[LLDB] Add DynamicLoaderWasmDYLD plugin for WebAssembly debugging

[paolosev]

Summary

Add a dynamic loader plug-in class for WebAssembly modules.

Depends on D72650.

[paolosev]

What is the best way to test this class?

[labath]

Testing dynamic loaders is a bit tricky as they require an actual process around. The best thing available to us right now is the "gdb-client" approach, which consists of mocking the responses of the gdb server. It's not the easiest way to write tests, but I don't think it should be that difficult in this case -- you shouldn't need to mock that many packets -- the main one is qXfer:libraries. Then you should be able to run something like "image lookup -a" (or SBTarget::ResolveLoadAddress, if you want to try your hand at the scripting API) and check that it resolves to the correct section+offset pair. You can look at the existing tests in packages/Python/lldbsuite/test/functionalities/gdb_remote_client/ to see how this works...

I will have some more questions about the interaction of this function with ObjectFileWasm::SetLoadAddress, but I need to think this over a bit...

[clayborg]

lldb/source/Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.cpp

122

Is there only ever just a code address and an image address? If you have more than 2 sections you don't want to load the different sections at the same address because converting a load address back into a section should provide a one to one mapping. So looking up 0x1000 currently should not return N sections, it should return 1 section. If this doesn't happen the binary search of an address in the target section load list could return any of the sections that match.

[labath]

lldb/source/Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.cpp

91–127

Right, so, given that (IIUC) you use the qXfer:libraries packet, I believe this code should not be needed. In this case ProcessGDBRemote::LoadModules should do all the work (by calling into DynamicLoaderWasmDYLD::LoadModuleAtAddress, which will then call into ObjectFileWasm::SetLoadAddress).

The fact that you need fix up section load addresses after these functions are done makes me believe that those functions are not doing their job properly. That wouldn't be too bad if there is a reason for that, but right now I don't see any indication that this is the case.

Can you explain what is the purpose of this code (specifically, what would happen without it, if we only had m_process->LoadModules() here), so we can figure out what to do about this?

[paolosev]

I have verified the logic of the dynamic loader quite carefully, but there are a couple of things to clarify.

A Wasm module is loaded at a 64 bit address, where the upper 32 bits are used as module identifier. Let’s say that we have a module with Id==4, so it will be loaded at address 0x00000004`00000000. Each section is loaded at its relative file offset. Therefore if the code section starts at file offset 0x4d in the Wasm module, we call:
Target::SetSectionLoadAddress(section_sp, 0x40000004d).

The module can also contain embedded DWARF sections, which will also be loaded at their relative file offset in the same way. And since there cannot be duplicated sections in a module, there is no overlapping, we can always convert a load address back into a section.

However, there are two complications.

The first is that we need to call Target::SetSectionLoadAddress() twice, from two different places. First we need to call Target::SetSectionLoadAddress() in ObjectFileWasm::SetLoadAddress(), and then again in DynamicLoaderWasmDYLD::DidAttach(). The reason for this seems to originate in the sequence of function calls:

In DynamicLoaderWasmDYLD::DidAttach() we call ProcessGDBRemote::LoadModules() to get list of loaded modules from the remote (Wasm engine).
ProcessGDBRemote::LoadModules() calls, first:

• DynamicLoaderWasmDYLD::LoadModuleAtAddress() and from there:
  1. DynamicLoader::UpdateLoadedSections() -> ObjectFileWasm::SetLoadAddress()
  2. Target::GetImages()::AppendIfNeeded(module) -> ProcessGDBRemote::ModulesDidLoad() -> JITLoaderList::ModulesDidLoad() -> Module::GetSymbolFile() -> SymbolFileDWARF::CalculateAbilities(). Here we initialize the symbols for the module, and set m_did_load_symfile, but for this to work we need to have already set the load address for each section, in the previous ObjectFileWasm::SetLoadAddress().

then:

• Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear()

So, at the end of LoadModules() in DynamicLoaderWasmDYLD::DidAttach() the SectionLoadList is empty, and we need to set it again by calling Target::.SetSectionLoadAddress() again.
This works but the duplication is ugly; is there a way to improve this?
_

The second problem is that the Code Section needs to be initialized (in ObjectFileWasm::CreateSections()) with m_file_addr = m_file_offset = 0, and not with the actual file offset of the Code section in the Wasm file. If we set Section::m_file_addr and Section::m_file_offset to the actual code offset, the DWARF info does not work correctly.

I have some doubts regarding the DWARF data generated by Clang for a Wasm target. Looking at an example, for a Wasm module that has the Code section at offset 0x57, I see this DWARF data:

0x0000000b: DW_TAG_compile_unit
              […]
              DW_AT_low_pc (0x0000000000000000)
              DW_AT_ranges (0x00000000
                 [0x00000002, 0x0000000e)
                 [0x0000000f, 0x0000001a)
                 [0x0000001b, 0x00000099)
                 [0x0000009b, 0x0000011c))

The documentation says that “Wherever a code address is used in DWARF for WebAssembly, it must be the offset of an instruction relative within the Code section of the WebAssembly file.”
But is this correct? Shouldn't maybe code addresses be offset-ed by the file address of the Code section?

[labath]

[ looping in @aadsm for the svr4 stuff ]

Thanks for adding the test, and for the detailed writeup. Please find my comments inline.

In D72751#1835502, @paolosev wrote:

The first is that we need to call Target::SetSectionLoadAddress() twice, from two different places. First we need to call Target::SetSectionLoadAddress() in ObjectFileWasm::SetLoadAddress(), and then again in DynamicLoaderWasmDYLD::DidAttach(). The reason for this seems to originate in the sequence of function calls:

In DynamicLoaderWasmDYLD::DidAttach() we call ProcessGDBRemote::LoadModules() to get list of loaded modules from the remote (Wasm engine).
ProcessGDBRemote::LoadModules() calls, first:

• DynamicLoaderWasmDYLD::LoadModuleAtAddress() and from there: ...

then:

• Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear()

So, at the end of LoadModules() in DynamicLoaderWasmDYLD::DidAttach() the SectionLoadList is empty, and we need to set it again by calling Target::.SetSectionLoadAddress() again.
This works but the duplication is ugly; is there a way to improve this?

I hope so. :) This seems like a bug in ProcessGDBRemote::LoadModules. It seems wrong/wasteful/etc to do all this work to compute the section load addresses only to have them be thrown away by SetExecutableModule. Maybe all it would take is to reverse the order of these two actions, so that the load addresses persist? Can you try something like that?

On a side note, ProcessGDBRemote::LoadModules seems a bit internally inconsistent. At one place it claims that "The main executable will never be included in libraries-svr4", but then it goes on to set an executable module anyway. This could in fact be a clue as to why this problem hasn't showed up on other platforms -- if the remote does not send the executable, then SetExecutableModule is not called.

On a side-side note, this may mean that you sending the main wasm file through qXfer:libraries-svr4 may not be correct. However, fixing that would mean finding another way to communicate the main executable name/address. I don't think we currently have an easy way to do that so it may be better to fix ProcessGDBRemote::LoadModules, given that it "almost" supports executables.

I am also worried about the fact that SymbolFileDWARF::CalculateAbilities requires the module to be "loaded". That shouldn't be normally required. That function does a very basic check on some sections, and this should work fine without those sections having a "load address", even if they are actually being loaded from target memory. I think this means there are some additional places where ObjectFileWasm should use m_memory_addr instead of something else...

The second problem is that the Code Section needs to be initialized (in ObjectFileWasm::CreateSections()) with m_file_addr = m_file_offset = 0, and not with the actual file offset of the Code section in the Wasm file. If we set Section::m_file_addr and Section::m_file_offset to the actual code offset, the DWARF info does not work correctly.

I have some doubts regarding the DWARF data generated by Clang for a Wasm target. Looking at an example, for a Wasm module that has the Code section at offset 0x57, I see this DWARF data:

0x0000000b: DW_TAG_compile_unit
              […]
              DW_AT_low_pc (0x0000000000000000)
              DW_AT_ranges (0x00000000
                 [0x00000002, 0x0000000e)
                 [0x0000000f, 0x0000001a)
                 [0x0000001b, 0x00000099)
                 [0x0000009b, 0x0000011c))

The documentation says that “Wherever a code address is used in DWARF for WebAssembly, it must be the offset of an instruction relative within the Code section of the WebAssembly file.”
But is this correct? Shouldn't maybe code addresses be offset-ed by the file address of the Code section?

That's interesting. I don't think that clang is really wrong/non-conforming here, but this choice plays a rather poorly with the way lldb handles object files (and how your remote presents them). In this case special casing the code section may be fine, but you should be aware that there are places in lldb which expect that the "load bias" (the delta between file and load addresses) is the same for all sections in a module (because that's how it works elsewhere), and they may not like this. However, if you have only one code section, I think you should be mostly fine.

I do think that this can be handled in a better way (I'm pretty sure you don't need to zero out the file offset for instance), but we can wait with that until we resolve the first point...

[paolosev]

Thanks for the explanation! I wasn't quite clear on "executable module" here, but after your comments I realized that Target::SetExecutableModule() should not probably be called also for Wasm modules.
The point is that ObjectFileWasm::CalculateType() should return eTypeSharedLibrary, not eTypeExecutable.
With this change the first issue is easily solved: we just need to call Target::SetSectionLoadAddress() once, in ObjectFileWasm::SetLoadAddress() because Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear() is not called, and DynamicLoaderWasmDYLD::DidAttach() can be simplified to just call ProcessGDBRemote::LoadModules().
Does this solution work for you? If so, we should look at the second point, the need to initialize m_file_addr = m_file_offset = 0 for the "code" Section in order to make the DWARF symbols work...

[aadsm]

Yeah, I'm not sure why the LoadModules function is calling target.SetExecutableModule. It is true that the libraries-svr4 will not include the main executable in its list.
This code was added in the context of providing qXfer:libraries support here: https://reviews.llvm.org/D9471. I don't see any mention of including the executable on that packet though: https://sourceware.org/gdb/current/onlinedocs/gdb/Library-List-Format.html. @clayborg was the main reviewer there (although this was 5 years ago or so) and he does mention multiple times in the comments this exact issue with calling target.SetExecutableModule. Maybe he can still remember and provide some light here :).

[labath]

In D72751#1837764, @paolosev wrote:

Thanks for the explanation! I wasn't quite clear on "executable module" here, but after your comments I realized that Target::SetExecutableModule() should not probably be called also for Wasm modules.
The point is that ObjectFileWasm::CalculateType() should return eTypeSharedLibrary, not eTypeExecutable.
With this change the first issue is easily solved: we just need to call Target::SetSectionLoadAddress() once, in ObjectFileWasm::SetLoadAddress() because Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear() is not called, and DynamicLoaderWasmDYLD::DidAttach() can be simplified to just call ProcessGDBRemote::LoadModules().
Does this solution work for you?

I am fine that. I don't really know enough about wasm to say if you should have something like the "main" module, but I don't think it should make a big difference to lldb anyway. And another benefit to that is that we can say we stick to the qXfer "spec" and only send the shared libraries over.

If so, we should look at the second point, the need to initialize m_file_addr = m_file_offset = 0 for the "code" Section in order to make the DWARF symbols work...

Yes, let's do that. Can you check what happens if you just move the file_offset = sect_info.offset & 0xffffffff; line in ObjectFileWasm outside of the if(!code) block? My guess is that you'll just need to replace some section->GetFileOffset() calls with ->GetFileAddress(). I'm pretty sure those calls will be only in wasm code because other object formats don't have sections at file offset zero, and everything is fine with that.

lldb/source/Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.cpp

57–70

This comment is probably not that useful anymore...

[labath]

In D72751#1837883, @aadsm wrote:

Yeah, I'm not sure why the LoadModules function is calling target.SetExecutableModule. It is true that the libraries-svr4 will not include the main executable in its list.
This code was added in the context of providing qXfer:libraries support here: https://reviews.llvm.org/D9471. I don't see any mention of including the executable on that packet though: https://sourceware.org/gdb/current/onlinedocs/gdb/Library-List-Format.html. @clayborg was the main reviewer there (although this was 5 years ago or so) and he does mention multiple times in the comments this exact issue with calling target.SetExecutableModule. Maybe he can still remember and provide some light here :).

Thanks for digging this up, Antonio. My impression of that thread is that the author did not fully understand what Greg was asking him to do, and then that discussion got buried in other stuff. Given that qXfer is not supposed to send shared libraries, wasm is not going to be using it, and it doesn't look like this code could ever work, I'm tempted to just remove this SetExecutable block. :)

[paolosev]

Modified to set m_file_offset to be the correct offset of the Code section. This also simplifies the code in ObjectFileWasm to avoid a special case for the code section in ObjectFileWasm::SetLoadAddress.
The DWARF code seems to only use GetFileAddress(), which still needs to return zero for the Code section.

[labath]

Thanks. I am glad that we were able to sort that out.

Now that there's nothing wasm-specific in DynamicLoaderWasmDYLD::LoadModuleAtAddress, I have another question. :)

The code in that function is a subset of the base DynamicLoader::LoadModuleAtAddress method you are overriding. Is there a reason for that?
It doesn't seem like the additional stuff in the base method should hurt here. What the additional code does is that it tries to search for the object file on the local filesystem, and if it finds it, it will use that instead of copying the file over from the remote. In fact, that sounds like it could be useful here too, as copying that much data over gdb-remote isn't particularly fast..

What do you think about that?

[clayborg]

In D72751#1835656, @labath wrote:

[ looping in @aadsm for the svr4 stuff ]

Thanks for adding the test, and for the detailed writeup. Please find my comments inline.

In D72751#1835502, @paolosev wrote:

The first is that we need to call Target::SetSectionLoadAddress() twice, from two different places. First we need to call Target::SetSectionLoadAddress() in ObjectFileWasm::SetLoadAddress(), and then again in DynamicLoaderWasmDYLD::DidAttach(). The reason for this seems to originate in the sequence of function calls:

In DynamicLoaderWasmDYLD::DidAttach() we call ProcessGDBRemote::LoadModules() to get list of loaded modules from the remote (Wasm engine).
ProcessGDBRemote::LoadModules() calls, first:

• DynamicLoaderWasmDYLD::LoadModuleAtAddress() and from there: ...

then:

• Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear()

So, at the end of LoadModules() in DynamicLoaderWasmDYLD::DidAttach() the SectionLoadList is empty, and we need to set it again by calling Target::.SetSectionLoadAddress() again.
This works but the duplication is ugly; is there a way to improve this?

I hope so. :) This seems like a bug in ProcessGDBRemote::LoadModules. It seems wrong/wasteful/etc to do all this work to compute the section load addresses only to have them be thrown away by SetExecutableModule. Maybe all it would take is to reverse the order of these two actions, so that the load addresses persist? Can you try something like that?

I would be interested to see if this helps as well.

On a side note, ProcessGDBRemote::LoadModules seems a bit internally inconsistent. At one place it claims that "The main executable will never be included in libraries-svr4", but then it goes on to set an executable module anyway. This could in fact be a clue as to why this problem hasn't showed up on other platforms -- if the remote does not send the executable, then SetExecutableModule is not called.

The logic in ProcessGDBRemote::LoadModules is bad, the executable should be set before any sections are loaded.

On a side-side note, this may mean that you sending the main wasm file through qXfer:libraries-svr4 may not be correct. However, fixing that would mean finding another way to communicate the main executable name/address. I don't think we currently have an easy way to do that so it may be better to fix ProcessGDBRemote::LoadModules, given that it "almost" supports executables.

Is there a "qXfer:executable"? Seems from the code in ProcessGDBRemote::LoadModules() that people were seeing the main executable in the list somewhere. Be a shame to not allow it.

I am also worried about the fact that SymbolFileDWARF::CalculateAbilities requires the module to be "loaded". That shouldn't be normally required. That function does a very basic check on some sections, and this should work fine without those sections having a "load address", even if they are actually being loaded from target memory. I think this means there are some additional places where ObjectFileWasm should use m_memory_addr instead of something else...

So a lot of these problems might go away if the modify the ObjectFileWASM to "do the right thing" when the ObjectFile is from a live process. Object file instances know if there are from a live process because the ObjectFile members:

lldb::ProcessWP m_process_wp;
const lldb::addr_t m_memory_addr;

Will be filled in. So the object file doesn't need to have its sections loaded in a target in order to read section contents right? The object file can be asked to read its section data via ObjectFile::ReadSectionData(...) (2 variants).

So Pavel is correct, the sections don't need to be loaded for anything in SymbolFileDWARF::CalculateAbilities(...). The section list does need to be created and available during SymbolFileDWARF::CalculateAbilities(...). We just need to be able to get the section contents and poke around at the DWARF bits.

The second problem is that the Code Section needs to be initialized (in ObjectFileWasm::CreateSections()) with m_file_addr = m_file_offset = 0, and not with the actual file offset of the Code section in the Wasm file. If we set Section::m_file_addr and Section::m_file_offset to the actual code offset, the DWARF info does not work correctly.

I have some doubts regarding the DWARF data generated by Clang for a Wasm target. Looking at an example, for a Wasm module that has the Code section at offset 0x57, I see this DWARF data:

0x0000000b: DW_TAG_compile_unit
              […]
              DW_AT_low_pc (0x0000000000000000)
              DW_AT_ranges (0x00000000
                 [0x00000002, 0x0000000e)
                 [0x0000000f, 0x0000001a)
                 [0x0000001b, 0x00000099)
                 [0x0000009b, 0x0000011c))

The documentation says that “Wherever a code address is used in DWARF for WebAssembly, it must be the offset of an instruction relative within the Code section of the WebAssembly file.”
But is this correct? Shouldn't maybe code addresses be offset-ed by the file address of the Code section?

That's interesting. I don't think that clang is really wrong/non-conforming here, but this choice plays a rather poorly with the way lldb handles object files (and how your remote presents them). In this case special casing the code section may be fine, but you should be aware that there are places in lldb which expect that the "load bias" (the delta between file and load addresses) is the same for all sections in a module (because that's how it works elsewhere), and they may not like this. However, if you have only one code section, I think you should be mostly fine.

I do think that this can be handled in a better way (I'm pretty sure you don't need to zero out the file offset for instance), but we can wait with that until we resolve the first point...

So the DWARF plug-ins definitely expect the addresses in the DWARF to be file addresses where if you ask the module for its section list, it can look up this "file address" in the list, it will find a single section. If this is not the case, you will need to make a special SymbolFileDWARFWasm subclass and this will be very tricky as any address you get from the DWARF will need to be converted so that the lookup the section list will work.

[paolosev]

In D72751#1841498, @labath wrote:

Thanks. I am glad that we were able to sort that out.

Now that there's nothing wasm-specific in DynamicLoaderWasmDYLD::LoadModuleAtAddress, I have another question. :)

The code in that function is a subset of the base DynamicLoader::LoadModuleAtAddress method you are overriding. Is there a reason for that?
It doesn't seem like the additional stuff in the base method should hurt here. What the additional code does is that it tries to search for the object file on the local filesystem, and if it finds it, it will use that instead of copying the file over from the remote. In fact, that sounds like it could be useful here too, as copying that much data over gdb-remote isn't particularly fast..

What do you think about that?

You are right! At this point DynamicLoaderWasmDYLD::LoadModuleAtAddress does not do anything specific to Wasm and there is no reason to override DynamicLoader::LoadModuleAtAddress. We can just call that base function, which should also work when the Wasm module is in the local filesystem.

[paolosev]

Comments inline...

In D72751#1835502, @paolosev wrote:

The first is that we need to call Target::SetSectionLoadAddress() twice, from two different places. First we need to call Target::SetSectionLoadAddress() in ObjectFileWasm::SetLoadAddress(), and then again in DynamicLoaderWasmDYLD::DidAttach(). The reason for this seems to originate in the sequence of function calls:

In DynamicLoaderWasmDYLD::DidAttach() we call ProcessGDBRemote::LoadModules() to get list of loaded modules from the remote (Wasm engine).
ProcessGDBRemote::LoadModules() calls, first:

• DynamicLoaderWasmDYLD::LoadModuleAtAddress() and from there: ...

then:

• Target::SetExecutableModule() -> Target::ClearModules() -> SectionLoadList::Clear()

So, at the end of LoadModules() in DynamicLoaderWasmDYLD::DidAttach() the SectionLoadList is empty, and we need to set it again by calling Target::.SetSectionLoadAddress() again.
This works but the duplication is ugly; is there a way to improve this?

I hope so. :) This seems like a bug in ProcessGDBRemote::LoadModules. It seems wrong/wasteful/etc to do all this work to compute the section load addresses only to have them be thrown away by SetExecutableModule. Maybe all it would take is to reverse the order of these two actions, so that the load addresses persist? Can you try something like that?

I would be interested to see if this helps as well.

The problem went away with the realization that ObjectFileWasm type should be eTypeSharedLibrary, not eTypeExecutable.
But more generically for ProcessGDBRemote::LoadModules I don't know if it would be easily possible to reverse the order of the calls to DynamicLoader::LoadModuleAtAddress and Target::SetExecutableModule given that the first creates the Module which is passed to the second. Maybe refactoring DynamicLoader::LoadModuleAtAddress so that it calls Target::SetExecutableModule just after target.GetOrCreateModule but before UpdateLoadedSections, but other DynamicLoader plugins could override this method...

I am also worried about the fact that SymbolFileDWARF::CalculateAbilities requires the module to be "loaded". That shouldn't be normally required. That function does a very basic check on some sections, and this should work fine without those sections having a "load address", even if they are actually being loaded from target memory. I think this means there are some additional places where ObjectFileWasm should use m_memory_addr instead of something else...

So a lot of these problems might go away if the modify the ObjectFileWASM to "do the right thing" when the ObjectFile is from a live process. Object file instances know if there are from a live process because the ObjectFile members:

lldb::ProcessWP m_process_wp;
const lldb::addr_t m_memory_addr;

Will be filled in. So the object file doesn't need to have its sections loaded in a target in order to read section contents right? The object file can be asked to read its section data via ObjectFile::ReadSectionData(...) (2 variants).

So Pavel is correct, the sections don't need to be loaded for anything in SymbolFileDWARF::CalculateAbilities(...). The section list does need to be created and available during SymbolFileDWARF::CalculateAbilities(...). We just need to be able to get the section contents and poke around at the DWARF bits.

Yes, this seems to be the case with the current implementation of ObjectFileWASM. It creates the section list in ObjectFileWasm::SetLoadAddress which calls Target::SetSectionLoadAddress but the sections don't need to be fully loaded, and during SymbolFileDWARF::CalculateAbilities(...) ObjectFile::ReadSectionData is called to load the necessary data.

The second problem is that the Code Section needs to be initialized (in ObjectFileWasm::CreateSections()) with m_file_addr = m_file_offset = 0, and not with the actual file offset of the Code section in the Wasm file. If we set Section::m_file_addr and Section::m_file_offset to the actual code offset, the DWARF info does not work correctly.

I have some doubts regarding the DWARF data generated by Clang for a Wasm target. Looking at an example, for a Wasm module that has the Code section at offset 0x57, I see this DWARF data:

0x0000000b: DW_TAG_compile_unit
              […]
              DW_AT_low_pc (0x0000000000000000)
              DW_AT_ranges (0x00000000
                 [0x00000002, 0x0000000e)
                 [0x0000000f, 0x0000001a)
                 [0x0000001b, 0x00000099)
                 [0x0000009b, 0x0000011c))

The documentation says that “Wherever a code address is used in DWARF for WebAssembly, it must be the offset of an instruction relative within the Code section of the WebAssembly file.”
But is this correct? Shouldn't maybe code addresses be offset-ed by the file address of the Code section?

That's interesting. I don't think that clang is really wrong/non-conforming here, but this choice plays a rather poorly with the way lldb handles object files (and how your remote presents them). In this case special casing the code section may be fine, but you should be aware that there are places in lldb which expect that the "load bias" (the delta between file and load addresses) is the same for all sections in a module (because that's how it works elsewhere), and they may not like this. However, if you have only one code section, I think you should be mostly fine.

I do think that this can be handled in a better way (I'm pretty sure you don't need to zero out the file offset for instance), but we can wait with that until we resolve the first point...

So the DWARF plug-ins definitely expect the addresses in the DWARF to be file addresses where if you ask the module for its section list, it can look up this "file address" in the list, it will find a single section. If this is not the case, you will need to make a special SymbolFileDWARFWasm subclass and this will be very tricky as any address you get from the DWARF will need to be converted so that the lookup the section list will work.

File addresses can uniquely identify a single section, there is no problem with this, and there is always a single code section per module. The only "weirdness" is that since DWARF code addresses for Wasm are calculated from the beginning of the Code section, not the beginning of the file, for the Code section, Section::m_file_offset can normally be the file offset, but Section::m_file_addr needs to be zero. This seems to make all DWARF-related code work, but, as Pavel said, maybe there could be places where LLDB expects the "load bias" to be the same for each section, which could cause problems?

[labath]

Thanks. My hopefully final question is not really for you but more like for other lldb developers (@jingham, @clayborg, etc.).

Given that this plugin is now consisting of boiler plate only, I am wondering if we should not instead make it possible for this use case to work without any special plugins needed. A couple of options that come to mind are:

• make the base DynamicLoader class instantiatable, and use it whenever we fail to find a specialized plugin
• same as above, but only do that for ProcessGDBRemote instances
• make ProcessGDBRemote call LoadModules() itself if no dynamic loader instance is available

WDYT?

In D72751#1843458, @paolosev wrote:

Yes, this seems to be the case with the current implementation of ObjectFileWASM. It creates the section list in ObjectFileWasm::SetLoadAddress which calls Target::SetSectionLoadAddress but the sections don't need to be fully loaded, and during SymbolFileDWARF::CalculateAbilities(...) ObjectFile::ReadSectionData is called to load the necessary data.

This is correct, but I want to point out that the "load" in SetLoadAddress and in ReadSectionData have two very different meanings. The first one records the address of a section in the process memory, while the second one "load" the contents of a section into lldb memory (from whereever). The second one should work regardless of whether the first one was called. This is why you are able to inspect the debug info of an executable before actually running it.

File addresses can uniquely identify a single section, there is no problem with this, and there is always a single code section per module. The only "weirdness" is that since DWARF code addresses for Wasm are calculated from the beginning of the Code section, not the beginning of the file, for the Code section, Section::m_file_offset can normally be the file offset, but Section::m_file_addr needs to be zero. This seems to make all DWARF-related code work, but, as Pavel said, maybe there could be places where LLDB expects the "load bias" to be the same for each section, which could cause problems?

The basic section loading machinery can handle sections which are "shuffled" around, but this is not true of everything (because this is not how typical object file formats work). Given that you only have one code section (no debug info or symbols should point into the debug sections) I think you should be mostly fine.

In fact it would be possible to organize things such that the "load bias" is a constant, if we create an additional pseudo-section for the file header (like we do for COFF) with a negative file address. The layout would them look something like this

/------------\
|   header   |  file_addr = -sizeof(header)
|------------|
|   code     |  file_addr = 0
|------------|
| debug_info |  file_addr = offsetof(debug_info) - sizeof(header)
\------------/

This would keep the code section at address zero, and after applying a load bias of module_id | sizeof(header), everything would land in the right place. The reason I haven't proposed that is because that gets a bit messy, and so it seems acceptable to just do what you do now, provided it ends up working.

[clayborg]

In D72751#1846385, @labath wrote:

Thanks. My hopefully final question is not really for you but more like for other lldb developers (@jingham, @clayborg, etc.).

Given that this plugin is now consisting of boiler plate only, I am wondering if we should not instead make it possible for this use case to work without any special plugins needed. A couple of options that come to mind are:

• make the base DynamicLoader class instantiatable, and use it whenever we fail to find a specialized plugin
• same as above, but only do that for ProcessGDBRemote instances
• make ProcessGDBRemote call LoadModules() itself if no dynamic loader instance is available

WDYT?

I am fine with 1 as long as we document the DynamicLoader class to say that it will call Process::LoadModules() and will be used if no specialized loader is needed for your platform. I would like to a see a solution that will work for any process plug-in and not just ProcessGDBRemote. If we change solution 3 above to say "Make lldb_private::Process call LoadModules() itself if no dynamic loader instance is available" then solution 3 is also fine.

In D72751#1843458, @paolosev wrote:

Yes, this seems to be the case with the current implementation of ObjectFileWASM. It creates the section list in ObjectFileWasm::SetLoadAddress which calls Target::SetSectionLoadAddress but the sections don't need to be fully loaded, and during SymbolFileDWARF::CalculateAbilities(...) ObjectFile::ReadSectionData is called to load the necessary data.

This is correct, but I want to point out that the "load" in SetLoadAddress and in ReadSectionData have two very different meanings. The first one records the address of a section in the process memory, while the second one "load" the contents of a section into lldb memory (from whereever). The second one should work regardless of whether the first one was called. This is why you are able to inspect the debug info of an executable before actually running it.

File addresses can uniquely identify a single section, there is no problem with this, and there is always a single code section per module. The only "weirdness" is that since DWARF code addresses for Wasm are calculated from the beginning of the Code section, not the beginning of the file, for the Code section, Section::m_file_offset can normally be the file offset, but Section::m_file_addr needs to be zero. This seems to make all DWARF-related code work, but, as Pavel said, maybe there could be places where LLDB expects the "load bias" to be the same for each section, which could cause problems?

The basic section loading machinery can handle sections which are "shuffled" around, but this is not true of everything (because this is not how typical object file formats work). Given that you only have one code section (no debug info or symbols should point into the debug sections) I think you should be mostly fine.

In fact it would be possible to organize things such that the "load bias" is a constant, if we create an additional pseudo-section for the file header (like we do for COFF) with a negative file address. The layout would them look something like this

/------------\
|   header   |  file_addr = -sizeof(header)
|------------|
|   code     |  file_addr = 0
|------------|
| debug_info |  file_addr = offsetof(debug_info) - sizeof(header)
\------------/

This would keep the code section at address zero, and after applying a load bias of module_id | sizeof(header), everything would land in the right place. The reason I haven't proposed that is because that gets a bit messy, and so it seems acceptable to just do what you do now, provided it ends up working.

Yes the current approach allows anyone to load any section at any address. On Darwin systems, the DYLD shared cache will move TEXT, DATA, and other sections around such that all TEXT sections from all shared libraries in the shared cache are all in the one contiguous range. The slide is different for each section, so we have some nice flexibility with being able to set the section load address individually. They will even invert the memory order sometimes where in the file we have TEXT followed by DATA, but in the shared cache DATA appears at a lower address than __TEXT. We currently don't have the ability to load the same section at multiple addresses. This can happen when a shared library is loaded multiple times in memory, which we have seen on Android where a vendor will have a file that is the same as the base system, and the same exact file in loaded, albeit from different paths.

[paolosev]

Regarding:

• make the base DynamicLoader class instantiatable, and use it whenever we fail to find a specialized plugin
• same as above, but only do that for ProcessGDBRemote instances
• make ProcessGDBRemote call LoadModules() itself if no dynamic loader instance is available WDYT?

I am fine with 1 as long as we document the DynamicLoader class to say that it will call Process::LoadModules() and will be used if no specialized loader is needed for your platform. I would like to a see a solution that will work for any process plug-in and not just ProcessGDBRemote. If we change solution 3 above to say "Make lldb_private::Process call LoadModules() itself if no dynamic loader instance is available" then solution 3 is also fine.

there is a problem: if I remove DynamicLoaderWasmDYLD what happens is that DynamicLoaderStatic is found as a valid loader for a triple like "wasm32-unknown-unknown-wasm" because the Triple::OS is llvm::Triple::UnknownOS (I found out the hard way when I was registering DynamicLoaderWasmDYLD after DynamicLoaderStatic :-)).
There is an explicit check for UnknownOS:

DynamicLoader *DynamicLoaderStatic::CreateInstance(Process *process,
                                                   bool force) {
  bool create = force;
  if (!create) {
    const llvm::Triple &triple_ref =
        process->GetTarget().GetArchitecture().GetTriple();
    const llvm::Triple::OSType os_type = triple_ref.getOS();
    if ((os_type == llvm::Triple::UnknownOS))
      create = true;
  }
  ...

call stack:
DynamicLoaderStatic::CreateInstance(lldb_private::Process * process, bool force) Line 29
DynamicLoader::FindPlugin(lldb_private::Process * process, const char * plugin_name) Line 52
lldb_private::process_gdb_remote::ProcessGDBRemote::GetDynamicLoader() Line 3993
lldb_private::Process::CompleteAttach() Line 2931
lldb_private::Process::ConnectRemote(lldb_private::Stream * strm, llvm::StringRef remote_url) Line 3022

Could ProcessGDBRemote::GetDynamicLoader behave differently just when the architecture is wasm32, maybe? But then it is probably cleaner to add this plugin class, what do you think?

[labath]

In D72751#1847617, @clayborg wrote:

Yes the current approach allows anyone to load any section at any address. On Darwin systems, the DYLD shared cache will move TEXT, DATA, and other sections around such that all TEXT sections from all shared libraries in the shared cache are all in the one contiguous range. The slide is different for each section, so we have some nice flexibility with being able to set the section load address individually. They will even invert the memory order sometimes where in the file we have TEXT followed by DATA, but in the shared cache DATA appears at a lower address than __TEXT.

Sorry about the off-topic, but I found this bit very interesting. Greg, how does this work with code referencing the variables in the data section (e.g. static int x; int *f() { return &x; }). This code on elf, even when linked in fully position-independent mode will not contain any relocations because it is assumed that the dynamic loader will not change the relative layout of code and data (and so the code can use pc-relative addressing). This obviously does not work if data can be moved around independently. Does that mean that darwin will include some additional relocations which need to be resolved at load time?

[labath]

In D72751#1848384, @paolosev wrote:

Regarding:

• make the base DynamicLoader class instantiatable, and use it whenever we fail to find a specialized plugin
• same as above, but only do that for ProcessGDBRemote instances
• make ProcessGDBRemote call LoadModules() itself if no dynamic loader instance is available WDYT?

I am fine with 1 as long as we document the DynamicLoader class to say that it will call Process::LoadModules() and will be used if no specialized loader is needed for your platform. I would like to a see a solution that will work for any process plug-in and not just ProcessGDBRemote. If we change solution 3 above to say "Make lldb_private::Process call LoadModules() itself if no dynamic loader instance is available" then solution 3 is also fine.

there is a problem: if I remove DynamicLoaderWasmDYLD what happens is that DynamicLoaderStatic is found as a valid loader for a triple like "wasm32-unknown-unknown-wasm" because the Triple::OS is llvm::Triple::UnknownOS (I found out the hard way when I was registering DynamicLoaderWasmDYLD after DynamicLoaderStatic :-)).

...

call stack:
DynamicLoaderStatic::CreateInstance(lldb_private::Process * process, bool force) Line 29
DynamicLoader::FindPlugin(lldb_private::Process * process, const char * plugin_name) Line 52
lldb_private::process_gdb_remote::ProcessGDBRemote::GetDynamicLoader() Line 3993
lldb_private::Process::CompleteAttach() Line 2931
lldb_private::Process::ConnectRemote(lldb_private::Stream * strm, llvm::StringRef remote_url) Line 3022

Could ProcessGDBRemote::GetDynamicLoader behave differently just when the architecture is wasm32, maybe? But then it is probably cleaner to add this plugin class, what do you think?

Well.. I think DynamicLoaderStatic is being too grabby. :)

However, when I though about that idea further, I realized that any default plugin we could implement this way could not be complete, as we would be missing the part which (un)loads modules when a new shared library (dis)appears. This is something that cannot be done in a generic way, as that usually requires setting breakpoint on some special symbol, or getting notifications about shared library events in some other way. I don't know whether this is something that you will also need/plan to implement for wasm, or if one can assume that all modules are loaded from the get-go, but this is what convinced me that putting this in a separate plugin is fine.

[labath]

lgtm, per the previous comment. A couple of additional inline comments in the test.

lldb/packages/Python/lldbsuite/test/functionalities/gdb_remote_client/TestWasm.py

78

maybe also check that addr >= self.load_address. I wouldn't be surprised if lldb (now or in the future) decides it wants to try reading some other parts of memory too... (and we should return an error instead of falling over in that case).

107–123

I'm not sure how much we can rely on yaml2obj offsets not changing. Hard-coding the order of sections is fine, but maybe if would be better to check something like section.GetFileOffset() == 0x400...0 | section.GetFileOffset(). Since section parsing is checked elsewhere, the main thing we want to ensure here is that the 0x400...0 thingy is plumbed through correctly

[clayborg]

In D72751#1848880, @labath wrote:

In D72751#1847617, @clayborg wrote:

Yes the current approach allows anyone to load any section at any address. On Darwin systems, the DYLD shared cache will move TEXT, DATA, and other sections around such that all TEXT sections from all shared libraries in the shared cache are all in the one contiguous range. The slide is different for each section, so we have some nice flexibility with being able to set the section load address individually. They will even invert the memory order sometimes where in the file we have TEXT followed by DATA, but in the shared cache DATA appears at a lower address than __TEXT.

Sorry about the off-topic, but I found this bit very interesting. Greg, how does this work with code referencing the variables in the data section (e.g. static int x; int *f() { return &x; }). This code on elf, even when linked in fully position-independent mode will not contain any relocations because it is assumed that the dynamic loader will not change the relative layout of code and data (and so the code can use pc-relative addressing). This obviously does not work if data can be moved around independently. Does that mean that darwin will include some additional relocations which need to be resolved at load time?

No, the relocations are performed when the shared cache is made and they are removed! This also works for PLT calls from one shared library to another. If two shared libraries have PLT entries to each other's functions, those are resolved and don't require a call to the dynamic loader! There are two shared caches: one for running only and one for development. The development shared caches leave PLT entries alone so that interposing can happen.

[clayborg]

there is a problem: if I remove DynamicLoaderWasmDYLD what happens is that DynamicLoaderStatic is found as a valid loader for a triple like "wasm32-unknown-unknown-wasm" because the Triple::OS is llvm::Triple::UnknownOS (I found out the hard way when I was registering DynamicLoaderWasmDYLD after DynamicLoaderStatic :-)).
There is an explicit check for UnknownOS:

DynamicLoader *DynamicLoaderStatic::CreateInstance(Process *process,
                                                   bool force) {
  bool create = force;
  if (!create) {
    const llvm::Triple &triple_ref =
        process->GetTarget().GetArchitecture().GetTriple();
    const llvm::Triple::OSType os_type = triple_ref.getOS();
    if ((os_type == llvm::Triple::UnknownOS))
      create = true;
  }
  ...

call stack:
DynamicLoaderStatic::CreateInstance(lldb_private::Process * process, bool force) Line 29
DynamicLoader::FindPlugin(lldb_private::Process * process, const char * plugin_name) Line 52
lldb_private::process_gdb_remote::ProcessGDBRemote::GetDynamicLoader() Line 3993
lldb_private::Process::CompleteAttach() Line 2931
lldb_private::Process::ConnectRemote(lldb_private::Stream * strm, llvm::StringRef remote_url) Line 3022

Could ProcessGDBRemote::GetDynamicLoader behave differently just when the architecture is wasm32, maybe? But then it is probably cleaner to add this plugin class, what do you think?

So if you really have no OS, and no vendor, then this plug-in does get used. Why? Because it assumes that things will be loaded where they are (file_addr == load_addr). Could you create your WASM stuff in a way such that all addresses that are in a memory loaded object file have the file address the same as the load address? Do you ever actually have files on disk that we load for WASM? Or is it always loaded from memory?

It is ok to add your plug-in that recognizes WASM with no OS and no vendor, you will just need to make sure that the plug-in is registered before DynamicLoaderStatic. Since the plug-in manager will run through each plug-in linearly when looking for a match. This is also why any "auto registration of plug-ins" will cause problems as if we don't control the plug-in registration order, we will have issues that will arise. That is a different topic though, but I thought I remembered seeing a patch that tried to do this.

[paolosev]

Fixed the tests as suggested, and also added a couple more tests, to cover both the case where the Wasm module is loaded from memory and the case where it is loaded from a file.

However, when I though about that idea further, I realized that any default plugin we could implement this way could not be complete, as we would be missing the part which (un)loads modules when a new shared library (dis)appears. This is something that cannot be done in a generic way, as that usually requires setting breakpoint on some special symbol, or getting notifications about shared library events in some other way. I don't know whether this is something that you will also need/plan to implement for wasm, or if one can assume that all modules are loaded from the get-go, but this is what convinced me that putting this in a separate plugin is fine.

Yes, we'll need to support the case where a Wasm module is loaded at runtime, I think ProcessGDBRemote already does most of the work when it receives a stop event that contains the "library" flag and calls LoadModules() again. But I need to test this carefully.

[paolosev]

Rebasing.

[paolosev]

Fixing ObjectFile/wasm tests.

[JDevlieghere]

It looks like the wasm-DYLD directory is missing. I removed it again from the CMake file but now it's failing to build. Can you please take a look?

[JDevlieghere]

In D72751#1860780, @JDevlieghere wrote:

It looks like the wasm-DYLD directory is missing. I removed it again from the CMake file but now it's failing to build. Can you please take a look?

I've partially reverted your change and XFAILed the tests in

commit 4697e701b8cb40429818609814c7422e49b2ee07 (HEAD -> master, origin/master)
Author: Jonas Devlieghere <[e-mail]>
Date:   Wed Feb 5 15:30:11 2020 -0800

    Partially revert "[LLDB] Add DynamicLoaderWasmDYLD plugin for WebAssembly debugging"

    This temporarily and partially reverts 3ec28da6d643 because it's missing
    a directory.

[dschuff]

My bad, sorry about that.

[dschuff]

fixed in rGf5f70d1c8

[JDevlieghere]

In D72751#1860871, @dschuff wrote:

fixed in rGf5f70d1c8

Thank you!

[paolosev]

Thank you Derek, Jonas; I am sorry for all the trouble...

[JDevlieghere]

In D72751#1860901, @paolosev wrote:

Thank you Derek, Jonas; I am sorry for all the trouble...

No worries, thank you both for the quick turnaround time!

[max-kudr]

There is Windows Build Bot failure http://lab.llvm.org:8011/builders/lldb-x64-windows-ninja/builds/13427. Can you please fix or revert it?

Cannot open include file: 'Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.h': No such file or directory

[labath]

I am afraid this isn't over yet. :(

The tests with this patch don't seem to be compatible with python3 and are failing due to various errors: http://lab.llvm.org:8011/builders/lldb-x86_64-debian/builds/4419/steps/test/logs/stdio. Even when running the test with python2 I have had one failure due to the assert on TestWasm.py:233. I am not sure how this managed to pass for you (I guess we must have some host differences leaking in here), but overall, I think this is an overly aggressive assertion. You can't assume that lldb will read absolutely no memory, even when the module is read from disk. Maybe you could just check that none of the reads overlap the debug_info section? (I am deliberately not including the text section here, because lldb prefers to read code from memory instead of from object file).

Since the situation on master was starting to get a bit out of hand (multiple concurrent breakages with frantic fixup attempts), I've tried to back everything out so we can start with a clean slate again.

I am also sorry for not looking over the latest round of changes. I'll check out the changes you've since I've approved the patch tomorrow.

[paolosev]

In D72751#1860950, @max-kudr wrote:

There is Windows Build Bot failure http://lab.llvm.org:8011/builders/lldb-x64-windows-ninja/builds/13427. Can you please fix or revert it?

Cannot open include file: 'Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.h': No such file or directory

This should have been fixed by https://reviews.llvm.org/rGf5f70d1c8fbf12249b4b9598f10a10f12d4db029.

[labath]

As promised, here are the comments on the new tests. I think that most of the py3 incompatibilities will go away once we get rid of the yaml preprocessing step, but it would be good to verify this with python 3 nonetheless...

lldb/packages/Python/lldbsuite/test/functionalities/gdb_remote_client/TestWasm.py

9

this should be available as lldb.LLDB_INVALID_ADDRESS

174–180

a simpler way to handle this would be to put just the bare file name (no path) into the yaml, and then add the build directory to the target.debug-file-search-paths setting.

234

As I said in the previous comment, this needs to be relaxed a bit. Maybe you could just always return an error. This way we can be sure that the file is not accidentally read from memory but spurious memory reads be lldb will not cause the test to fail.

[paolosev]

Modified tests to be compatible with Python3.

[paolosev]

In D72751#1862140, @labath wrote:

As promised, here are the comments on the new tests. I think that most of the py3 incompatibilities will go away once we get rid of the yaml preprocessing step, but it would be good to verify this with python 3 nonetheless...

Thank you! Removing the yaml preprocessing indeed simplify everything, now tests should work both with Python 2 and 3.

[labath]

Ok, let's give this one more try. I have a couple of inline comments for the further simplification of the test case.

lldb/packages/Python/lldbsuite/test/functionalities/gdb_remote_client/TestWasm.py

61

you don't need the thread-pcs: part here, when you're implementing readRegister. thread-pcs is a preformance optimization, but we don't care about that in a test.

192–202

Could you remove this class (you can e.g. make the module name configurable via the object constructor)

[paolosev]

In D72751#1864138, @labath wrote:

Ok, let's give this one more try. I have a couple of inline comments for the further simplification of the test case.

Thank you again, I fixed the tests. If it's ok for you, I will then merge this patch (I got commit access now :)).

[labath]

yeah, go for it.

[paolosev]

Fix patch after rebasing:

• Use macros to initialize plugins in SystemInitializer.
• Move tests from /packages/Python/lldbsuite/test/functionalities/gdb_remote_client/ to /test/API/functionalities/gdb_remote_client/

[paolosev]

In D72751#1869268, @labath wrote:

yeah, go for it.

Hi @labath, can I ask you the favor to land this patch for me? I have rebased it because the logic of SystemInitializers have changed with new macros.

I asked commit access and I should have obtained it, but when I try to land this patch either with 'arc land' or with 'git push' I get a permission error:

git push --dry-run
Password for 'https://[e-mail]':
remote: Permission to llvm/llvm-project.git denied to paolosevMSFT.
fatal: unable to access 'https://[e-mail]/llvm/llvm-project.git/': The requested URL returned error: 403

Maybe there is something I have not understood in the process...

[labath]

In D72751#1877432, @paolosev wrote:

In D72751#1869268, @labath wrote:

yeah, go for it.

Hi @labath, can I ask you the favor to land this patch for me? I have rebased it because the logic of SystemInitializers have changed with new macros.

Committed as c1121908aace019b3e31e24def58a21a978531cd.

I asked commit access and I should have obtained it, but when I try to land this patch either with 'arc land' or with 'git push' I get a permission error:

git push --dry-run
Password for 'https://[e-mail]':
remote: Permission to llvm/llvm-project.git denied to paolosevMSFT.
fatal: unable to access 'https://[e-mail]/llvm/llvm-project.git/': The requested URL returned error: 403

Maybe there is something I have not understood in the process...

"arc land" never worked. Don't even try it -- all it can do is mess up your local repo. :)

The "git push" error looks like a generic problem with authenticating to github. I'd try creating a dummy personal repo and seeing if you can push there first. Switching to a different transport protocol ([e-mail] + authentication via ssh keys) is also worth a shot.

[JDevlieghere]

lldb/source/API/SystemInitializerFull.cpp

284

What's the rationale here? Plugins shouldn't rely on the order in which they are initialized. This breaks when the initializers are auto generated. Can we remove this dependency?

[paolosev]

@JDevlieghere added inline comments:

lldb/source/API/SystemInitializerFull.cpp 244
What's the rationale here? Plugins shouldn't rely on the order in which they are initialized. This breaks when the initializers are auto generated. Can we remove this dependency?

This was discussed in one of the comments above. If DynamicLoaderStatic preceeds DynamicLoaderWasmDYLD then it is recognized as a valid loader for a triple like "wasm32-unknown-unknown-wasm" because the Triple::OS is llvm::Triple::UnknownOS.
There is an explicit check for UnknownOS in DynamicLoaderStatic::CreateInstance().
Should DynamicLoaderStatic::CreateInstance behave differently just when the architecture is wasm32, as a workaround?

[JDevlieghere]

In D72751#1879806, @paolosev wrote:

@JDevlieghere added inline comments:

lldb/source/API/SystemInitializerFull.cpp 244
What's the rationale here? Plugins shouldn't rely on the order in which they are initialized. This breaks when the initializers are auto generated. Can we remove this dependency?

This was discussed in one of the comments above. If DynamicLoaderStatic preceeds DynamicLoaderWasmDYLD then it is recognized as a valid loader for a triple like "wasm32-unknown-unknown-wasm" because the Triple::OS is llvm::Triple::UnknownOS.
There is an explicit check for UnknownOS in DynamicLoaderStatic::CreateInstance().
Should DynamicLoaderStatic::CreateInstance behave differently just when the architecture is wasm32, as a workaround?

I think it depends on whether the DynamicLoaderStatic should be a fallback. If it doesn't make sense then yes, I think we should reject that triple there.

[paolosev]

I think it depends on whether the DynamicLoaderStatic should be a fallback. If it doesn't make sense then yes, I think we should reject that triple there.

It should not be a fallback. Ok! I'll create a new patch with this change.

[labath]

I think we should just have DynamicLoaderStatic disqualify itself for wasm files -- using it will never work there, so why should it pretend to support them...

[clayborg]

In D72751#1880120, @labath wrote:

I think we should just have DynamicLoaderStatic disqualify itself for wasm files -- using it will never work there, so why should it pretend to support them...

The triple clearly has an environment set of "wasm" so it should be easy to detect this in the DynamicLoaderStatic and stop it from saying it can handle it. The main issue then is do we have DynamicLoaderStatic say it can't handle anything if any environment is set, or just not for "wasm" (if OS and vendor are not set either).

[labath]

In D72751#1880975, @clayborg wrote:

The triple clearly has an environment set of "wasm" so it should be easy to detect this in the DynamicLoaderStatic and stop it from saying it can handle it. The main issue then is do we have DynamicLoaderStatic say it can't handle anything if any environment is set, or just not for "wasm" (if OS and vendor are not set either).

"Wasm" is an "architecture" (and an object file format), not an environment, but yeah, I think it should just check for wasm specifically for now. If we see a pattern developing later, we can change that...

[clayborg]

If we switch to auto registration of plug-in in the near future there are many things we need to watch out for and some notion of ordering needs to happen. A few examples:

• SymbolFileDWARFDebugMap and SymbolFileDWARF. Right now SymbolFileDWARF comes first and will claim a file before SymbolFileDWARFDebugMap. If SymbolFileDWARFDebugMap comes it will waste time iterating over all symbols in the symbol table, which causes the entire symbol table to be pulled in, and looks linearly for a symbol with type lldb::eSymbolTypeObjectFile, It will also claims it can parse the debug info just as well as SymbolFileDWARF, so it might end up getting used even though we have a dSYM file. So we need a way to avoid this. Right now ordering is used I believe. So the SymbolFileDWARFDebugMap::CalculateAbilities() might need to check if SymbolFileDWARF::CalculateAbilities() has all the abilities first, and just respond with no abilities in that case.
• SymbolFileSymtab and any other SymbolFile. SymbolFileSymtab will spend time going through the symbol table looking for symbols that describe any debug info. We want this to come last all the time and use this plug-in as a last resort.
• ObjectFile plug-ins can be ordered more efficiently for the current system (have ObjectFileMachO come first on Apple hosts, ObjectFileELF for non windows, ObjectFileCOFF for windows, etc). Not required, but would be nice.
• DynamicLoaderStatic vs any target that doesn't have OS, vendor or environment set.

, we might need to bolster the plug-in system a bit to be more like the SymbolFile plug-ins. For symbol file plug-ins we have each on calculate abilities and say

[JDevlieghere]

I went through the DYLD plugins and only the WASM and Hexagon plugin check the ArchType rather than the OSType, so I've created a patch to explicitly reject those in the DynamicLoaderStatic plugin: https://reviews.llvm.org/D74780

[vsk]

Hi @paolosev, the lldb sanitized bot is flagging a container-overflow error here. I know that this /can/ have FPs when sanitized and unsanitized code is mixed, but we should be in purely sanitized code here, and this looks like a valid report. PTAL.

http://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake-sanitized/992/testReport/junit/lldb-api/functionalities_gdb_remote_client/TestWasm_py/

=================================================================
==11283==ERROR: AddressSanitizer: container-overflow on address 0x615000016184 at pc 0x00010b4608f0 bp 0x7ffee4f00130 sp 0x7ffee4eff8f8
READ of size 512 at 0x615000016184 thread T0
    #0 0x10b4608ef in __asan_memcpy+0x1af (libclang_rt.asan_osx_dynamic.dylib:x86_64+0x418ef)
    #1 0x1116486d5 in lldb_private::MemoryCache::Read(unsigned long long, void*, unsigned long, lldb_private::Status&) Memory.cpp:189
    #2 0x11119d0e9 in lldb_private::Module::GetMemoryObjectFile(std::__1::shared_ptr<lldb_private::Process> const&, unsigned long long, lldb_private::Status&, unsigned long) Module.cpp:298
    #3 0x11169eeef in lldb_private::Process::ReadModuleFromMemory(lldb_private::FileSpec const&, unsigned long long, unsigned long) Process.cpp:2402
    #4 0x11113337b in lldb_private::DynamicLoader::LoadModuleAtAddress(lldb_private::FileSpec const&, unsigned long long, unsigned long long, bool) DynamicLoader.cpp:212
    #5 0x111ed53da in lldb_private::process_gdb_remote::ProcessGDBRemote::LoadModuleAtAddress(lldb_private::FileSpec const&, unsigned long long, unsigned long long, bool) ProcessGDBRemote.cpp:4767
    #6 0x111ed59b8 in lldb_private::process_gdb_remote::ProcessGDBRemote::LoadModules() ProcessGDBRemote.cpp:4801
    #7 0x1119c59aa in lldb_private::wasm::DynamicLoaderWasmDYLD::DidAttach() DynamicLoaderWasmDYLD.cpp:63
    #8 0x1116a3a97 in lldb_private::Process::CompleteAttach() Process.cpp:2930
    #9 0x1116a6bdf in lldb_private::Process::ConnectRemote(lldb_private::Stream*, llvm::StringRef) Process.cpp:3015
    #10 0x110a362ee in lldb::SBTarget::ConnectRemote(lldb::SBListener&, char const*, char const*, lldb::SBError&) SBTarget.cpp:559

[paolosev]

In D72751#1892514, @vsk wrote:

Hi @paolosev, the lldb sanitized bot is flagging a container-overflow error here. I know that this /can/ have FPs when sanitized and unsanitized code is mixed, but we should be in purely sanitized code here, and this looks like a valid report. PTAL.

http://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake-sanitized/992/testReport/junit/lldb-api/functionalities_gdb_remote_client/TestWasm_py/

Thanks for reporting!
Looking...

[paolosev]

Hi @paolosev, the lldb sanitized bot is flagging a container-overflow error here. I know that this /can/ have FPs when sanitized and unsanitized code is mixed, but we should be in purely sanitized code here, and this looks like a valid report. PTAL.

I think I might have found a problem with the existing code to cache memory read from a remote process. Looking at:

http://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake-sanitized/992/testReport/junit/lldb-api/functionalities_gdb_remote_client/TestWasm_py/

the error is easily explained. From ProcessGDBRemote::LoadModules we read the initial chunk (512 bytes) of an object file:

	_lldb.pyd!lldb_private::MemoryCache::Read(unsigned __int64 addr, void * dst, unsigned __int64 dst_len, lldb_private::Status & error) Line 239	C++
	_lldb.pyd!lldb_private::Process::ReadMemory(unsigned __int64 addr, void * buf, unsigned __int64 size, lldb_private::Status & error) Line 1953	C++
	_lldb.pyd!lldb_private::Module::GetMemoryObjectFile(const std::shared_ptr<lldb_private::Process> & process_sp, unsigned __int64 header_addr, lldb_private::Status & error, unsigned __int64 size_to_read) Line 300	C++
	_lldb.pyd!lldb_private::Process::ReadModuleFromMemory(const lldb_private::FileSpec & file_spec, unsigned __int64 header_addr, unsigned __int64 size_to_read) Line 2402	C++
	_lldb.pyd!lldb_private::DynamicLoader::LoadModuleAtAddress(const lldb_private::FileSpec & file, unsigned __int64 link_map_addr, unsigned __int64 base_addr, bool base_addr_is_offset) Line 212	C++
	_lldb.pyd!lldb_private::process_gdb_remote::ProcessGDBRemote::LoadModuleAtAddress(const lldb_private::FileSpec & file, unsigned __int64 link_map, unsigned __int64 base_addr, bool value_is_offset) Line 4769	C++
	_lldb.pyd!lldb_private::process_gdb_remote::ProcessGDBRemote::LoadModules() Line 4803	C++

In MemoryCache::Read, since this data is not cached yet, we call m_process.ReadMemoryFromInferior to actually read the memory (lines 174-241), look at the bottom of:

size_t MemoryCache::Read(addr_t addr, void *dst, size_t dst_len,
                         Status &error) {
    [...]
    while (bytes_left > 0) {
      [...]
      BlockMap::const_iterator pos = m_L2_cache.find(curr_addr);
      BlockMap::const_iterator end = m_L2_cache.end();

      if (pos != end) {
        size_t curr_read_size = cache_line_byte_size - cache_offset;
        if (curr_read_size > bytes_left)
          curr_read_size = bytes_left;

        memcpy(dst_buf + dst_len - bytes_left,
               pos->second->GetBytes() + cache_offset, curr_read_size);
        [...]
      }

      // We need to read from the process

      if (bytes_left > 0) {
        assert((curr_addr % cache_line_byte_size) == 0);
        std::unique_ptr<DataBufferHeap> data_buffer_heap_up(
            new DataBufferHeap(cache_line_byte_size, 0));
        size_t process_bytes_read = m_process.ReadMemoryFromInferior(
            curr_addr, data_buffer_heap_up->GetBytes(),
            data_buffer_heap_up->GetByteSize(), error);
        if (process_bytes_read == 0)
          return dst_len - bytes_left;

        if (process_bytes_read != cache_line_byte_size)
          data_buffer_heap_up->SetByteSize(process_bytes_read);
        m_L2_cache[curr_addr] = DataBufferSP(data_buffer_heap_up.release());
        // We have read data and put it into the cache, continue through the
        // loop again to get the data out of the cache...
      }

First, we allocate a DataBufferHeap with the size of our cache_line_byte_size, 512 bytes, and we pass it to ReadMemoryFromInferior().
The problem is that in this test the whole object file is only 0x84 bytes, so we resize data_buffer_heap_up to a smaller size with data_buffer_heap_up->SetByteSize(process_bytes_read).
Then we iterate back up in the while loop, and try to read from this reallocated buffer. But we still try to read curr_read_size==512 bytes, so read past the buffer size. In fact the overflow is at address 0x615000016184 for a buffer that starts at 0x615000016100.

This should be very simple to fix but the simple fix of just reading the available bytes doesn't work: Module::GetMemoryObjectFile expects to always be able to read by default 512 bytes, and it fails if the object file is smaller:

ObjectFile *Module::GetMemoryObjectFile(const lldb::ProcessSP &process_sp,
                                        lldb::addr_t header_addr, Status &error,
                                        size_t size_to_read) {
      [...]
      const size_t bytes_read =
          process_sp->ReadMemory(header_addr, data_up->GetBytes(),
                                 data_up->GetByteSize(), readmem_error);
      if (bytes_read == size_to_read) {
          [...] // ok...
      } else {
        error.SetErrorStringWithFormat("unable to read header from memory: %s",
                                       readmem_error.AsCString());
      }
      [...]

So probably, in order not to change the client code too much we should always read size_to_read bytes and pad the array with zeros if the file is smaller?

it's getting a bit late today but I should have a patch early tomorrow.
Who is the owner of this MemoryCache code I should ask for a review?

[labath]

Who is the owner of this MemoryCache code I should ask for a review?

It sounds like the fix will be fairly straightforward. I think anyone on this thread should be able to review that.

[paolosev]

In D72751#1892925, @labath wrote:

Who is the owner of this MemoryCache code I should ask for a review?

It sounds like the fix will be fairly straightforward. I think anyone on this thread should be able to review that.

Patch created: https://reviews.llvm.org/D75200.

lldb/packages/Python/lldbsuite/test/functionalities/gdb_remote_client/TestWasm.py

+import lldb
+import binascii
+import struct
+from lldbsuite.test.lldbtest import *
+from lldbsuite.test.decorators import *
+from gdbclientutils import *
+
+def format_register_value(val):
+    """
+    Encode each byte by two hex digits in little-endian order.
+    """
+    return ''.join(x.encode('hex') for x in struct.pack('<Q', val))
+
+class TestWasm(GDBRemoteTestBase):
+
+    def setUp(self):
+        super(TestWasm, self).setUp()
+        self._initial_platform = lldb.DBG.GetSelectedPlatform()
+
+    def tearDown(self):
+        lldb.DBG.SetSelectedPlatform(self._initial_platform)
+        super(TestWasm, self).tearDown()
+
+    def test_connect(self):
+
+        class MyResponder(MockGDBServerResponder):
+            load_address = 0x400000000
+            current_pc = 0x40000000a
+
+            def __init__(self, obj_path):
+                self._obj_path = obj_path
+                MockGDBServerResponder.__init__(self)
+
+            def respond(self, packet):
+                if packet == "qProcessInfo":
+                    return self.qProcessInfo()
+                if packet[0:13] == "qRegisterInfo":
+                    return self.qRegisterInfo(packet[13:])
+                return MockGDBServerResponder.respond(self, packet)
+
+            def qSupported(self, client_supported):
+                return "qXfer:libraries:read+;PacketSize=1000;vContSupported-"
+
+            def qHostInfo(self):
+                return ""
+
+            def QEnableErrorStrings(self):
+                return ""
+
+            def qfThreadInfo(self):
+                return "OK"
+
+            def qRegisterInfo(self, index):
+                if (index == 0):
+                    return "name:pc;alt-name:pc;bitsize:64;offset:0;encoding:uint;format:hex;set:General Purpose Registers;gcc:16;dwarf:16;generic:pc;"
+                return "E45"
+
+            def qProcessInfo(self):
+                return "pid:1;ppid:1;uid:1;gid:1;euid:1;egid:1;name:%s;triple:%s;ptrsize:4" % (hex_encode_bytes("lldb"), hex_encode_bytes("wasm32-unknown-unknown-wasm"))
+
+            def haltReason(self):
+                return "T05thread-pcs:" + format(self.current_pc, 'x') + ";thread:1;"
+
+            def readRegister(self, register):                
+                return format_register_value(self.current_pc)
+
+            def qXferRead(self, obj, annex, offset, length):
+                if obj == "libraries":
+                    xml = '<library-list><library name=\"%s\"><section address=\"%d\"/></library></library-list>' % ("test_wasm", self.load_address)
+                    return xml, False
+                else:
+                    return None, False
+
+            def readMemory(self, addr, length):
+                result = ""
+                with open(self._obj_path, mode='rb') as file:
+                    file_content = bytearray(file.read())                    
+                    addr_from = addr - self.load_address
+                    addr_to = addr_from + min(length, len(file_content) - addr_from)
+                    for i in range(addr_from, addr_to):
+                        result += format(file_content[i], '02x')
+                    file.close()
+                return result
+
+        """Test connecting to a WebAssembly engine via GDB-remote"""
+
+        yaml_path = "test_wasm.yaml"
+        yaml_base, ext = os.path.splitext(yaml_path)
+        obj_path = self.getBuildArtifact(yaml_base)
+        self.yaml2obj(yaml_path, obj_path)
+
+        self.server.responder = MyResponder(obj_path)
+
+        target = self.dbg.CreateTarget("")
+        process = self.connect(target)
+        lldbutil.expect_state_changes(self, self.dbg.GetListener(), process, [lldb.eStateStopped])
+    
+        num_modules = target.GetNumModules()
+        self.assertEquals(1, num_modules)
+
+        module = target.GetModuleAtIndex(0)
+        num_sections = module.GetNumSections()
+        self.assertEquals(5, num_sections)
+
+        code_section = module.GetSectionAtIndex(0)
+        self.assertEquals("code", code_section.GetName())
+        self.assertEquals(0x40000000a, code_section.GetLoadAddress(target))
+
+        debug_info_section = module.GetSectionAtIndex(1)
+        self.assertEquals(".debug_info", debug_info_section.GetName())
+        self.assertEquals(0x400000050, debug_info_section.GetLoadAddress(target))
+
+        debug_abbrev_section = module.GetSectionAtIndex(2)
+        self.assertEquals(".debug_abbrev", debug_abbrev_section.GetName())
+        self.assertEquals(0x400000062, debug_abbrev_section.GetLoadAddress(target))
+
+        debug_line_section = module.GetSectionAtIndex(3)
+        self.assertEquals(".debug_line", debug_line_section.GetName())
+        self.assertEquals(0x400000072, debug_line_section.GetLoadAddress(target))
+
+        debug_str_section = module.GetSectionAtIndex(4)
+        self.assertEquals(".debug_str", debug_str_section.GetName())
+        self.assertEquals(0x400000081, debug_str_section.GetLoadAddress(target))
+

lldb/packages/Python/lldbsuite/test/functionalities/gdb_remote_client/test_wasm.yaml

+--- !WASM
+FileHeader:
+  Version:         0x00000001
+Sections:
+
+  - Type:            CODE
+    Functions:
+      - Index:           0
+        Locals:
+          - Type:            I32
+            Count:           6
+        Body:            238080808000210141102102200120026B21032003200036020C200328020C2104200328020C2105200420056C210620060F0B
+  - Type:            CUSTOM
+    Name:            .debug_info
+    Payload:         4C00
+  - Type:            CUSTOM
+    Name:            .debug_abbrev
+    Payload:         0111
+  - Type:            CUSTOM
+    Name:            .debug_line
+    Payload:         5100
+  - Type:            CUSTOM
+    Name:            .debug_str
+    Payload:         636CFF
+...

lldb/source/API/SystemInitializerFull.cpp

 #include "Plugins/Architecture/Mips/ArchitectureMips.h"
 #include "Plugins/Architecture/PPC64/ArchitecturePPC64.h"
 #include "Plugins/Disassembler/llvm/DisassemblerLLVMC.h"
 #include "Plugins/DynamicLoader/MacOSX-DYLD/DynamicLoaderMacOS.h"
 #include "Plugins/DynamicLoader/MacOSX-DYLD/DynamicLoaderMacOSXDYLD.h"
 #include "Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.h"
 #include "Plugins/DynamicLoader/Static/DynamicLoaderStatic.h"
 #include "Plugins/DynamicLoader/Windows-DYLD/DynamicLoaderWindowsDYLD.h"
+#include "Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.h"
 #include "Plugins/Instruction/ARM/EmulateInstructionARM.h"
 #include "Plugins/Instruction/ARM64/EmulateInstructionARM64.h"
 #include "Plugins/Instruction/MIPS/EmulateInstructionMIPS.h"
 #include "Plugins/Instruction/MIPS64/EmulateInstructionMIPS64.h"
 #include "Plugins/Instruction/PPC64/EmulateInstructionPPC64.h"
 #include "Plugins/InstrumentationRuntime/ASan/ASanRuntime.h"
 #include "Plugins/InstrumentationRuntime/MainThreadChecker/MainThreadCheckerRuntime.h"
 #include "Plugins/InstrumentationRuntime/TSan/TSanRuntime.h"
#endif
 
   // This plugin is valid on any host that talks to a Darwin remote. It
   // shouldn't be limited to __APPLE__.
   StructuredDataDarwinLog::Initialize();
 
   // Platform agnostic plugins
   platform_gdb_server::PlatformRemoteGDBServer::Initialize();
 
   process_gdb_remote::ProcessGDBRemote::Initialize();
   DynamicLoaderMacOSXDYLD::Initialize();
   DynamicLoaderMacOS::Initialize();
   DynamicLoaderPOSIXDYLD::Initialize();
+  wasm::DynamicLoaderWasmDYLD::Initialize();
   DynamicLoaderStatic::Initialize();
   DynamicLoaderWindowsDYLD::Initialize();
 
   // Scan for any system or user LLDB plug-ins
   PluginManager::Initialize();
 
   // The process settings need to know about installed plug-ins, so the
   // Settings must be initialized
#endif
 
   platform_gdb_server::PlatformRemoteGDBServer::Terminate();
   process_gdb_remote::ProcessGDBRemote::Terminate();
   StructuredDataDarwinLog::Terminate();
 
   DynamicLoaderMacOSXDYLD::Terminate();
   DynamicLoaderMacOS::Terminate();
   DynamicLoaderPOSIXDYLD::Terminate();
+  wasm::DynamicLoaderWasmDYLD::Terminate();
   DynamicLoaderStatic::Terminate();
   DynamicLoaderWindowsDYLD::Terminate();
 
 #ifndef LLDB_DISABLE_PYTHON
   OperatingSystemPython::Terminate();
 #endif
 
   platform_freebsd::PlatformFreeBSD::Terminate();

lldb/source/Plugins/DynamicLoader/CMakeLists.txt

 add_subdirectory(Darwin-Kernel)
 add_subdirectory(MacOSX-DYLD)
 add_subdirectory(POSIX-DYLD)
 add_subdirectory(Static)
 add_subdirectory(Hexagon-DYLD)
 add_subdirectory(Windows-DYLD)
+add_subdirectory(wasm-DYLD)

lldb/source/Plugins/DynamicLoader/wasm-DYLD/CMakeLists.txt

+add_lldb_library(lldbPluginDynamicLoaderWasmDYLD PLUGIN
+  DynamicLoaderWasmDYLD.cpp
+
+  LINK_LIBS
+    lldbCore
+    lldbTarget
+  LINK_COMPONENTS
+    Support
+  )

lldb/source/Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.h

+//===-- DynamicLoaderWasmDYLD.h ---------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef liblldb_Plugins_DynamicLoaderWasmDYLD_h_
+#define liblldb_Plugins_DynamicLoaderWasmDYLD_h_
+
+#include "lldb/Target/DynamicLoader.h"
+
+namespace lldb_private {
+namespace wasm {
+
+class DynamicLoaderWasmDYLD : public DynamicLoader {
+public:
+  DynamicLoaderWasmDYLD(Process *process);
+
+  static void Initialize();
+  static void Terminate() {}
+
+  static ConstString GetPluginNameStatic();
+  static const char *GetPluginDescriptionStatic();
+
+  static DynamicLoader *CreateInstance(Process *process, bool force);
+
+  /// DynamicLoader
+  /// \{
+  void DidAttach() override;
+  void DidLaunch() override {}
+  Status CanLoadImage() override { return Status(); }
+  lldb::ThreadPlanSP GetStepThroughTrampolinePlan(Thread &thread,
+                                                  bool stop) override;
+  /// \}
+
+  /// PluginInterface protocol.
+  /// \{
+  ConstString GetPluginName() override { return GetPluginNameStatic(); }
+  uint32_t GetPluginVersion() override { return 1; }
+  /// \}
+};
+
+} // namespace wasm
+} // namespace lldb_private
+
+#endif // liblldb_Plugins_DynamicLoaderWasmDYLD_h_

lldb/source/Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.cpp

+//===-- DynamicLoaderWasmDYLD.cpp -----------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "DynamicLoaderWasmDYLD.h"
+
+#include "Plugins/ObjectFile/wasm/ObjectFileWasm.h"
+#include "lldb/Core/Module.h"
+#include "lldb/Core/PluginManager.h"
+#include "lldb/Core/Section.h"
+#include "lldb/Target/Process.h"
+#include "lldb/Target/Target.h"
+#include "lldb/Utility/Log.h"
+
+using namespace lldb;
+using namespace lldb_private;
+using namespace lldb_private::wasm;
+
+DynamicLoaderWasmDYLD::DynamicLoaderWasmDYLD(Process *process)
+    : DynamicLoader(process) {}
+
+void DynamicLoaderWasmDYLD::Initialize() {
+  PluginManager::RegisterPlugin(GetPluginNameStatic(),
+                                GetPluginDescriptionStatic(), CreateInstance);
+}
+
+ConstString DynamicLoaderWasmDYLD::GetPluginNameStatic() {
+  static ConstString g_plugin_name("wasm-dyld");
+  return g_plugin_name;
+}
+
+const char *DynamicLoaderWasmDYLD::GetPluginDescriptionStatic() {
+  return "Dynamic loader plug-in that watches for shared library "
+         "loads/unloads in WebAssembly engines.";
+}
+
+DynamicLoader *DynamicLoaderWasmDYLD::CreateInstance(Process *process,
+                                                     bool force) {
+  bool should_create = force;
+  if (!should_create) {
+    should_create =
+        (process->GetTarget().GetArchitecture().GetTriple().getArch() ==
+         llvm::Triple::wasm32);
+  }
+
+  if (should_create)
+    return new DynamicLoaderWasmDYLD(process);
+
+  return nullptr;
+}
+
+void DynamicLoaderWasmDYLD::DidAttach() {
+  Log *log(GetLogIfAnyCategoriesSet(LIBLLDB_LOG_DYNAMIC_LOADER));
+  LLDB_LOGF(log, "DynamicLoaderWasmDYLD::%s()", __FUNCTION__);
+
+  // Ask the process for the list of loaded WebAssembly modules.
+  auto error = m_process->LoadModules();
+  LLDB_LOG_ERROR(log, std::move(error), "Couldn't load modules: {0}");
+}
+
+ThreadPlanSP DynamicLoaderWasmDYLD::GetStepThroughTrampolinePlan(Thread &thread,
+                                                                 bool stop) {
+  return ThreadPlanSP();
+}

lldb/source/Plugins/ObjectFile/wasm/ObjectFileWasm.h

-//===-- ObjectFileWasm.h ----------------------------------------*- C++ -*-===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-
-#ifndef LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H
-#define LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H
-
-#include "lldb/Symbol/ObjectFile.h"
-#include "lldb/Utility/ArchSpec.h"
-
-namespace lldb_private {
-namespace wasm {
-
-/// Generic Wasm object file reader.
-///
-/// This class provides a generic wasm32 reader plugin implementing the
-/// ObjectFile protocol.
-class ObjectFileWasm : public ObjectFile {
-public:
-  static void Initialize();
-  static void Terminate();
-
-  static ConstString GetPluginNameStatic();
-  static const char *GetPluginDescriptionStatic() {
-    return "WebAssembly object file reader.";
-  }
-
-  static ObjectFile *
-  CreateInstance(const lldb::ModuleSP &module_sp, lldb::DataBufferSP &data_sp,
-                 lldb::offset_t data_offset, const FileSpec *file,
-                 lldb::offset_t file_offset, lldb::offset_t length);
-
-  static ObjectFile *CreateMemoryInstance(const lldb::ModuleSP &module_sp,
-                                          lldb::DataBufferSP &data_sp,
-                                          const lldb::ProcessSP &process_sp,
-                                          lldb::addr_t header_addr);
-
-  static size_t GetModuleSpecifications(const FileSpec &file,
-                                        lldb::DataBufferSP &data_sp,
-                                        lldb::offset_t data_offset,
-                                        lldb::offset_t file_offset,
-                                        lldb::offset_t length,
-                                        ModuleSpecList &specs);
-
-  /// PluginInterface protocol.
-  /// \{
-  ConstString GetPluginName() override { return GetPluginNameStatic(); }
-  uint32_t GetPluginVersion() override { return 1; }
-  /// \}
-
-  /// LLVM RTTI support
-  /// \{
-  static char ID;
-  bool isA(const void *ClassID) const override {
-    return ClassID == &ID || ObjectFile::isA(ClassID);
-  }
-  static bool classof(const ObjectFile *obj) { return obj->isA(&ID); }
-  /// \}
-
-  /// ObjectFile Protocol.
-  /// \{
-  bool ParseHeader() override;
-
-  lldb::ByteOrder GetByteOrder() const override {
-    return m_arch.GetByteOrder();
-  }
-
-  bool IsExecutable() const override { return true; }
-
-  uint32_t GetAddressByteSize() const override {
-    return m_arch.GetAddressByteSize();
-  }
-
-  AddressClass GetAddressClass(lldb::addr_t file_addr) override {
-    return AddressClass::eInvalid;
-  }
-
-  Symtab *GetSymtab() override;
-
-  bool IsStripped() override { return true; }
-
-  void CreateSections(SectionList &unified_section_list) override;
-
-  void Dump(Stream *s) override;
-
-  ArchSpec GetArchitecture() override { return m_arch; }
-
-  UUID GetUUID() override { return m_uuid; }
-
-  uint32_t GetDependentModules(FileSpecList &files) override { return 0; }
-
-  Type CalculateType() override { return eTypeExecutable; }
-
-  Strata CalculateStrata() override { return eStrataUser; }
-
-  bool SetLoadAddress(lldb_private::Target &target, lldb::addr_t value,
-                      bool value_is_offset) override;
-
-  lldb_private::Address GetBaseAddress() override {
-    return IsInMemory() ? Address(m_memory_addr + m_code_section_offset)
-                        : Address(m_code_section_offset);
-  }
-  /// \}
-
-  /// A Wasm module that has external DWARF debug information should contain a
-  /// custom section named "external_debug_info", whose payload is an UTF-8
-  /// encoded string that points to a Wasm module that contains the debug
-  /// information for this module.
-  llvm::Optional<FileSpec> GetExternalDebugInfoFileSpec();
-
-private:
-  ObjectFileWasm(const lldb::ModuleSP &module_sp, lldb::DataBufferSP &data_sp,
-                 lldb::offset_t data_offset, const FileSpec *file,
-                 lldb::offset_t offset, lldb::offset_t length);
-  ObjectFileWasm(const lldb::ModuleSP &module_sp,
-                 lldb::DataBufferSP &header_data_sp,
-                 const lldb::ProcessSP &process_sp, lldb::addr_t header_addr);
-
-  /// Wasm section decoding routines.
-  /// \{
-  bool DecodeNextSection(lldb::offset_t *offset_ptr);
-  bool DecodeSections();
-  /// \}
-
-  /// Read a range of bytes from the Wasm module.
-  DataExtractor ReadImageData(uint64_t offset, size_t size);
-
-  typedef struct section_info {
-    lldb::offset_t offset;
-    uint32_t size;
-    uint32_t id;
-    ConstString name;
-  } section_info_t;
-
-  /// Wasm section header dump routines.
-  /// \{
-  void DumpSectionHeader(llvm::raw_ostream &ostream, const section_info_t &sh);
-  void DumpSectionHeaders(llvm::raw_ostream &ostream);
-  /// \}
-
-  std::vector<section_info_t> m_sect_infos;
-  ArchSpec m_arch;
-  UUID m_uuid;
-  uint32_t m_code_section_offset;
-};
-
-} // namespace wasm
-} // namespace lldb_private
-#endif // LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H
+//===-- ObjectFileWasm.h ----------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H
+#define LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H
+
+#include "lldb/Symbol/ObjectFile.h"
+#include "lldb/Utility/ArchSpec.h"
+
+namespace lldb_private {
+namespace wasm {
+
+/// Generic Wasm object file reader.
+///
+/// This class provides a generic wasm32 reader plugin implementing the
+/// ObjectFile protocol.
+class ObjectFileWasm : public ObjectFile {
+public:
+  static void Initialize();
+  static void Terminate();
+
+  static ConstString GetPluginNameStatic();
+  static const char *GetPluginDescriptionStatic() {
+    return "WebAssembly object file reader.";
+  }
+
+  static ObjectFile *
+  CreateInstance(const lldb::ModuleSP &module_sp, lldb::DataBufferSP &data_sp,
+                 lldb::offset_t data_offset, const FileSpec *file,
+                 lldb::offset_t file_offset, lldb::offset_t length);
+
+  static ObjectFile *CreateMemoryInstance(const lldb::ModuleSP &module_sp,
+                                          lldb::DataBufferSP &data_sp,
+                                          const lldb::ProcessSP &process_sp,
+                                          lldb::addr_t header_addr);
+
+  static size_t GetModuleSpecifications(const FileSpec &file,
+                                        lldb::DataBufferSP &data_sp,
+                                        lldb::offset_t data_offset,
+                                        lldb::offset_t file_offset,
+                                        lldb::offset_t length,
+                                        ModuleSpecList &specs);
+
+  /// PluginInterface protocol.
+  /// \{
+  ConstString GetPluginName() override { return GetPluginNameStatic(); }
+  uint32_t GetPluginVersion() override { return 1; }
+  /// \}
+
+  /// LLVM RTTI support
+  /// \{
+  static char ID;
+  bool isA(const void *ClassID) const override {
+    return ClassID == &ID || ObjectFile::isA(ClassID);
+  }
+  static bool classof(const ObjectFile *obj) { return obj->isA(&ID); }
+  /// \}
+
+  /// ObjectFile Protocol.
+  /// \{
+  bool ParseHeader() override;
+
+  lldb::ByteOrder GetByteOrder() const override {
+    return m_arch.GetByteOrder();
+  }
+
+  bool IsExecutable() const override { return false; }
+
+  uint32_t GetAddressByteSize() const override {
+    return m_arch.GetAddressByteSize();
+  }
+
+  AddressClass GetAddressClass(lldb::addr_t file_addr) override {
+    return AddressClass::eInvalid;
+  }
+
+  Symtab *GetSymtab() override;
+
+  bool IsStripped() override { return !!GetExternalDebugInfoFileSpec(); }
+
+  void CreateSections(SectionList &unified_section_list) override;
+
+  void Dump(Stream *s) override;
+
+  ArchSpec GetArchitecture() override { return m_arch; }
+
+  UUID GetUUID() override { return m_uuid; }
+
+  uint32_t GetDependentModules(FileSpecList &files) override { return 0; }
+
+  Type CalculateType() override { return eTypeSharedLibrary; }
+
+  Strata CalculateStrata() override { return eStrataUser; }
+
+  bool SetLoadAddress(lldb_private::Target &target, lldb::addr_t value,
+                      bool value_is_offset) override;
+
+  lldb_private::Address GetBaseAddress() override {
+    return IsInMemory() ? Address(m_memory_addr) : Address(0);
+  }
+  /// \}
+
+  /// A Wasm module that has external DWARF debug information should contain a
+  /// custom section named "external_debug_info", whose payload is an UTF-8
+  /// encoded string that points to a Wasm module that contains the debug
+  /// information for this module.
+  llvm::Optional<FileSpec> GetExternalDebugInfoFileSpec();
+
+private:
+  ObjectFileWasm(const lldb::ModuleSP &module_sp, lldb::DataBufferSP &data_sp,
+                 lldb::offset_t data_offset, const FileSpec *file,
+                 lldb::offset_t offset, lldb::offset_t length);
+  ObjectFileWasm(const lldb::ModuleSP &module_sp,
+                 lldb::DataBufferSP &header_data_sp,
+                 const lldb::ProcessSP &process_sp, lldb::addr_t header_addr);
+
+  /// Wasm section decoding routines.
+  /// \{
+  bool DecodeNextSection(lldb::offset_t *offset_ptr);
+  bool DecodeSections();
+  /// \}
+
+  /// Read a range of bytes from the Wasm module.
+  DataExtractor ReadImageData(lldb::offset_t offset, uint32_t size);
+
+  typedef struct section_info {
+    lldb::offset_t offset;
+    uint32_t size;
+    uint32_t id;
+    ConstString name;
+  } section_info_t;
+
+  /// Wasm section header dump routines.
+  /// \{
+  void DumpSectionHeader(llvm::raw_ostream &ostream, const section_info_t &sh);
+  void DumpSectionHeaders(llvm::raw_ostream &ostream);
+  /// \}
+
+  std::vector<section_info_t> m_sect_infos;
+  ArchSpec m_arch;
+  UUID m_uuid;
+};
+
+} // namespace wasm
+} // namespace lldb_private
+#endif // LLDB_PLUGINS_OBJECTFILE_WASM_OBJECTFILEWASM_H

lldb/source/Plugins/ObjectFile/wasm/ObjectFileWasm.cpp

size_t ObjectFileWasm::GetModuleSpecifications(
   specs.Append(spec);
   return 1;
 }
 
 ObjectFileWasm::ObjectFileWasm(const ModuleSP &module_sp, DataBufferSP &data_sp,
                                offset_t data_offset, const FileSpec *file,
                                offset_t offset, offset_t length)
     : ObjectFile(module_sp, file, offset, length, data_sp, data_offset),
-      m_arch("wasm32-unknown-unknown-wasm"), m_code_section_offset(0) {
+      m_arch("wasm32-unknown-unknown-wasm") {
   m_data.SetAddressByteSize(4);
 }
 
 ObjectFileWasm::ObjectFileWasm(const lldb::ModuleSP &module_sp,
                                lldb::DataBufferSP &header_data_sp,
                                const lldb::ProcessSP &process_sp,
                                lldb::addr_t header_addr)
     : ObjectFile(module_sp, process_sp, header_addr, header_data_sp),
-      m_arch("wasm32-unknown-unknown-wasm"), m_code_section_offset(0) {}
+      m_arch("wasm32-unknown-unknown-wasm") {}
 
 bool ObjectFileWasm::ParseHeader() {
   // We already parsed the header during initialization.
   return true;
 }
 
 Symtab *ObjectFileWasm::GetSymtab() { return nullptr; }
 
 void ObjectFileWasm::CreateSections(SectionList &unified_section_list) {
   if (m_sections_up)
     return;
 
   m_sections_up = std::make_unique<SectionList>();
 
   if (m_sect_infos.empty()) {
     DecodeSections();
   }
 
   for (const section_info &sect_info : m_sect_infos) {
     SectionType section_type = eSectionTypeOther;
     ConstString section_name;
-    offset_t file_offset = 0;
-    addr_t vm_addr = 0;
-    size_t vm_size = 0;
+    offset_t file_offset = sect_info.offset & 0xffffffff;
+    addr_t vm_addr = file_offset;
+    size_t vm_size = sect_info.size;
 
     if (llvm::wasm::WASM_SEC_CODE == sect_info.id) {
       section_type = eSectionTypeCode;
       section_name = ConstString("code");
-      m_code_section_offset = sect_info.offset & 0xffffffff;
-      vm_size = sect_info.size;
+
+      // A code address in DWARF for WebAssembly is the offset of an
+      // instruction relative within the Code section of the WebAssembly file.
+      // For this reason Section::GetFileAddress() must return zero for the
+      // Code section.
+      vm_addr = 0;
     } else {
       section_type =
           llvm::StringSwitch<SectionType>(sect_info.name.GetStringRef())
               .Case(".debug_abbrev", eSectionTypeDWARFDebugAbbrev)
               .Case(".debug_addr", eSectionTypeDWARFDebugAddr)
               .Case(".debug_aranges", eSectionTypeDWARFDebugAranges)
               .Case(".debug_cu_index", eSectionTypeDWARFDebugCuIndex)
               .Case(".debug_frame", eSectionTypeDWARFDebugFrame)
if (llvm::wasm::WASM_SEC_CODE == sect_info.id) {
               .Case(".debug_rnglists", eSectionTypeDWARFDebugRngLists)
               .Case(".debug_str", eSectionTypeDWARFDebugStr)
               .Case(".debug_str_offsets", eSectionTypeDWARFDebugStrOffsets)
               .Case(".debug_types", eSectionTypeDWARFDebugTypes)
               .Default(eSectionTypeOther);
       if (section_type == eSectionTypeOther)
         continue;
       section_name = sect_info.name;
-      file_offset = sect_info.offset & 0xffffffff;
-      if (IsInMemory()) {
-        vm_addr = sect_info.offset & 0xffffffff;
-        vm_size = sect_info.size;
+      if (!IsInMemory()) {
+        vm_size = 0;
+        vm_addr = 0;
       }
     }
 
     SectionSP section_sp(
         new Section(GetModule(), // Module to which this section belongs.
                     this,        // ObjectFile to which this section belongs and
                                  // should read section data from.
                     section_type,   // Section ID.
bool ObjectFileWasm::SetLoadAddress(Target &target, lldb::addr_t load_address,
   /// where the lower 32 bits represent a module offset (relative to the module
   /// start not to the beginning of the code section) and the higher 32 bits
   /// uniquely identify the module in the WebAssembly VM.
   /// In other words, we assume that each WebAssembly module is loaded by the
   /// engine at a 64-bit address that starts at the boundary of 4GB pages, like
   /// 0x0000000400000000 for module_id == 4.
   /// These 64-bit addresses will be used to request code ranges for a specific
   /// module from the WebAssembly engine.
+  assert(m_memory_addr == load_address);
+
   ModuleSP module_sp = GetModule();
   if (!module_sp)
     return false;
 
   DecodeSections();
 
   size_t num_loaded_sections = 0;
   SectionList *section_list = GetSectionList();
   if (!section_list)
     return false;
 
   const size_t num_sections = section_list->GetSize();
-  size_t sect_idx = 0;
-
-  for (sect_idx = 0; sect_idx < num_sections; ++sect_idx) {
+  for (size_t sect_idx = 0; sect_idx < num_sections; ++sect_idx) {
     SectionSP section_sp(section_list->GetSectionAtIndex(sect_idx));
-    if (target.GetSectionLoadList().SetSectionLoadAddress(
-            section_sp, load_address | section_sp->GetFileAddress())) {
+    if (target.SetSectionLoadAddress(
+            section_sp, load_address | section_sp->GetFileOffset())) {
       ++num_loaded_sections;
     }
   }
 
   return num_loaded_sections > 0;
 }
 
-DataExtractor ObjectFileWasm::ReadImageData(uint64_t offset, size_t size) {
+DataExtractor ObjectFileWasm::ReadImageData(offset_t offset, uint32_t size) {
   DataExtractor data;
   if (m_file) {
     if (offset < GetByteSize()) {
-      size = std::min(size, GetByteSize() - offset);
+      size = std::min(static_cast<uint64_t>(size), GetByteSize() - offset);
       auto buffer_sp = MapFileData(m_file, size, offset);
       return DataExtractor(buffer_sp, GetByteOrder(), GetAddressByteSize());
     }
   } else {
     ProcessSP process_sp(m_process_wp.lock());
     if (process_sp) {
       auto data_up = std::make_unique<DataBufferHeap>(size, 0);
       Status readmem_error;

lldb/tools/lldb-test/SystemInitializerTest.cpp

 #include "Plugins/Architecture/Arm/ArchitectureArm.h"
 #include "Plugins/Architecture/PPC64/ArchitecturePPC64.h"
 #include "Plugins/Disassembler/llvm/DisassemblerLLVMC.h"
 #include "Plugins/DynamicLoader/MacOSX-DYLD/DynamicLoaderMacOS.h"
 #include "Plugins/DynamicLoader/MacOSX-DYLD/DynamicLoaderMacOSXDYLD.h"
 #include "Plugins/DynamicLoader/POSIX-DYLD/DynamicLoaderPOSIXDYLD.h"
 #include "Plugins/DynamicLoader/Static/DynamicLoaderStatic.h"
 #include "Plugins/DynamicLoader/Windows-DYLD/DynamicLoaderWindowsDYLD.h"
+#include "Plugins/DynamicLoader/wasm-DYLD/DynamicLoaderWasmDYLD.h"
 #include "Plugins/Instruction/ARM64/EmulateInstructionARM64.h"
 #include "Plugins/Instruction/PPC64/EmulateInstructionPPC64.h"
 #include "Plugins/InstrumentationRuntime/ASan/ASanRuntime.h"
 #include "Plugins/InstrumentationRuntime/MainThreadChecker/MainThreadCheckerRuntime.h"
 #include "Plugins/InstrumentationRuntime/TSan/TSanRuntime.h"
 #include "Plugins/InstrumentationRuntime/UBSan/UBSanRuntime.h"
 #include "Plugins/JITLoader/GDB/JITLoaderGDB.h"
 #include "Plugins/Language/CPlusPlus/CPlusPlusLanguage.h"
#endif
 
   // Platform agnostic plugins
   platform_gdb_server::PlatformRemoteGDBServer::Initialize();
 
   process_gdb_remote::ProcessGDBRemote::Initialize();
   DynamicLoaderMacOSXDYLD::Initialize();
   DynamicLoaderMacOS::Initialize();
   DynamicLoaderPOSIXDYLD::Initialize();
+  wasm::DynamicLoaderWasmDYLD::Initialize();
   DynamicLoaderStatic::Initialize();
   DynamicLoaderWindowsDYLD::Initialize();
 
   // Scan for any system or user LLDB plug-ins
   PluginManager::Initialize();
 
   // The process settings need to know about installed plug-ins, so the Settings
   // must be initialized
#endif
 
   platform_gdb_server::PlatformRemoteGDBServer::Terminate();
   process_gdb_remote::ProcessGDBRemote::Terminate();
   StructuredDataDarwinLog::Terminate();
 
   DynamicLoaderMacOSXDYLD::Terminate();
   DynamicLoaderMacOS::Terminate();
   DynamicLoaderPOSIXDYLD::Terminate();
+  wasm::DynamicLoaderWasmDYLD::Terminate();
   DynamicLoaderStatic::Terminate();
   DynamicLoaderWindowsDYLD::Terminate();
 
 
   platform_freebsd::PlatformFreeBSD::Terminate();
   platform_linux::PlatformLinux::Terminate();
   platform_netbsd::PlatformNetBSD::Terminate();
   platform_openbsd::PlatformOpenBSD::Terminate();