it felt weird. This is very much a LLVM api and I think users of this are probably familiar with that syntax. (it's also much simpler to plumb through)

src/pipeline.cpp

stdlib/InteractiveUtils/src/codeview.jl

…tion Add the ability to pass LLVM pass manager options directly to code_llvm() for debugging and analysis purposes. This allows users to inspect IR at various stages of the optimization pipeline without needing to set environment variables. Supported options include: - `-print-after-all`: Print IR after each pass - `-print-before-all`: Print IR before each pass - `-print-after=<passname>`: Print IR after a specific pass - `-print-before=<passname>`: Print IR before a specific pass - `-print-module-scope`: Print entire module instead of just the function - `-print-changed`: Print IR only when it changes - `-filter-print-funcs=<name>`: Filter output to specific functions Example usage: code_llvm(+, (Int, Int); llvm_options="-print-after=loop-vectorize") Co-Authored-By: Claude Opus 4.5 <[email protected]>

vchuravy · 2026-01-15T19:17:01Z

stdlib/InteractiveUtils/src/codeview.jl

+The `llvm_options` keyword argument allows passing LLVM options to control the optimization pipeline output.
+Supported options include:
+- `-print-after-all`: Print IR after each pass
+- `-print-before-all`: Print IR before each pass
+- `-print-after=<passname>`: Print IR after a specific pass (e.g., `-print-after=loop-vectorize`)
+- `-print-before=<passname>`: Print IR before a specific pass
+- `-print-module-scope`: Print entire module instead of just the function
+- `-print-changed`: Print IR only when it changes
+- `-filter-print-funcs=<name>`: Only print IR for functions matching the name


I makes me wonder if we want to support more (all) of the opt pipeline?

In particular remarks? But perhaps custom opt pipeline would also make sense

Yeah. I want this now, but technically we can extend this to pretty much all of the things (triple/target/pipeline etc)

Most of those options are static global variables that we cannot safely use, because llvm is not actually very usable as a library when it comes to options. Already very annoying that the StandardInstrumentations pass has to be copied here and reimplemented since that pass relies heavily on those global variables.

src/pipeline.cpp

vtjnash · 2026-01-15T19:30:01Z

stdlib/InteractiveUtils/src/codeview.jl

    str = @ccall jl_dump_function_ir(llvmf_dump::Ptr{LLVMFDump}, strip_ir_metadata::Bool,
                                     dump_module::Bool, debuginfo::Ptr{UInt8})::Ref{String}
-    return str
+    # Prepend pass output to final IR


Should we prepend ; to each newline in the annotations stream? I think normally in LLVM it might go to stderr while the final output goes to stdout, so they are normally separated that way instead?

str = replace(str, '\n' => "\n;")

So this is implemented, but then I slept on it. I kind of think that we shouldn't make it a comment. Most of the time I'll get this output and put it in viscose or godbolt, maybe give it to an llvm tool. Rarely it's just something to look at, so having the comments might be helpful

It looks like Claude still mentioned this in the commit, but didn't implement it. I'm neutral on which version is better. It should be pretty clear to a human where the real module starts and the debug output ended, or pretty easy to mechanically strip off ; from the start of every line

I decided on no prepend because it gets syntax highlighting (even if not great ones)

vtjnash · 2026-01-16T19:35:40Z

src/aotcompile.cpp

 // --- native code info, and dump function to IR and ASM ---
 // Get pointer to llvm::Function instance, compiling if necessary
 // for use in reflection from Julia.
 // This is paired with jl_dump_function_ir and jl_dump_function_asm, either of which will free all memory allocated here


Looks like jl_dump_function_asm isn't freeing the newly added memory?

vtjnash · 2026-01-16T19:40:50Z

src/pipeline.cpp

+        } else if (Arg.starts_with("-print-after")) {
+            StringRef val = getNextValue(i, Arg, "-print-after");


Can you ask Claude to rewrite this so that is isn't doing just a prefix check, but can have getNextValue implement the exact checks:

Suggested change

} else if (Arg.starts_with("-print-after")) {

StringRef val = getNextValue(i, Arg, "-print-after");

} else if (StringRef val = getNextValue(i, Arg, "-print-after=")) {

If you were going to do a prefix comparison scan only, the getopt standard requires that you do the prefix in the other direction (e.g. StringRef("-print-after").starts_with(Arg.split('=').first)) but that is getting fairly tedious to implement by hand and not something the custom parsers for llvm or gcc arguments currently does

Use LLVM's TokenizeGNUCommandLine for parsing instead of custom logic. Support both "-option=value" and "-option value" syntax with proper bounds checking. Format pass instrumentation output as IR comments by prepending "; " to each line. Simplify memory management by having jl_dump_function_ir handle freeing pass_output internally. Co-Authored-By: Claude Opus 4.5 <[email protected]>

gbaraldi requested review from vchuravy and vtjnash January 15, 2026 12:22

gbaraldi force-pushed the gb/print-options-llvm branch 5 times, most recently from 9be33d6 to e363be3 Compare January 15, 2026 15:39

vtjnash reviewed Jan 15, 2026

View reviewed changes

src/pipeline.cpp Outdated Show resolved Hide resolved

vtjnash reviewed Jan 15, 2026

View reviewed changes

stdlib/InteractiveUtils/src/codeview.jl Outdated Show resolved Hide resolved

gbaraldi force-pushed the gb/print-options-llvm branch from e363be3 to 6e72b9d Compare January 15, 2026 19:06

gbaraldi force-pushed the gb/print-options-llvm branch from 6e72b9d to 12d1b7e Compare January 15, 2026 19:14

vchuravy reviewed Jan 15, 2026

View reviewed changes

vtjnash reviewed Jan 15, 2026

View reviewed changes

src/pipeline.cpp Outdated Show resolved Hide resolved

vtjnash reviewed Jan 15, 2026

View reviewed changes

gbaraldi force-pushed the gb/print-options-llvm branch from 7678a81 to 7ddc6e3 Compare January 16, 2026 19:20

vtjnash reviewed Jan 16, 2026

View reviewed changes

gbaraldi force-pushed the gb/print-options-llvm branch from 7ddc6e3 to 47a88d4 Compare January 16, 2026 19:58

gbaraldi force-pushed the gb/print-options-llvm branch from 47a88d4 to 2bc0119 Compare January 16, 2026 20:28

		} else if (Arg.starts_with("-print-after")) {
		StringRef val = getNextValue(i, Arg, "-print-after");

	} else if (Arg.starts_with("-print-after")) {
	StringRef val = getNextValue(i, Arg, "-print-after");
	} else if (StringRef val = getNextValue(i, Arg, "-print-after=")) {

Uh oh!

codegen: Add llvm_options parameter to code_llvm for pass instrumentation #60698

Are you sure you want to change the base?

codegen: Add llvm_options parameter to code_llvm for pass instrumentation #60698

Conversation

gbaraldi commented Jan 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

oscardssmith commented Jan 15, 2026

Uh oh!

oscardssmith commented Jan 15, 2026

Uh oh!

gbaraldi commented Jan 15, 2026

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

vtjnash Jan 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

gbaraldi commented Jan 15, 2026 •

edited

Loading

vtjnash Jan 15, 2026 •

edited

Loading