[Clang] implement -fno-eliminate-unused-debug-types

[nickdesaulniers]

Summary

Fixes pr/11710.
Signed-off-by: Nick Desaulniers <[e-mail]>

[nickdesaulniers]

• add support for structs, unions, and enums

[nickdesaulniers]

For C++, I'd imagine:

class foo {};
using my_foo = foo;
template <typename bar>
struct baz {
  bar my_bar;
};

but it seems that g++ doesn't emit debug info the for the templated struct.

[nickdesaulniers]

• add C++ using and class support

[nickdesaulniers]

CXXRecordDecl is a funny case, the AST looks like:

|-CXXRecordDecl 0x5c661a8 <clang/baz.cpp:1:1, col:12> col:7 class foo definition
| |-DefinitionData pass_in_registers empty aggregate standard_layout trivially_copyable pod trivial literal has_constexpr_non_copy_
| `-CXXRecordDecl 0x5c662d0 <col:1, col:7> col:7 implicit class foo
...

There's probably a bunch more C++ cases I'm not thinking of. Reading through gcc's test suite for -fno-eliminate-unused-debug-types, this feature doesn't have tests for anything with templates.

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5401–5403

The difference between using DebugInfo vs getModuleDebugInfo in this method is *killing* me. Same with return vs break.

[nickdesaulniers]

probably should add a test for enum class in C++

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5596

I think it makes sense at this point to rename EmitExplicitCastType to RetainType or EmitType. WDYT?

[nickdesaulniers]

• add enum class test case

[nickdesaulniers]

Bumping for review

[dblaikie]

clang/lib/CodeGen/CodeGenModule.cpp

5344

Since this is only for top-level decls, have you checked this works for types inside namespaces, local types in functions (well, that one might not be supported by GCC - so probably good to test/compare GCC's behavior here too), etc?

5400–5403

Perhaps this could be modified to use the same general purpose utility like the other type emissions (as speculated in another comment "EmitUnusedType" or the like)

5401–5403

Feel free to send separate patches to clean these things up.

5598

Rather than testing DebugUnusedType for every call - might be best to add a "EmitUnusedType" function that tests the flag and does the emission? (same way EmitExplicitCastType already checks for a certain level of debug info before emitting)

clang/lib/Frontend/CompilerInvocation.cpp

806

Could this be rolled into the debug-info-kind? (a kind beyond "standalone")

[nickdesaulniers]

clang/lib/Frontend/CompilerInvocation.cpp

806

That sounds like a good idea. Looking at the definition of DebugInfoKind (llvm/llvm-project/clang/include/clang/Basic/DebugInfoOptions.h), it seems that DebugInfoKind is an enum that defines a "level" of debug info to emit? Looking at the guard in CGDebugInfo::EmitExplicitCastType (llvm/llvm-project/clang/lib/CodeGen/CGDebugInfo.cpp), it calls CodeGenOptions::hasReducedDebugInfo() which does a comparison against a certain level. That seems to indicate the order of the enumerations is important. Do you have a suggestion what order I should add the new enum?

I'm guessing after LimitedDebugInfo or DebugInfoConstructor, but before FullDebugInfo? (I find the name of the method hasReducedDebugInfo confusing in that regard).

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5401–5403

https://reviews.llvm.org/D80840

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5598

It can be; what I don't like about that approach is that we have to determine the QualType here to pass in, at which point such function might just return without doing anything. (ie. we have to do slightly more work in the case where the debugging of unused types was *not* requested). But that's a minor nit and we can live with it?

clang/lib/Frontend/CompilerInvocation.cpp

806

Ok, so I have a diff that implements this approach. I feel like I should maybe publish it as a child commit to this one, to be able to more easily discuss it?

Two problems I run into:

1. as alluded to in my previous response in this thread, DebugInfoKind is an enum that specifies a level. It tends to "drag" other debug flags in based on the ordering. Looking at extending the switch in CGDebugInfo::CreateCompileUnit (clang/lib/CodeGen/CGDebugInfo.cpp), it's not at all clear to me which we existing case should we choose?
2. we want the flag -fno-eliminate-unused-debug-types to match GCC for compatibility. We can additionally add a new debug info kind like "standalone" (clang/lib/Frontend/CompilerInvocation.cpp), but it's not clear how the two flags together should interact.

The suggestion for a new debug info kind feels like a recommendation to add a new "level" of debug info, but -fno-eliminate-unused-debug-types feels like it should be mutually exclusive of debug info kind? (I guess GCC does *require* -g with -fno-eliminate-unused-debug-types)

@dblaikie maybe you have more recommendations on this?

[dblaikie]

clang/lib/CodeGen/CodeGenModule.cpp

5596

Yep - now that I understand the confusingly named hasReducedDebugInfo

"RetainType" sounds good to me, maybe "EmitAndRetainType" might be worthwhile for the extra clarity/symmetry with other "Emit" functions.

clang/lib/Frontend/CompilerInvocation.cpp

806

This value would probably go "after" "full" (full isn't full enough, as you've found - you need a mode that's even "fullerer")

Perhaps renaming "Full" to "Referenced" and then introducing this new kind as the new "Full" or maybe under a somewhat different name to avoid ambiguity.

Any suggestions on a less confusing name for "hasReducedDebugInfo"? (I think it was basically "Has less than full debug info"... but, yep, it doesn't do that at all - it's "HasTypeAndVariableDebugInfo" by the looks of it/by the comment)

[dblaikie]

clang/lib/CodeGen/CodeGenModule.cpp

5598

I don't believe getting the QualType from a TypeDecl is an expensive operation/enough to try to avoid it here. If it is I'd rather have a "EmitUnusedType(TypeDecl*)" function, and then the conditional QualType retrieval could be in there, written once rather than several times.

clang/lib/Frontend/CompilerInvocation.cpp

806

Ok, so I have a diff that implements this approach. I feel like I should maybe publish it as a child commit to this one, to be able to more easily discuss it?

Two problems I run into:

1. as alluded to in my previous response in this thread, DebugInfoKind is an enum that specifies a level. It tends to "drag" other debug flags in based on the ordering. Looking at extending the switch in CGDebugInfo::CreateCompileUnit (clang/lib/CodeGen/CGDebugInfo.cpp), it's not at all clear to me which we existing case should we choose?
2. we want the flag -fno-eliminate-unused-debug-types to match GCC for compatibility. We can additionally add a new debug info kind like "standalone" (clang/lib/Frontend/CompilerInvocation.cpp), but it's not clear how the two flags together should interact.

I'm not suggesting adding a new driver-level flag for this, but implementing the GCC flag name in terms of the debug-info-kind. Probably by having "-fno-eliminate-unused-debug-types" override "-fstandalone-debug" - whichever comes later wins. Because they are part of a progression. Though admittedly that might get a smidge confusing about exactly whit no/yes versions of these two flags override each other - but I think that's inevitable confusion with the nature of these flags.

What does GCC do for its -f[no-]emit-class-debug-always (which is somewhat similar to -fstandalone-debug) V -f[no-]eliminate-unused-debug-types? I'm not sure we'll want to emulate the behavior exxactly, but it's probably a good place to start to see if there's an existing model that looks OK here.

The suggestion for a new debug info kind feels like a recommendation to add a new "level" of debug info, but -fno-eliminate-unused-debug-types feels like it should be mutually exclusive of debug info kind? (I guess GCC does *require* -g with -fno-eliminate-unused-debug-types)

Sorry, I'm not following here - perhaps you could explain in different words? At the driver/user level, "debug info kind" doesn't exist - there's flags like "-g", "-gmlt", "-fstandalone-debug", etc. That are mapped to debug-info-kind on the -cc1 command line. I'm suggesting "-fno-eliminate-unused-debug-types" is another of those driver flags that is taken into account when mapping down to the cc1 "debug info kind".

[MaskRay]

clang/include/clang/Basic/CodeGenOptions.def

291

This flag is now more than "typedefs"

clang/include/clang/Driver/Options.td

3325–3326

This group is for ignored options. Add this near other f options.

I fixed the FIXME about BooleanFFlag with TableGen special identifier NAME. The way we define boolean options still look verbose to me, so I just created D80883.

[probinson]

In D80242#2045362, @nickdesaulniers wrote:

For C++, I'd imagine:

class foo {};
using my_foo = foo;
template <typename bar>
struct baz {
  bar my_bar;
};

but it seems that g++ doesn't emit debug info the for the templated struct.

There's no debug info for templates, only for instantiations of templates. HTH.

[nickdesaulniers]

• rebase on 1b6d29e06b07e518025b6f06445ad3275d6f5684 and D80840 (no other changes yet)

[nickdesaulniers]

• missed one part of the rebase

[nickdesaulniers]

clang/lib/Frontend/CompilerInvocation.cpp

806

Are debug info kinds a progression or mutually exclusive? I think they're a progression, but I have concerns with modeling -fno-eliminate-unused-debug-types as part of that progression.

I don't think we should consider -fno-eliminate-unused-debug-types to be part of the progression of debug info levels via debug info kind. Users must specify -g in addition to -fno-eliminate-unused-debug-types to get any debug info, at which point they'll get both full debug info via -g and unused type debug info via -fno-eliminate-unused-debug-types.

To me, -fno-eliminate-unused-debug-types and -fno-eliminate-unused-debug-symbols seem orthogonal to debug info level. For example, how would you order -fno-eliminate-unused-debug-types vs -fno-eliminate-unused-debug-symbols in a progression? One does not imply the other; they are mutually exclusive.

I agree in a progression -fno-eliminate-unused-debug-types should come after full debug info, but then it seems weird for -fno-eliminate-unused-debug-symbols if we ever wanted to add support for that (YAGNI?) So maybe a progression isn't the best representation?

---

I should probably fix the documentation in this commit for -fstandalone-debug to mention this, and add documentation for this flag in the first place.

---

Regarding -femit-class-debug-always g++ with -femit-class-debug-always for the test case in this CL doesn't produce any debug info. Looks like clang does not currently support -femit-class-debug-always.

[nickdesaulniers]

• add documentation

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5344

class declared in namespace works with this patch. I'll add a test.

local type in function, ex.

void foo(void) {
  struct bar {};
}

is not in the current diff, but is supported by gcc. Let me fix that and add a test. Great suggestion.

[nickdesaulniers]

• handle non top level decls, add tests

[nickdesaulniers]

Bumping for review. Are there additional reviewers we can entrust code review to, to help share the code review responsibility?

[probinson]

+debug-info tag

Needs a Driver test to show that the Right Thing gets passed to cc1.

clang/docs/CommandGuide/clang.rst

438

The change to the command-line reference says "defined" not "declared", these should be consistent.

clang/docs/UsersManual.rst

2357

"declared" vs "defined" again.

clang/lib/Driver/ToolChains/Clang.cpp

5190

Looks like this makes the "no" option take precedence, even if it comes before "yes" option. Should check for both in one call, so you really do take the last one:

Args.AddLastArg(CmdArgs, options::OPT_feliminate_unused_debug_types, options::OPT_fno_eliminate_unused_debug_types);

[nickdesaulniers]

• rebase
• fix docs (use declare consistently, avoid define)
• add driver test
• fix command line check

[nickdesaulniers]

• update comment to CODEGENOPT
• avoid BooleanFFlag group

[dblaikie]

• fix docs (use declare consistently, avoid define)

Defined would be the right word here (declaring a type is code like "struct foo;" - and this change probably doesn't/shouldn't emit debug info for type declarations, yes?)

clang/include/clang/Driver/Options.td

1541–1543

I /think/ @MaskRay added some kind of helper to make it easier to define the f/fno pair together, perhaps? Ah, right, mentioned in another part of this review, D80883 - though that'll only be applicable if this doesn't become a debug-info-kind.

clang/lib/CodeGen/CodeGenModule.cpp

5400–5403

Ping on this ^

5596

Ping on this ^

5598

Ping on this^

clang/lib/Frontend/CompilerInvocation.cpp

806

Are debug info kinds a progression or mutually exclusive?

Both? But yes, generally considered a progression.

I don't think we should consider -fno-eliminate-unused-debug-types to be part of the progression of debug info levels via debug info kind. Users must specify -g in addition to -fno-eliminate-unused-debug-types to get any debug info, at which point they'll get both full debug info via -g and unused type debug info via -fno-eliminate-unused-debug-types.

The same is true of the existing -fstandalone-debug (similar to GCC's -femit-class-debug-always)

To me, -fno-eliminate-unused-debug-types and -fno-eliminate-unused-debug-symbols seem orthogonal to debug info level. For example, how would you order -fno-eliminate-unused-debug-types vs -fno-eliminate-unused-debug-symbols in a progression? One does not imply the other; they are mutually exclusive.

I'd probably implement -fno-eliminate-unused-debug-symbols as a separate thing, not in that progression & would be happy to design the current feature without considering that future possibility in any case.

I should probably fix the documentation in this commit for -fstandalone-debug to mention this, and add documentation for this flag in the first place.

Mention what, sorry?

Regarding -femit-class-debug-always g++ with -femit-class-debug-always for the test case in this CL doesn't produce any debug info.

Sorry, more specifically I meant "-g -femit-class-debug-always" and "-g -fno-eliminate-unused-debug-types" and I meant on more than only this test case.

Basically: how do those two flags interact in GCC? Are they both considered separately (in the no-X and X variants) and then the "more verbose" of the two wins (ie: -femit-class-debug-always -fno-eliminate-unused-debug-types -feliminate-unused-debug-types still produces "Class debug always" behavior?)?

Looks like clang does not currently support -femit-class-debug-always.

clang/test/CodeGen/debug-info-unused-types.c

21–30

Might be a smidge easier to read without the enumerators (& this doesn't check the enumerators are in the "elements" list of the enumeration anyway) - presumably the type emission codepath is well enough tested that we only need to to briefly test that the types were emitted at all? (might even be easy enough to stop after the tag and name))

36–45

Could this be simplified down to:

NODBG-NOT: DICompositeType

?

clang/test/CodeGen/debug-info-unused-types.cpp

16–30

Similar feedback as for the C example above

[nickdesaulniers]

In D80242#2093785, @dblaikie wrote:

• fix docs (use declare consistently, avoid define)

Defined would be the right word here (declaring a type is code like "struct foo;" - and this change probably doesn't/shouldn't emit debug info for type declarations, yes?)

Oh, looks like that's a difference between my current implementation and GCC. GCC does not emit debug info with -g -fno-eliminate-unused-debug-types for type declarations; the current version of my patch does. Let me fix that+add tests+update docs.

[nickdesaulniers]

Ok, I think I have all feedback addresses locally. One issue I'm hitting is that I'm regressing clang/test/OpenMP/parallel_codegen.cpp. Specifically, it seems it has code like:

template <typename T>                                                                                                                                                                                                                                               
int tmain(T argc) {                                                                                                                                                                                                                                                 
  typedef double (*chunk_t)[argc[0][0]];
...

I'm not sure what type chunk_t is exactly, but my brain can't parse it. Function pointer that returns double (no parameter list?) array? WTF

[nickdesaulniers]

• rebase
• make feature progression of debug-info-kind
• fix docs (define, not declare)
• add driver tests
• use OptOutFFlag tablegen helper
• rename helper to EmitAndRetainType
• dont emit debug info for declarations
• rewrite tests

[nickdesaulniers]

• git clang-format HEAD~

[dblaikie]

clang/lib/CodeGen/CGDecl.cpp

102–115

All of these might be able to be collapsed together - using "cast<TagDecl>(D).hasDefinition()" ... to have a common implementation.

104–106

Any particular reason this one's handled differently from the others?

111

I think the right thing to call here (& the other places in this patch) is "hasDefinition" rather than "getDefinition" (this will avoid the definition being deserialized when it's not needed yet (then if we don't end up emitting the type because the flag hasn't been given, it won't be deserialized unnecessarily))

clang/lib/CodeGen/CodeGenModule.cpp

5404

I think this might not work as intended if the type trips over the other class deduplication optimizations (try manually testing with a type that has an external VTable?) - and perhaps this codepath should use the EmitAndRetain functionality.

[nickdesaulniers]

• make cases more uniform, cleanup some missed fixes

clang/lib/CodeGen/CGDecl.cpp

102–115

TagDecl doesn't have a method hasDefinition. Only CXXRecordDecl does.

111

TagDecl doesn't have a method hasDefinition. Only CXXRecordDecl does.

clang/lib/CodeGen/CodeGenModule.cpp

5404

completeUnusedClass eventually calls into CGDebugInfo::completeClass which makes use of the TypeCache. What does it mean for a class to have "an external vtable?" Can you give me an example?

[nickdesaulniers]

clang/lib/CodeGen/CGDecl.cpp

102–115

Also, how to get the QualType differs between Decl::Enum an Decl::Record; getEnumType vs getRecordType respectively.

[nickdesaulniers]

• fix clang/test/OpenMP/parallel_codegen.cpp

[nickdesaulniers]

Bumping for review.

[nickdesaulniers]

bumping for review

[probinson]

clang/docs/ClangCommandLineReference.rst

2159

I think this description goes counter to how other options are described, which basically is about the positive form of the option. If that's the case, this wants to be something more like "Suppress (or emit) debug info ..."

clang/include/clang/Driver/Options.td

947

The positive form of the option suppresses info, so "Emit " and "Do not emit " are reversed here.

[nickdesaulniers]

• fix documentation double negation

[nickdesaulniers]

Bumping for review

[dblaikie]

clang/lib/CodeGen/CGDecl.cpp

102–115

Looks like ASTContext::getTypeDeclType could be used, if it makes things especially tidier. But perhaps the TagDecl/hasDefinition difference is enough for it not to be especially significant.

111

Ah, fair enough!

clang/lib/CodeGen/CodeGenModule.cpp

5404

Here's some sample code that uses a type in the DWARF but due to the vtable being external (an external vtable is one that's guaranteed to be emitted in some other object file - due to the key function optimization in the Itanium ABI - since all virtual functions must be defined if a type is ODR used, the compiler can make a shared assumption that the vtable will only be emitted with a single consistently chosen virtual function (ideally: the first non-inline one) and emit the vtable only in that object and nowhere else) the type is emitted as a declaration (when using "-fno-standalone-debug"):

struct t1 {
  virtual void f1();
};
t1 v1;

clang/lib/Driver/ToolChains/Clang.cpp

3834–3837

How does GCC compose this with -gmlt, for instance? I'd suspect that the desired behavior might be for -gmlt to override this -f flag? So maybe this should be predicated on "if (EmitDwarf && DebugInfoKind >= codegenoptions::LimitedDebugInfo && ... "?

(so if someone wants to use only -gmlt in certain parts of their build that doesn't get upgraded by the presence of this -f flag)

clang/test/CodeGen/debug-info-unused-types.c

47–61

Maybe give these more specific names (similarly in the other test) - and/or you could use a pattern in the naming that might make the -NOT lines easier?

eg:

struct decl_struct;
enum decl_enum;
...
// NODBG-NOT: name: "decl_

(also best to include a bit more context in the -NOT checks, otherwise there's risk that a users build directory (even with these longer names) or other paths might accidentally fail the checks)

clang/test/CodeGen/debug-info-unused-types.cpp

16–22

Probably good to check these end up in the retained types list

24–30

With a uniform naming scheme, perhaps this could be // NODBG-NOT: name: "unused_ ?

[nickdesaulniers]

• rebase on 227db86a1b7dd, update NOT checks

[nickdesaulniers]

clang/lib/CodeGen/CodeGenModule.cpp

5404

Cool, thanks for the example. For your example:
clang -g -emit-llvm -S foo.cpp -o -:

...
DIGlobalVariable(name: "v1"
...
DICompositeType(tag: DW_TAG_structure_type, name: "t1"
...

clang -g -fno-eliminate-unused-debug-types -emit-llvm -S foo.cpp -o -:

...
DIGlobalVariable(name: "v1"
...
DICompositeType(tag: DW_TAG_structure_type, name: "t1"
...
DIDerivedType(tag: DW_TAG_member, name: "_vptr$t1"
DIDerivedType(tag: DW_TAG_pointer_type
DIDerivedType(tag: DW_TAG_pointer_type, name: "__vtbl_ptr_type"
DIBasicType(name: "int"
DISubprogram(name: "f1"
DISubroutineType
DIDerivedType(tag: DW_TAG_pointer_type
...

So even though f1 was not defined, we still emit debug info for it.

Comparing the output of llvm-dwarfdump on $CC -g -fno-eliminate-unused-debug-types -c <example.cpp> for g++ vs clang++; we're definitely producing more debug info (the vtable info).

That said, the creation of v1 technically means that t1 is no longer an "unused type." If we comment out its creation, then compare $CC -g -fno-eliminate-unused-debug-types -c <example.cpp> for g++ vs clang++, g++ produces no debug info (That seems like a bug in g++ to me), while clang++ with this patch produces the debug info for the CXXRecord, and additional info for the vtable ptr and signature of f1.

I'm not sure what's expected in this case (we should at least emit the DW_TAG_structure_type for t1, but I'm not sure about the DW_TAG_subprogram for f1. I assume that's because we've enabled more-than-full-debuginfo behavior? WDYT?

clang/lib/Driver/ToolChains/Clang.cpp

3834–3837

GCC does not yet support -gmlt.
https://godbolt.org/z/Wrv6sK

IIUC, -gmlt ("minimum line tables") is meant to minimize the amount of debug info produced. I'm not sure what the user expects if they simultaneous ask for more debug info (via -fno-eliminate-unused-types) and less debug info (via -gmlt).

The current behavior is that -fno-eliminate-unused-types "wins" out over -gmlt, resulting in the "fullest" amount of debug info being produced. (Would you like me to add tests for this behavior to clang/test/Driver/debug-options.c for this behavior?)

If we don't want that behavior, we should reconsider whether we want UnusedTypeInfo to be part of the DebugInfoKind "progression." I also don't understand what the suggested added condition to the predicate is supposed to do; if EmitDebugInfo is set, doesn't the default value of DebugInfoConstructor (after 227db86a1b7dd6f96f7df14890fcd071bc4fe1f5, rebased this patch on top of) mean that we wouldn't use UnusedTypeInfo with `-g -fno-eliminate-unused-type-info? Or am I misunderstanding the suggestion?

clang/test/CodeGen/debug-info-unused-types.cpp

16–22

What's the best way to check this? In particular, I worry about the !<number> number changing in the future, resulting in this test spuriously failing. For this test case, we have:

!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, producer: "Nick Desaulniers clang version 12.0.0 ([e-mail]:llvm/llvm-project.git e541558d11ca82b3664b907b350da5187f23c0c9)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, retainedTypes: !15, splitDebugInlining: false, nameTableKind: None)

which tells us the retainedTypes is !15. !15 looks like:

!15 = !{!16, !17, !3, !5, !18, !8}

In this case, should I just add the numbers? If we used regex patterns [0-9], I worry that we couldn't ensure the list would match. Is it ok to just hardcode these?

[dblaikie]

clang/lib/CodeGen/CodeGenModule.cpp

5404

I think it'd be nice to handle this the same as GCC (though technically we don't offer the same customization that GCC offers - I imagine the GCC behavior of -femit-class-debug-always, which Clang doesn't have a configuration option for (well, not specifically - again, comes back to that "independent configuration" or "overriding behavior" implementation choice) - I'd sort of think of it like that flag is always enabled)

But I guess if we think of it as a series of escalations, then, yeah, it makes sense in the Clang way of thinking of it, as -fno-eliminate-unused-debug-types as overriding the default -fno-emit-class-debug-always... *shrug* don't feel /super/ strongly either way I guess. If the current implementation is such that -fno-eliminate-unused-debug-types overrides all other type homing techniques (vtable based, explicit template specialization decl/def, ctor based, and not-required-to-be-complete (define a type but only use pouinters to that type and never dereference them)) that's fine by me.

clang/lib/Driver/ToolChains/Clang.cpp

3834–3837

GCC's nearest equivalent is -g1 by the looks of it - I'd hazard a guess that -fno-eliminate-unused-types has no effect when compiling with -g1 (as -fstandalone-debug has no effect when -gmlt is specified)

I thin it's probably appropriate to implement similar behavior in Clang - that -fstandalone-debug/no-standalone-debug/no-eliminate-unused-types only applies to users that already asked for some kind of class debug info. It looks like that functionality's already implemented correctly (at least in one or two small cases I tested) for -fstandalone-debug and -fno-eliminate-unused-types should follow that.

What should happen between -fstandalone-debug and -fno-eliminate-unused-types is an open question I don't feel too strongly about which order they work in.

clang/test/CodeGen/debug-info-unused-types.cpp

16–22

You can use captures in FileCheck to make references a little more flexible. Surprised there aren't /any/ tests in LLVM testing the retainedTypes list... test/CodeGen/debug-info-extern-call.c used to use the retained types list & still has some testing remnants of it you could draw from, something like::

// CHECK: !DICompileUnit({{.*}}retainedTypes: [[RETTYPES:![0-9]+]]
// CHECK: [[RETTYPES]] = !{[[T1:![0-9]+]], [[T2:![0-9]+]], ...}
...
// CHECK: [[T1]] = !DICompositeType(...

It doesn't protect the test from changes in the order of the retained types list but otherwise resilient to extraneous metadata nodes being added or removed.

[nickdesaulniers]

• rebase, add -g1 test case, don't emit unless -g2 or higher

clang/lib/Driver/ToolChains/Clang.cpp

3834–3837

Oh! Indeed!

gcc -g1 -fno-eliminate-unused-debug-types foo.c -c -o foo.o

produces no additional debug info. -g2 does. Let me fix that and add tests.

clang/test/CodeGen/debug-info-unused-types.cpp

16–22

So one issue I'm running into with this approach is that the retained type list is emitted in between the other debug nodes (some before it, some after it). The capture logic seems to have an issue referring to captures that have yet to be defined yet.

note: uses undefined variable(s): "TYPE7"

but if I move the CHECK for the retained types last, then FileCheck errors because it seems it only doesn't one pass for the CHECK statements (IIUC), so it's "too late" to parse.

Is there another way to test this, or am I holding it wrong?

[dblaikie]

clang/lib/Driver/ToolChains/Clang.cpp

3834–3838

Would this be tidier if it were rolled into the if/checking around 380 since it's a very similar option?

clang/test/CodeGen/debug-info-unused-types.cpp

16–22

Order dependence is a currently unavoidable part of FileCheck tests, so if the content you're trying to test (I'm not entirely clear from your description - so guessing a bit) looks something like:

!1 = DICompileUnit ... retainedTypes: !3
!2 = DICompositeType "x"
!3 = !{!2, !4}
!4 = DICompositeType "y"

Then you'd CHECK this something like:

; CHECK: DICompileUnit {{.*}} retainedTypes: [[RET_TY:![0-9]*]]
; CHECK: [[X:![0-9]*]] = DICompositeType "x"
; CHECK: [[RET_TY]] = !{[[X]], [[Y:![0-9]*]]}
; CHECK: [[Y]] = DICompositeType "y"

Yes, this does mean that if the types were emitted in a different order, the test would fail - but that's a limitation we're living with for pretty much all LLVM and Clang testing at the moment.

25–31

Maybe simpler to test the NODBG by NODBG-NOT: DI{{[a-zA-Z]*}}Type ? Not sure if that'd quite work, but might be adequate.

[nickdesaulniers]

• rebase, add captures to tests, simplify NODBG-NOT cases

[nickdesaulniers]

clang/lib/Driver/ToolChains/Clang.cpp

3834–3838

Sorry, I recently rebased, is 380 still where you were thinking...or...? (Maybe you can add some code context in case it moves again)? Maybe 490 (DebugLevelToInfoKind)? Otherwise I don't really see other patterns that check EmitDwarf.

clang/test/CodeGen/debug-info-unused-types.cpp

25–31

DISubroutineType unfortunately would match. !DI{{CompositeType|Enumerator|DerivedType}} does work though!

[nickdesaulniers]

bumping for review

[dblaikie]

clang/lib/Driver/ToolChains/Clang.cpp

3786–3794

Re: https://reviews.llvm.org/D80242#inline-783632

This is the spot that seems like the place to handle UnusedTypeInfo, with something like:

if (DebugInfoKind == codegenoptions::LimitedDebugInfo) {
  if (Args.hasFlag(options::OPT_fno_eliminate_unused_debug_types,
                   options::OPT_feliminate_unused_debug_types, false))
    DebugInfoKind = codegenoptions::UnusedTypeInfo;
 else if (NeedFullDebug)
    DebugInfoKind = codegenoptions::FullDebugInfo;
}

clang/test/CodeGen/debug-info-unused-types.cpp

25–31

Sweet - thanks!

[nickdesaulniers]

• rebase, centralize DebugInfoKind update as per @dblaikie

[dblaikie]

Looks good - thanks!

[nickdesaulniers]

• git clang-format

[nickdesaulniers]

build failure on windows: http://lab.llvm.org:8011/builders/llvm-clang-win-x-aarch64/builds/1886/steps/test-check-clang/logs/FAIL%3A%20Clang%3A%3Adebug-info-unused-types.cpp

[nickdesaulniers]

fix pushed in cbd8ec93709376fbf404c99f4eee399790e26db7

[nickdesaulniers]

Looks like another related failure for http://green.lab.llvm.org/green/job/clang-stage1-RA/13460/consoleFull#-42777206a1ca8a51-895e-46c6-af87-ce24fa4cd561 ?

[nickdesaulniers]

So it looks like %clang++ should be %clangxx for windows, then for OSX the default value of debug-info-kind is different, while the tests also seem to fail there.

[nickdesaulniers]

• rebase, move Codegen tests to _cc1, move -g1 test to driver test, use more regex for ![0-9]+, manually test --target=x86_64-apple-macosx10.15.0 which has a different default debug info kind.

[nickdesaulniers]

• squash and reupload

[nickdesaulniers]

Hi @dblaikie , sorry for messing up the upload, I think it's good now, but to see the latest change, you'll need to look at the diff between the 2 prior versions (https://reviews.llvm.org/D80242?vs=284053&id=284477#toc).

I reverted this on Friday. It looks like it failed the windows tests because %clang++ is not a thing on windows. I moved the two tests under clang/tests/Codegen to use %clang_cc1 instead and moved the -g1 test to the clang/test/Driver/ test.

This also failed osx builds, which default to --target=x86_64-apple-macosx10.15.0. It looks like the default -debug-info-kind on that target is standalone, not limited. I made the NO_DEBUG_UNUSED_TYPES-NOT use a regex for other values for this to be, since we only really care that it's not unused-types. I manually tested all added test cases with --target=x86_64-apple-macosx10.15.0 or --triple=x86_64-apple-macosx10.15.0 added. I also used your suggestions from here to use more general regex for two of the debug info nodes.

Please take a look.

[dblaikie]

clang/test/Driver/debug-options.c

371–373

Why does this test -g1 in particular? (that, I think, would always be line-tables-only, on all Clang platforms?) Rather than -g like the positive test?

[nickdesaulniers]

clang/test/Driver/debug-options.c

371–373

From the previous diff (https://lists.llvm.org/pipermail/llvm-dev/2020-August/144082.html) we added a test to clang/test/CodeGen/ tests for -fno... and -g1 together. Since I moved those to %clang_cc1, I moved the -fno... + -g1 tests here.

[dblaikie]

Looks good, thanks!

clang/test/Driver/debug-options.c

371–373

Ah, right right (link might be incorrect though - so including the one where it was discussed: https://reviews.llvm.org/D80242#inline-774452 ) - thanks!

[nickdesaulniers]

clang/test/Driver/debug-options.c

371–373

LOL, yeah sorry, thanks!

clang/docs/ClangCommandLineReference.rst

 .. option:: -fstack-size-section, -fno-stack-size-section
 
 Emit section containing metadata on function stack sizes
 
 .. option:: -fstandalone-debug, -fno-limit-debug-info, -fno-standalone-debug
 
 Emit full debug info for all types used by the program
 
+.. option:: -feliminate-unused-debug-types, -fno-eliminate-unused-debug-types
+
+Suppress (or emit) debug info for types that are unused but defined by the program.
+
 .. option:: -fstrict-aliasing, -fno-strict-aliasing
 
 .. option:: -fstrict-enums, -fno-strict-enums
 
 Enable optimizations based on the strict definition of an enum's value range
 
 .. option:: -fstrict-float-cast-overflow, -fno-strict-float-cast-overflow
 

clang/docs/CommandGuide/clang.rst

.. option:: -fstandalone-debug -fno-standalone-debug
   module that contains the vtable for the class.
 
   The :option:`-fstandalone-debug` option turns off these optimizations.
   This is useful when working with 3rd-party libraries that don't come with
   debug information.  This is the default on Darwin.  Note that Clang will
   never emit type information for types that are not referenced at all by the
   program.
 
+.. option:: -feliminate-unused-debug-types
+
+  By default, Clang does not emit type information for types that are defined
+  but not used in a program. To retain the debug info for these unused types,
+  the negation **-fno-eliminate-unused-debug-types** can be used.
+
 .. option:: -fexceptions
 
   Enable generation of unwind information. This allows exceptions to be thrown
   through Clang compiled stack frames.  This is on by default in x86-64.
 
 .. option:: -ftrapv
 
   Generate code to catch integer overflow errors.  Signed integer overflow is

clang/docs/UsersManual.rst

.. option:: -fno-standalone-debug
    On Darwin **-fstandalone-debug** is enabled by default. The
    **-fno-standalone-debug** option can be used to get to turn on the
    vtable-based optimization described above.
 
 .. option:: -g
 
   Generate complete debug info.
 
+.. option:: -feliminate-unused-debug-types
+
+  By default, Clang does not emit type information for types that are defined
+  but not used in a program. To retain the debug info for these unused types,
+  the negation **-fno-eliminate-unused-debug-types** can be used.
+
 Controlling Macro Debug Info Generation
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Debug info for C preprocessor macros increases the size of debug information in
 the binary. Macro debug info generated by Clang can be controlled by the flags
 listed below.
 
 .. option:: -fdebug-macro

clang/include/clang/Basic/CodeGenOptions.h

#include "clang/Basic/CodeGenOptions.def"
 
   /// Check if CSIR profile use is on.
   bool hasProfileCSIRUse() const { return getProfileUse() == ProfileCSIRInstr; }
 
   /// Check if type and variable info should be emitted.
   bool hasReducedDebugInfo() const {
     return getDebugInfo() >= codegenoptions::DebugInfoConstructor;
   }
+
+  /// Check if maybe unused type info should be emitted.
+  bool hasMaybeUnusedDebugInfo() const {
+    return getDebugInfo() >= codegenoptions::UnusedTypeInfo;
+  }
 };
 
 }  // end namespace clang
 
 #endif

clang/include/clang/Basic/CodeGenOptions.def

 
 CODEGENOPT(SplitDwarfInlining, 1, 1) ///< Whether to include inlining info in the
                                      ///< skeleton CU to allow for symbolication
                                      ///< of inline stack frames without .dwo files.
 CODEGENOPT(DebugFwdTemplateParams, 1, 0) ///< Whether to emit complete
                                          ///< template parameter descriptions in
                                          ///< forward declarations (versus just
                                          ///< including them in the name).
-
 CODEGENOPT(EmitLLVMUseLists, 1, 0) ///< Control whether to serialize use-lists.
 
 CODEGENOPT(WholeProgramVTables, 1, 0) ///< Whether to apply whole-program
                                       ///  vtable optimization.
 
 CODEGENOPT(VirtualFunctionElimination, 1, 0) ///< Whether to apply the dead
                                              /// virtual function elimination
                                              /// optimization.
 
 /// Whether to use public LTO visibility for entities in std and stdext
 /// namespaces. This is enabled by clang-cl's /MT and /MTd flags.
 CODEGENOPT(LTOVisibilityPublicStd, 1, 0)
 
 /// The user specified number of registers to be used for integral arguments,
 /// or 0 if unspecified.
 VALUE_CODEGENOPT(NumRegisterParameters, 32, 0)
 
 /// The threshold to put data into small data section.
 VALUE_CODEGENOPT(SmallDataLimit, 32, 0)
 
 /// The lower bound for a buffer to be considered for stack protection.
 VALUE_CODEGENOPT(SSPBufferSize, 32, 0)
 
 /// The kind of generated debug info.
-ENUM_CODEGENOPT(DebugInfo, codegenoptions::DebugInfoKind, 3, codegenoptions::NoDebugInfo)
+ENUM_CODEGENOPT(DebugInfo, codegenoptions::DebugInfoKind, 4, codegenoptions::NoDebugInfo)
 
 /// Whether to generate macro debug info.
 CODEGENOPT(MacroDebugInfo, 1, 0)
 
 /// Tune the debug info for this debugger.
 ENUM_CODEGENOPT(DebuggerTuning, llvm::DebuggerKind, 2,
                 llvm::DebuggerKind::Default)
 

clang/include/clang/Basic/DebugInfoOptions.h

enum DebugInfoKind {
 
   /// Limit generated debug info to reduce size (-fno-standalone-debug). This
   /// emits forward decls for types that could be replaced with forward decls in
   /// the source code. For dynamic C++ classes type info is only emitted into
   /// the module that contains the classe's vtable.
   LimitedDebugInfo,
 
   /// Generate complete debug info.
-  FullDebugInfo
+  FullDebugInfo,
+
+  /// Generate debug info for types that may be unused in the source
+  /// (-fno-eliminate-unused-debug-types).
+  UnusedTypeInfo,
 };
 
 } // end namespace codegenoptions
 } // end namespace clang
 
 #endif

clang/include/clang/Driver/Options.td

 def fdwarf2_cfi_asm : Flag<["-"], "fdwarf2-cfi-asm">, Group<clang_ignored_f_Group>;
 def fno_dwarf2_cfi_asm : Flag<["-"], "fno-dwarf2-cfi-asm">, Group<clang_ignored_f_Group>;
 defm dwarf_directory_asm : OptOutFFlag<"dwarf-directory-asm", "", "">;
 def felide_constructors : Flag<["-"], "felide-constructors">, Group<f_Group>;
 def fno_elide_type : Flag<["-"], "fno-elide-type">, Group<f_Group>,
     Flags<[CC1Option]>,
     HelpText<"Do not elide types when printing diagnostics">;
 def feliminate_unused_debug_symbols : Flag<["-"], "feliminate-unused-debug-symbols">, Group<f_Group>;
+defm eliminate_unused_debug_types : OptOutFFlag<"eliminate-unused-debug-types",
+  "Do not emit ", "Emit ", " debug info for defined but unused types">;
 def femit_all_decls : Flag<["-"], "femit-all-decls">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Emit all declarations, even if unused">;
 def femulated_tls : Flag<["-"], "femulated-tls">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Use emutls functions to access thread_local variables">;
 def fno_emulated_tls : Flag<["-"], "fno-emulated-tls">, Group<f_Group>, Flags<[CC1Option]>;
 def fencoding_EQ : Joined<["-"], "fencoding=">, Group<f_Group>;
 def ferror_limit_EQ : Joined<["-"], "ferror-limit=">, Group<f_Group>, Flags<[CoreOption]>;
 defm exceptions : OptInFFlag<"exceptions", "Enable", "Disable", " support for exception handling">;
def fno_digraphs : Flag<["-"], "fno-digraphs">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Disallow alternative token representations '<:', ':>', '<%', '%>', '%:', '%:%:'">;
 def fno_declspec : Flag<["-"], "fno-declspec">, Group<f_clang_Group>,
   HelpText<"Disallow __declspec as a keyword">, Flags<[CC1Option]>;
 def fno_dollars_in_identifiers : Flag<["-"], "fno-dollars-in-identifiers">, Group<f_Group>,
   HelpText<"Disallow '$' in identifiers">, Flags<[CC1Option]>;
 def fno_elide_constructors : Flag<["-"], "fno-elide-constructors">, Group<f_Group>,
   HelpText<"Disable C++ copy constructor elision">, Flags<[CC1Option]>;
 def fno_eliminate_unused_debug_symbols : Flag<["-"], "fno-eliminate-unused-debug-symbols">, Group<f_Group>;
 def fno_gnu_keywords : Flag<["-"], "fno-gnu-keywords">, Group<f_Group>, Flags<[CC1Option]>;
 def fno_inline_functions : Flag<["-"], "fno-inline-functions">, Group<f_clang_Group>, Flags<[CC1Option]>;
 def fno_inline : Flag<["-"], "fno-inline">, Group<f_clang_Group>, Flags<[CC1Option]>;
 def fno_global_isel : Flag<["-"], "fno-global-isel">, Group<f_clang_Group>,
   HelpText<"Disables the global instruction selector">;
 def fno_experimental_isel : Flag<["-"], "fno-experimental-isel">, Group<f_clang_Group>,
   Alias<fno_global_isel>;
 def fno_experimental_new_pass_manager : Flag<["-"], "fno-experimental-new-pass-manager">,
   Group<f_clang_Group>, Flags<[CC1Option]>,
   HelpText<"Disables an experimental new pass manager in LLVM.">;
 def fveclib : Joined<["-"], "fveclib=">, Group<f_Group>, Flags<[CC1Option]>,

 def falign_jumps_EQ : Joined<["-"], "falign-jumps=">, Group<clang_ignored_gcc_optimization_f_Group>;
 
 // FIXME: This option should be supported and wired up to our diognostics, but
 // ignore it for now to avoid breaking builds that use it.
 def fdiagnostics_show_location_EQ : Joined<["-"], "fdiagnostics-show-location=">, Group<clang_ignored_f_Group>;
 
 defm fcheck_new : BooleanFFlag<"check-new">, Group<clang_ignored_f_Group>;
 defm caller_saves : BooleanFFlag<"caller-saves">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm reorder_blocks : BooleanFFlag<"reorder-blocks">, Group<clang_ignored_gcc_optimization_f_Group>;
-defm eliminate_unused_debug_types : BooleanFFlag<"eliminate-unused-debug-types">, Group<clang_ignored_f_Group>;
 defm branch_count_reg : BooleanFFlag<"branch-count-reg">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm default_inline : BooleanFFlag<"default-inline">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm fat_lto_objects : BooleanFFlag<"fat-lto-objects">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm float_store : BooleanFFlag<"float-store">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm friend_injection : BooleanFFlag<"friend-injection">, Group<clang_ignored_f_Group>;
 defm function_attribute_list : BooleanFFlag<"function-attribute-list">, Group<clang_ignored_f_Group>;
 defm gcse : BooleanFFlag<"gcse">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm gcse_after_reload: BooleanFFlag<"gcse-after-reload">, Group<clang_ignored_gcc_optimization_f_Group>;
 defm gcse_las: BooleanFFlag<"gcse-las">, Group<clang_ignored_gcc_optimization_f_Group>;

clang/lib/CodeGen/CGDebugInfo.h

public:
   void EmitExternalVariable(llvm::GlobalVariable *GV, const VarDecl *Decl);
 
   /// Emit C++ using directive.
   void EmitUsingDirective(const UsingDirectiveDecl &UD);
 
   /// Emit the type explicitly casted to.
   void EmitExplicitCastType(QualType Ty);
 
+  /// Emit the type even if it might not be used.
+  void EmitAndRetainType(QualType Ty);
+
   /// Emit C++ using declaration.
   void EmitUsingDecl(const UsingDecl &UD);
 
   /// Emit an @import declaration.
   void EmitImportDecl(const ImportDecl &ID);
 
   /// Emit C++ namespace alias.
   llvm::DIImportedEntity *EmitNamespaceAlias(const NamespaceAliasDecl &NA);

clang/lib/CodeGen/CGDebugInfo.cpp

case codegenoptions::DebugLineTablesOnly:
     EmissionKind = llvm::DICompileUnit::LineTablesOnly;
     break;
   case codegenoptions::DebugDirectivesOnly:
     EmissionKind = llvm::DICompileUnit::DebugDirectivesOnly;
     break;
   case codegenoptions::DebugInfoConstructor:
   case codegenoptions::LimitedDebugInfo:
   case codegenoptions::FullDebugInfo:
+  case codegenoptions::UnusedTypeInfo:
     EmissionKind = llvm::DICompileUnit::FullDebug;
     break;
   }
 
   uint64_t DwoId = 0;
   auto &CGOpts = CGM.getCodeGenOpts();
   // The DIFile used by the CU is distinct from the main source
   // file. Its directory part specifies what becomes the
void CGDebugInfo::finalize() {
   // up the final type in the type cache.
   for (auto &RT : RetainedTypes)
     if (auto MD = TypeCache[RT])
       DBuilder.retainType(cast<llvm::DIType>(MD));
 
   DBuilder.finalize();
 }
 
+// Don't ignore in case of explicit cast where it is referenced indirectly.
 void CGDebugInfo::EmitExplicitCastType(QualType Ty) {
-  if (!CGM.getCodeGenOpts().hasReducedDebugInfo())
-    return;
+  if (CGM.getCodeGenOpts().hasReducedDebugInfo())
+    if (auto *DieTy = getOrCreateType(Ty, TheCU->getFile()))
+      DBuilder.retainType(DieTy);
+}
 
+void CGDebugInfo::EmitAndRetainType(QualType Ty) {
+  if (CGM.getCodeGenOpts().hasMaybeUnusedDebugInfo())
     if (auto *DieTy = getOrCreateType(Ty, TheCU->getFile()))
-    // Don't ignore in case of explicit cast where it is referenced indirectly.
       DBuilder.retainType(DieTy);
 }
 
 llvm::DebugLoc CGDebugInfo::SourceLocToDebugLoc(SourceLocation Loc) {
   if (LexicalBlockStack.empty())
     return llvm::DebugLoc();
 
   llvm::MDNode *Scope = LexicalBlockStack.back();
   return llvm::DebugLoc::get(getLineNumber(Loc), getColumnNumber(Loc), Scope);
bool SupportsDWARFv4Ext =
       CGM.getCodeGenOpts().DwarfVersion == 4 &&
       (CGM.getCodeGenOpts().getDebuggerTuning() == llvm::DebuggerKind::LLDB ||
        CGM.getCodeGenOpts().getDebuggerTuning() == llvm::DebuggerKind::GDB);
 
   if (!SupportsDWARFv4Ext && CGM.getCodeGenOpts().DwarfVersion < 5)
     return llvm::DINode::FlagZero;
 
   return llvm::DINode::FlagAllCallsDescribed;
 }

clang/lib/CodeGen/CGDecl.cpp

void CodeGenFunction::EmitDecl(const Decl &D) {
   case Decl::FriendTemplate:
   case Decl::Block:
   case Decl::Captured:
   case Decl::ClassScopeFunctionSpecialization:
   case Decl::UsingShadow:
   case Decl::ConstructorUsingShadow:
   case Decl::ObjCTypeParam:
   case Decl::Binding:
     llvm_unreachable("Declaration should not be in declstmts!");
-  case Decl::Function:  // void X();
   case Decl::Record:    // struct/union/class X;
+  case Decl::CXXRecord: // struct/union/class X; [C++]
+    if (CGDebugInfo *DI = getDebugInfo())
+      if (cast<RecordDecl>(D).getDefinition())
+        DI->EmitAndRetainType(getContext().getRecordType(cast<RecordDecl>(&D)));
+    return;
   case Decl::Enum:      // enum X;
+    if (CGDebugInfo *DI = getDebugInfo())
+      if (cast<EnumDecl>(D).getDefinition())
+        DI->EmitAndRetainType(getContext().getEnumType(cast<EnumDecl>(&D)));
+    return;
+  case Decl::Function:     // void X();
   case Decl::EnumConstant: // enum ? { X = ? }
-  case Decl::CXXRecord: // struct/union/class X; [C++]
   case Decl::StaticAssert: // static_assert(X, ""); [C++0x]
   case Decl::Label:        // __label__ x;
   case Decl::Import:
   case Decl::MSGuid:    // __declspec(uuid("..."))
   case Decl::OMPThreadPrivate:
   case Decl::OMPAllocate:
   case Decl::OMPCapturedExpr:
   case Decl::OMPRequires:
void CodeGenFunction::EmitDecl(const Decl &D) {
   case Decl::OMPDeclareReduction:
     return CGM.EmitOMPDeclareReduction(cast<OMPDeclareReductionDecl>(&D), this);
 
   case Decl::OMPDeclareMapper:
     return CGM.EmitOMPDeclareMapper(cast<OMPDeclareMapperDecl>(&D), this);
 
   case Decl::Typedef:      // typedef int X;
   case Decl::TypeAlias: {  // using X = int; [C++0x]
-    const TypedefNameDecl &TD = cast<TypedefNameDecl>(D);
-    QualType Ty = TD.getUnderlyingType();
-
+    QualType Ty = cast<TypedefNameDecl>(D).getUnderlyingType();
+    if (CGDebugInfo *DI = getDebugInfo())
+      DI->EmitAndRetainType(Ty);
     if (Ty->isVariablyModifiedType())
       EmitVariablyModifiedType(Ty);
-
     return;
   }
   }
 }
 
 /// EmitVarDecl - This method handles emission of any variable declaration
 /// inside a function, including static vars etc.
 void CodeGenFunction::EmitVarDecl(const VarDecl &D) {

clang/lib/CodeGen/CodeGenModule.cpp

if (auto *OID = dyn_cast<ObjCImplDecl>(I)) {
         EmitTopLevelDecl(M);
     }
 
     EmitTopLevelDecl(I);
   }
 }
 
 /// EmitTopLevelDecl - Emit code for a single top level declaration.
 void CodeGenModule::EmitTopLevelDecl(Decl *D) {
   // Ignore dependent declarations.
   if (D->isTemplated())
     return;
 
   // Consteval function shouldn't be emitted.
   if (auto *FD = dyn_cast<FunctionDecl>(D))
     if (FD->isConsteval())
       return;
void CodeGenModule::EmitTopLevelDecl(Decl *D) {
   case Decl::ClassTemplateSpecialization: {
     const auto *Spec = cast<ClassTemplateSpecializationDecl>(D);
     if (CGDebugInfo *DI = getModuleDebugInfo())
       if (Spec->getSpecializationKind() ==
               TSK_ExplicitInstantiationDefinition &&
           Spec->hasDefinition())
         DI->completeTemplateDefinition(*Spec);
   } LLVM_FALLTHROUGH;
-  case Decl::CXXRecord:
-    if (CGDebugInfo *DI = getModuleDebugInfo())
+  case Decl::CXXRecord: {
+    CXXRecordDecl *CRD = cast<CXXRecordDecl>(D);
+    if (CGDebugInfo *DI = getModuleDebugInfo()) {
+      if (CRD->hasDefinition())
+        DI->EmitAndRetainType(getContext().getRecordType(cast<RecordDecl>(D)));
       if (auto *ES = D->getASTContext().getExternalSource())
         if (ES->hasExternalDefinitions(D) == ExternalASTSource::EK_Never)
-          DI->completeUnusedClass(cast<CXXRecordDecl>(*D));
+          DI->completeUnusedClass(*CRD);
+    }
     // Emit any static data members, they may be definitions.
-    for (auto *I : cast<CXXRecordDecl>(D)->decls())
+    for (auto *I : CRD->decls())
       if (isa<VarDecl>(I) || isa<CXXRecordDecl>(I))
         EmitTopLevelDecl(I);
     break;
+  }
     // No code generation needed.
   case Decl::UsingShadow:
   case Decl::ClassTemplate:
   case Decl::VarTemplate:
   case Decl::Concept:
   case Decl::VarTemplatePartialSpecialization:
   case Decl::FunctionTemplate:
   case Decl::TypeAliasTemplate:
void CodeGenModule::EmitTopLevelDecl(Decl *D) {
   case Decl::OMPDeclareMapper:
     EmitOMPDeclareMapper(cast<OMPDeclareMapperDecl>(D));
     break;
 
   case Decl::OMPRequires:
     EmitOMPRequiresDecl(cast<OMPRequiresDecl>(D));
     break;
 
+  case Decl::Typedef:
+  case Decl::TypeAlias: // using foo = bar; [C++11]
+    if (CGDebugInfo *DI = getModuleDebugInfo())
+      DI->EmitAndRetainType(
+          getContext().getTypedefType(cast<TypedefNameDecl>(D)));
+    break;
+
+  case Decl::Record:
+    if (CGDebugInfo *DI = getModuleDebugInfo())
+      if (cast<RecordDecl>(D)->getDefinition())
+        DI->EmitAndRetainType(getContext().getRecordType(cast<RecordDecl>(D)));
+    break;
+
+  case Decl::Enum:
+    if (CGDebugInfo *DI = getModuleDebugInfo())
+      if (cast<EnumDecl>(D)->getDefinition())
+        DI->EmitAndRetainType(getContext().getEnumType(cast<EnumDecl>(D)));
+    break;
+
   default:
     // Make sure we handled everything we should, every other kind is a
     // non-top-level decl.  FIXME: Would be nice to have an isTopLevelDeclKind
     // function. Need to recode Decl::Kind to do that easily.
     assert(isa<TypeDecl>(D) && "Unsupported decl kind");
     break;
   }
 }

clang/lib/Driver/ToolChains/Clang.cpp

case codegenoptions::DebugInfoConstructor:
     CmdArgs.push_back("-debug-info-kind=constructor");
     break;
   case codegenoptions::LimitedDebugInfo:
     CmdArgs.push_back("-debug-info-kind=limited");
     break;
   case codegenoptions::FullDebugInfo:
     CmdArgs.push_back("-debug-info-kind=standalone");
     break;
+  case codegenoptions::UnusedTypeInfo:
+    CmdArgs.push_back("-debug-info-kind=unused-types");
+    break;
   default:
     break;
   }
   if (DwarfVersion > 0)
     CmdArgs.push_back(
         Args.MakeArgString("-dwarf-version=" + Twine(DwarfVersion)));
   switch (DebuggerTuning) {
   case llvm::DebuggerKind::GDB:
static void RenderDebugOptions(const ToolChain &TC, const Driver &D,
   // figure out if we need to "upgrade" it to standalone debug info.
   // We parse these two '-f' options whether or not they will be used,
   // to claim them even if you wrote "-fstandalone-debug -gline-tables-only"
   bool NeedFullDebug = Args.hasFlag(
       options::OPT_fstandalone_debug, options::OPT_fno_standalone_debug,
       DebuggerTuning == llvm::DebuggerKind::LLDB ||
           TC.GetDefaultStandaloneDebug());
   if (const Arg *A = Args.getLastArg(options::OPT_fstandalone_debug))
     (void)checkDebugInfoOption(A, Args, D, TC);
-  if (DebugInfoKind == codegenoptions::LimitedDebugInfo && NeedFullDebug)
+
+  if (DebugInfoKind == codegenoptions::LimitedDebugInfo) {
+    if (Args.hasFlag(options::OPT_fno_eliminate_unused_debug_types,
+                     options::OPT_feliminate_unused_debug_types, false))
+      DebugInfoKind = codegenoptions::UnusedTypeInfo;
+    else if (NeedFullDebug)
       DebugInfoKind = codegenoptions::FullDebugInfo;
+  }
 
   if (Args.hasFlag(options::OPT_gembed_source, options::OPT_gno_embed_source,
                    false)) {
     // Source embedding is a vendor extension to DWARF v5. By now we have
     // checked if a DWARF version was stated explicitly, and have otherwise
     // fallen back to the target default, so if this is still not at least 5
     // we emit an error.
     const Arg *A = Args.getLastArg(options::OPT_gembed_source);
static void RenderDebugOptions(const ToolChain &TC, const Driver &D,
   // Adjust the debug info kind for the given toolchain.
   TC.adjustDebugInfoKind(DebugInfoKind, Args);
 
   // When emitting remarks, we need at least debug lines in the output.
   if (willEmitRemarks(Args) &&
       DebugInfoKind <= codegenoptions::DebugDirectivesOnly)
     DebugInfoKind = codegenoptions::DebugLineTablesOnly;
 
   RenderDebugEnablingArgs(Args, CmdArgs, DebugInfoKind, DWARFVersion,
                           DebuggerTuning);
 
   // -fdebug-macro turns on macro debug info generation.
   if (Args.hasFlag(options::OPT_fdebug_macro, options::OPT_fno_debug_macro,
                    false))
     if (checkDebugInfoOption(Args.getLastArg(options::OPT_fdebug_macro), Args,
                              D, TC))
       CmdArgs.push_back("-debug-info-macro");
 
   // -ggnu-pubnames turns on gnu style pubnames in the backend.
   const auto *PubnamesArg =
       Args.getLastArg(options::OPT_ggnu_pubnames, options::OPT_gno_gnu_pubnames,
#endif
   Args.AddLastArg(CmdArgs, options::OPT_fvisibility_inlines_hidden);
   Args.AddLastArg(CmdArgs, options::OPT_fvisibility_global_new_delete_hidden);
 
   Args.AddLastArg(CmdArgs, options::OPT_ftlsmodel_EQ);
 
   // Forward -f (flag) options which we can pass directly.
   Args.AddLastArg(CmdArgs, options::OPT_femit_all_decls);
   Args.AddLastArg(CmdArgs, options::OPT_fheinous_gnu_extensions);
   Args.AddLastArg(CmdArgs, options::OPT_fdigraphs, options::OPT_fno_digraphs);
   Args.AddLastArg(CmdArgs, options::OPT_fno_operator_names);
   Args.AddLastArg(CmdArgs, options::OPT_femulated_tls,
                   options::OPT_fno_emulated_tls);
 
   // AltiVec-like language extensions aren't relevant for assembling.
   if (!isa<PreprocessJobAction>(JA) || Output.getType() != types::TY_PP_Asm)
     Args.AddLastArg(CmdArgs, options::OPT_fzvector);
 

clang/lib/Frontend/CompilerInvocation.cpp

static bool ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args, InputKind IK,
   if (Arg *A = Args.getLastArg(OPT_debug_info_kind_EQ)) {
     unsigned Val =
         llvm::StringSwitch<unsigned>(A->getValue())
             .Case("line-tables-only", codegenoptions::DebugLineTablesOnly)
             .Case("line-directives-only", codegenoptions::DebugDirectivesOnly)
             .Case("constructor", codegenoptions::DebugInfoConstructor)
             .Case("limited", codegenoptions::LimitedDebugInfo)
             .Case("standalone", codegenoptions::FullDebugInfo)
+            .Case("unused-types", codegenoptions::UnusedTypeInfo)
             .Default(~0U);
     if (Val == ~0U)
       Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args)
                                                 << A->getValue();
     else
       Opts.setDebugInfo(static_cast<codegenoptions::DebugInfoKind>(Val));
   }
   if (Arg *A = Args.getLastArg(OPT_debugger_tuning_EQ)) {
static bool ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args, InputKind IK,
   Opts.LTOVisibilityPublicStd = Args.hasArg(OPT_flto_visibility_public_std);
   Opts.SplitDwarfFile = std::string(Args.getLastArgValue(OPT_split_dwarf_file));
   Opts.SplitDwarfOutput =
       std::string(Args.getLastArgValue(OPT_split_dwarf_output));
   Opts.SplitDwarfInlining = !Args.hasArg(OPT_fno_split_dwarf_inlining);
   Opts.DebugTypeExtRefs = Args.hasArg(OPT_dwarf_ext_refs);
   Opts.DebugExplicitImport = Args.hasArg(OPT_dwarf_explicit_import);
   Opts.DebugFwdTemplateParams = Args.hasArg(OPT_debug_forward_template_params);
   Opts.EmbedSource = Args.hasArg(OPT_gembed_source);
   Opts.ForceDwarfFrameSection = Args.hasArg(OPT_fforce_dwarf_frame);
 
   for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ)) {
     auto Split = StringRef(Arg).split('=');
     Opts.DebugPrefixMap.insert(
         {std::string(Split.first), std::string(Split.second)});
   }
 

clang/test/CodeGen/debug-info-unused-types.c

+// RUN: %clang -fno-eliminate-unused-debug-types -g -emit-llvm -S -o - %s | FileCheck %s
+// RUN: %clang -fno-eliminate-unused-debug-types -g1 -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang -feliminate-unused-debug-types -g -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang -g -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+typedef int my_int;
+struct foo {};
+enum bar { BAR };
+union baz {};
+
+void quux(void) {
+  typedef int x;
+  struct y {};
+  enum z { Z };
+  union w {};
+}
+
+// Check that debug info is emitted for the typedef, struct, enum, and union
+// when -fno-eliminate-unused-debug-types and -g are set.
+
+// CHECK: !DICompileUnit{{.+}}retainedTypes: [[RETTYPES:![0-9]+]]
+// CHECK: [[TYPE0:![0-9]+]] = !DICompositeType(tag: DW_TAG_enumeration_type, name: "bar"
+// CHECK: [[TYPE1:![0-9]+]] = !DIEnumerator(name: "BAR"
+// CHECK: [[TYPE2:![0-9]+]] = !DICompositeType(tag: DW_TAG_enumeration_type, name: "z"
+// CHECK: [[TYPE3:![0-9]+]] = !DIEnumerator(name: "Z"
+// CHECK: [[RETTYPES]] = !{[[TYPE4:![0-9]+]], [[TYPE5:![0-9]+]], [[TYPE0]], [[TYPE6:![0-9]+]], !17, [[TYPE7:![0-9]+]], [[TYPE2]], [[TYPE8:![0-9]+]]}
+// CHECK: [[TYPE4]] = !DIDerivedType(tag: DW_TAG_typedef, name: "my_int"
+// CHECK: [[TYPE5]] = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "foo"
+// CHECK: [[TYPE6]] = distinct !DICompositeType(tag: DW_TAG_union_type, name: "baz"
+// CHECK: [[TYPE7]] = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "y"
+// CHECK: [[TYPE8]] = distinct !DICompositeType(tag: DW_TAG_union_type, name: "w"
+
+// Check that debug info is not emitted for the typedef, struct, enum, and
+// union when -fno-eliminate-unused-debug-types and -g are not set. These are
+// the same checks as above with `NODBG-NOT` rather than `CHECK`.
+
+// NODBG-NOT: !DI{{CompositeType|Enumerator|DerivedType}}
+
+// Check that debug info is not emitted for declarations. Obnoxious
+// indentifiers are to avoid collisions with the SHA emittied as debug info.
+struct unused_struct;
+enum unused_enum;
+union unused_union;
+void b0(void) {
+  struct unused_local_struct;
+  enum unused_local_enum;
+  union unused_local_union;
+}
+
+// NODBG-NOT: name: "unused_

clang/test/CodeGen/debug-info-unused-types.cpp

+// RUN: %clang++ -fno-eliminate-unused-debug-types -g -emit-llvm -S -o - %s | FileCheck %s
+// RUN: %clang++ -fno-eliminate-unused-debug-types -g1 -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang++ -feliminate-unused-debug-types -g -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang++ -g -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+// RUN: %clang++ -emit-llvm -S -o - %s | FileCheck --check-prefix=NODBG %s
+using foo = int;
+class bar {};
+enum class baz { BAZ };
+
+void quux() {
+  using x = int;
+  class y {};
+  enum class z { Z };
+}
+
+// CHECK: !DICompileUnit{{.+}}retainedTypes: [[RETTYPES:![0-9]+]]
+// CHECK: [[TYPE0:![0-9]+]] = !DICompositeType(tag: DW_TAG_enumeration_type, name: "baz"
+// CHECK: [[TYPE1:![0-9]+]] = !DIEnumerator(name: "BAZ"
+// CHECK: [[TYPE2:![0-9]+]] = !DICompositeType(tag: DW_TAG_enumeration_type, name: "z"
+// CHECK: [[TYPE3:![0-9]+]] = !DIEnumerator(name: "Z"
+// CHECK: [[RETTYPES]] = !{[[TYPE4:![0-9]+]], [[TYPE5:![0-9]+]], [[TYPE0]], !5, [[TYPE6:![0-9]+]], [[TYPE2]]}
+// CHECK: [[TYPE4]] = !DIDerivedType(tag: DW_TAG_typedef, name: "foo"
+// CHECK: [[TYPE5]] = distinct !DICompositeType(tag: DW_TAG_class_type, name: "bar"
+// CHECK: [[TYPE6]] = distinct !DICompositeType(tag: DW_TAG_class_type, name: "y"
+
+// NODBG-NOT: !DI{{CompositeType|Enumerator|DerivedType}}
+
+class unused_class;
+enum class unused_enum_class;
+
+// NODBG-NOT: name: "unused_

clang/test/Driver/debug-options.c

 // RUN: %clang -### -gdwarf-2 -gembed-source %s 2>&1 | FileCheck -check-prefix=GEMBED_2 %s
 // RUN: %clang -### -gdwarf-5 -gno-embed-source %s 2>&1 | FileCheck -check-prefix=NOGEMBED_5 %s
 // RUN: %clang -### -gdwarf-2 -gno-embed-source %s 2>&1 | FileCheck -check-prefix=NOGEMBED_2 %s
 //
 // GEMBED_5:  "-gembed-source"
 // GEMBED_2:  error: invalid argument '-gembed-source' only allowed with '-gdwarf-5'
 // NOGEMBED_5-NOT:  "-gembed-source"
 // NOGEMBED_2-NOT:  error: invalid argument '-gembed-source' only allowed with '-gdwarf-5'
+//
+// RUN: %clang -### -g -fno-eliminate-unused-debug-types -c %s 2>&1 \
+// RUN:        | FileCheck -check-prefix=DEBUG_UNUSED_TYPES %s
+// DEBUG_UNUSED_TYPES: "-debug-info-kind=unused-types"
+// DEBUG_UNUSED_TYPES-NOT: "-debug-info-kind=limited"
+// RUN: %clang -### -g -feliminate-unused-debug-types -c %s 2>&1 \
+// RUN:        | FileCheck -check-prefix=NO_DEBUG_UNUSED_TYPES %s
+// NO_DEBUG_UNUSED_TYPES: "-debug-info-kind=limited"
+// NO_DEBUG_UNUSED_TYPES-NOT: "-debug-info-kind=unused-types"