Cleanup

cleanup_code(::Val{code_type}, code, dbinfo=true, cleanup_opts=(;))

Perform minor changes to code to improve readability and the quality of the differences.

dbinfo is a superset of debuginfo. It is compatible with all code types, but it may have no effect.

cleanup_opts are passed by the user, and have specific meaning depending on the code_type.

cleanup_code(::Val{:typed}, code, dbinfo, cleanup_opts)

Cleanup Julia typed IR code.

Accepted cleanup_opts and their default values:

• expand_llvmcall=true: replace raw inline LLVM IR with multiline blocks with syntax highlighting.

cleanup_code(::Val{:ptx}, code, dbinfo, cleanup_opts)

Cleanup PTX code.

Accepted cleanup_opts and their default values:

• demangle=true: demangle names within the code
• keep_loop_comments=false: keep loop comments generated by the LLVM backend
• keep_block_comments=false: keep code block comments generated by the LLVM backend
• keep_demoting_comments=false: keep variable demoting comments generated by the LLVM backend
• indent_calls=true: re-indent call instructions to be more recognizable and readable
• align_preds=8: align instruction guards (e.g. @p1) such that the instruction is placed at the n-th column.

cleanup_code(::Val{:sass}, code, dbinfo, cleanup_opts)

Cleanup SASS code.

If dbinfo, then location comments are kept.

Accepted cleanup_opts and their default values:

• demangle=true: demangle names within the code

cleanup_code(::Val{:gcn}, code, dbinfo, cleanup_opts)

Cleanup GCN code.

Accepted cleanup_opts and their default values:

• metadata=false: keep the metadata after the kernel's code
• kernel_metadata=false: keep the YAML amdhsa.kernels section. It is removed if metadata=false.
• demangle=true: demangle names within the code
• keep_loop_comments=false: keep loop comments generated by the LLVM backend
• keep_block_comments=false: keep code block comments generated by the LLVM backend
• keep_misc_comments=false: keep other comments generated by the LLVM backend
• align_operands=24: align the first operand of each instruction to the n-th column

cleanup_code(::Val{:spirv}, code, dbinfo, cleanup_opts)

Cleanup SPIRV code.

Accepted cleanup_opts and their default values:

• metadata=false: keep the meta operations around the main function's body
• demangle=true: demangle names within the code

replace_llvm_module_name(code::AbstractString)

Remove in code the trailing numbers in the LLVM module names, e.g. "julia_f_2007" => "f". This allows to remove false differences when comparing raw code, since each call to code_native (or code_llvm) triggers a new compilation using an unique LLVM module name, therefore each consecutive call is different even though the actual code does not change.

In Julia 1.11+, global variables names are also replaced with global_var_unique_gen_name_regex.

julia> f() = 1
f (generic function with 1 method)

julia> buf = IOBuffer();

julia> code_native(buf, f, Tuple{})  # Equivalent to `@code_native f()`

julia> code₁ = String(take!(buf));

julia> code_native(buf, f, Tuple{})

julia> code₂ = String(take!(buf));

julia> code₁ == code₂  # Different LLVM module names...
false

julia> replace_llvm_module_name(code₁) == replace_llvm_module_name(code₂)  # ...but same code
true

function_unique_gen_name_regex()
function_unique_gen_name_regex(function_name)

Regex matching all LLVM function names which might change from one compilation to another. As an example, in the outputs of @code_llvm below:

julia> f() = 1
f (generic function with 1 method)

julia> @code_llvm f()
...
define i64 @julia_f_855() #0 {
...

julia> @code_llvm f()
...
define i64 @julia_f_857() #0 {
...

the regex will match julia_f_855 and julia_f_857.

function_unique_gen_name_regex() should work for any function which does not have any characters in '",;- or spaces in its name. The function name is either in the capture group 1 or 2.

function_unique_gen_name_regex(function_name) should work with any generated name for the given function name.

It is 'globalUniqueGeneratedNames' in 'julia/src/codegen.cpp' which gives the unique number on the generated code. The regex matches most usages of this counter:

• from get_function_name

  • julia_<function_name>_<unique_num>
  • japi3_<function_name>_<unique_num>
  • japi1_<function_name>_<unique_num>

• from 'src/codegen.cpp#L4713'

  • j_<function_name>_<unique_num>
  • j1_<function_name>_<unique_num>

• from 'src/codegen.cpp#L6407'

  • jlcapi_<function_name>_<unique_num>

• from 'src/codegen.cpp#L7753'

  • jfptr_<function_name>_<unique_num>

• from 'src/codegen.cpp#L6185'

  • tojlinvoke<unique_num>

global_var_unique_gen_name_regex()
global_var_unique_gen_name_regex(global_name)

Regex matching all global variable names which might change from one compilation to another.

In LLVM IR, those variables are mentioned as such: @"+Core.GenericMemory#14067.jit". In native code, they look like this: ".L+Core.GenericMemory#13985.jit", with maybe some .set and .size sections at the end of the code (in x86 ASM).

global_var_unique_gen_name_regex() should work for any variable which does not have any characters in '",;- or spaces in its name.

global_var_unique_gen_name_regex(global_name) should work with any generated name for the given variable name.

It is 'globalUniqueGeneratedNames' in 'julia/src/codegen.cpp' which gives the unique number on the generated code. The regex matches only a single usage of this counter: in julia_pgv(ctx, cname, addr) at 'src/cgutils.cpp#L358' which is then added a ".jit" suffix in 'src/aotcompile.cpp#L2064' when doing code introspection.

demangle(name::AbstractString)

Demangle name into a C/C++ name using the demumble utility.

demangle_all(code::AbstractString)

Find and replace all mangled names in code with their demangled counterparts.

mangled_base_name(name::AbstractString)

Attempt to return the base name in the mangled function name. If it fails, nothing is returned.

clean_function_name(name_regex, code, replacement=nothing)

Replace occurences of name_regex in the code by replacement. replacement defaults to the demangled function name.

cleanup_inline_llvmcall_modules(c::Core.CodeInfo)
cleanup_inline_llvmcall_modules(c::Vector{Any})

Replace the LLVM-IR body of Base.llvmcall expressions in c to something more readable using an unescaped string, allowing to display the IR over multiple lines, with highlighting.

Only the LLVM function declaration is kept, other code (annotations, etc...) are stripped.