[AMDGPU] Support disassembly for AMDGPU kernel descriptors

[rochauha]

Summary

Decode AMDGPU Kernel descriptors as assembler directives.

[rochauha]

Minor change to the comment about Size.

[jhenderson]

In D80713#2108650, @rochauha wrote:

In D80713#2106033, @jhenderson wrote:

No idea about the functional logic of this code, but I do wonder whether you'd be better off dividing it into smaller patches for testability.

This patch is for entirely disassembling the kernel descriptor symbol. It happens to be so that many directives affect the kernel descriptor. I will also be adding a new test in this patch.

The problem is that you're trying to do it all at once. There's no need to implement the full disassembly at once. You can do it in a series of patches, one after the other that build towards the primary goal. For example, there are many if (<X is some value>) type cases, which you could omit - just assume those are all false, in earlier versions, and add them in (with corresponding testing) in a later patch. Similarly, you could assume that all results of x & y return some specific value, e.g. 0, and just print that for now. Yes, that means you won't support everything from the point at which this patch lands, but it will make each individual patch easier to reason with. This fits much better with LLVM's preferred approach - please see https://llvm.org/docs/DeveloperPolicy.html#incremental-development.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1264–1267

I don't see a response to this suggestion.

1325

next -> Next

1471

/*IsLittleEndian =*/true -> /*IsLittleEndian=*/true

(and same for AddressSize)

as already asked for once...

1661

Nit: missing full stop.

1663–1664

Put this comment above the if it is referring to, like the exisitng bit of comment.

[jhenderson]

Oh, you could also put the additional behaviour behind a switch in llvm-objdump. That'll mean you won't break all the existing tests too. Once you've finished the development, you could either remove the switch entirely or just change its default (which would allow users to disable the behaviour if they want the less verbose version of the output).

[rochauha]

• Updated to use Cursor. Now there isn't any neet to maintain CurrentIndex.
• Other small changes related to coding conventions.

[rochauha]

The problem is that you're trying to do it all at once. There's no need to implement the full disassembly at once. You can do it in a series of patches, one after the other that build towards the primary goal. For example, there are many if (<X is some value>) type cases, which you could omit - just assume those are all false, in earlier versions, and add them in (with corresponding testing) in a later patch. Similarly, you could assume that all results of x & y return some specific value, e.g. 0, and just print that for now. Yes, that means you won't support everything from the point at which this patch lands, but it will make each individual patch easier to reason with. This fits much better with LLVM's preferred approach - please see https://llvm.org/docs/DeveloperPolicy.html#incremental-development.

Oh, you could also put the additional behaviour behind a switch in llvm-objdump. That'll mean you won't break all the existing tests too. Once you've finished the development, you could either remove the switch entirely or just change its default (which would allow users to disable the behaviour if they want the less verbose version of the output).

Thanks for the feedback! I'll keep these points in mind :)

Regarding this patch - I think that because it has been reviewed to quite some extent, it doesn't need to be broken into smaller patches now?

[rochauha]

Following tests fail:

LLVM :: CodeGen/AMDGPU/call-encoding.ll
LLVM :: CodeGen/AMDGPU/s_code_end.ll
LLVM :: MC/AMDGPU/branch-comment.s
LLVM :: MC/AMDGPU/data.s
LLVM :: MC/AMDGPU/labels-branch-gfx9.s
LLVM :: MC/AMDGPU/labels-branch.s
LLVM :: MC/AMDGPU/offsetbug_once.s
LLVM :: MC/AMDGPU/offsetbug_one_and_one.s
LLVM :: MC/AMDGPU/offsetbug_twice.s
LLVM :: MC/AMDGPU/s_endpgm.s
LLVM :: Object/AMDGPU/objdump.s
LLVM :: tools/llvm-cov/ignore-filename-regex.test
LLVM :: tools/llvm-objdump/ELF/AMDGPU/source-lines.ll

These tests no longer fail.

[jhenderson]

In D80713#2116706, @rochauha wrote:

The problem is that you're trying to do it all at once. There's no need to implement the full disassembly at once. You can do it in a series of patches, one after the other that build towards the primary goal. For example, there are many if (<X is some value>) type cases, which you could omit - just assume those are all false, in earlier versions, and add them in (with corresponding testing) in a later patch. Similarly, you could assume that all results of x & y return some specific value, e.g. 0, and just print that for now. Yes, that means you won't support everything from the point at which this patch lands, but it will make each individual patch easier to reason with. This fits much better with LLVM's preferred approach - please see https://llvm.org/docs/DeveloperPolicy.html#incremental-development.

Oh, you could also put the additional behaviour behind a switch in llvm-objdump. That'll mean you won't break all the existing tests too. Once you've finished the development, you could either remove the switch entirely or just change its default (which would allow users to disable the behaviour if they want the less verbose version of the output).

Thanks for the feedback! I'll keep these points in mind :)

Regarding this patch - I think that because it has been reviewed to quite some extent, it doesn't need to be broken into smaller patches now?

Maybe I missed something from some of the other reviewers, but I haven't seen much in the way of commentary on the core functionality of the new code, so I wouldn't say that it has been reivewed to quite some extent. I've personally only skirted around style aspects. I also don't see any evidence of any testing of the new code.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1476

Bits or bytes?

[rochauha]

Added new test case.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1476

Bits.

If there is a bit that is wrong in a particular chunk of bytes, we consider that the entire chunk of bytes is invalid. We then update the Size value.

Further, we say that the first Size bytes in a symbol are invalid. Error handling in llvm-objdump will print these bytes using .byte directive. And then we fall back to decoding the remaining bytes in the symbol as instructions.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1476

We do this because most directives in the kernel descriptor affect single or a very few bits.

[scott.linder]

In D80713#2110898, @jhenderson wrote:

In D80713#2108650, @rochauha wrote:

In D80713#2106033, @jhenderson wrote:

No idea about the functional logic of this code, but I do wonder whether you'd be better off dividing it into smaller patches for testability.

This patch is for entirely disassembling the kernel descriptor symbol. It happens to be so that many directives affect the kernel descriptor. I will also be adding a new test in this patch.

The problem is that you're trying to do it all at once. There's no need to implement the full disassembly at once. You can do it in a series of patches, one after the other that build towards the primary goal. For example, there are many if (<X is some value>) type cases, which you could omit - just assume those are all false, in earlier versions, and add them in (with corresponding testing) in a later patch. Similarly, you could assume that all results of x & y return some specific value, e.g. 0, and just print that for now. Yes, that means you won't support everything from the point at which this patch lands, but it will make each individual patch easier to reason with. This fits much better with LLVM's preferred approach - please see https://llvm.org/docs/DeveloperPolicy.html#incremental-development.

I'm curious if this is the intent of the document you linked, though? It says "The remaining inter-related work should be decomposed into unrelated sets of changes if possible.", but the disassembly of this directive is not decomposable into unrelated changes; if only part of the descriptor is disassembled then the whole .amdhsa_kernel block is invalid. I suppose in the interest of review one could break down the logical/atomic series of commits even further into inter-dependent patches that must be applied together? I think that causes confusion when doing git-bisect etc. so doesn't seem ideal. Or maybe the patches should be squashed back together after review but before they are committed?

I do think there needs to be more direct testing of the branches taken in the code. If decomposing the patch beyond the logical/atomic decomposition makes it easier to do that as an author and/or a reviewer then I am OK with it. In this case, though, a lot of the length in terms of SLOC is just repetition and a little redundancy, so breaking it up further would only hide that and make it harder to see where things are best removed or factored out.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1227

This doesn't seem right. This symbol is not accurate without careful attention from an assembly author, so it will surely be incorrect when used by a disassembler.

We can accurately compute a VGPR count from the GRANULATED_WORKITEM_VGPR_COUNT, we just calculate e.g. (GRANULATED_WORKITEM_VGPR_COUNT + 1) * granularity where the + 1 is needed to account for the minimum allocation (i.e. a granulated count encoded as "0" actually indicates a 1 granule allocation) and where the scaling by granularity is device-specific. This will calculate the greatest value possible for .amdhsa_next_free_vgpr which will produce the same granulated count, but equally valid you could choose to calculate the minimum, or any value in between. All we care about is producing disassembly which results in the same descriptor. (Caveat: I would prove this to yourself by referencing https://llvm.org/docs/AMDGPUUsage.html#amdgpu-amdhsa-compute-pgm-rsrc1-gfx6-gfx10-table as I may have the actual calculation wrong; we just want to calculate the inverse of what the assembler does)

Your point that we can't recover the exact value the original author intended to place in the assembly text is true, but that is OK as long as we can always give them a valid input to the assembler which gives them the same output.

1232

This can be a char literal, i.e. '\n', same elsewhere.

1245

For this and the above case we should have tests to prove this out. I.e. assemble sources to a binary, disassemble and reassemble it, and then compare the two binaries. Ideally we would do this for some edge cases around VGPR/SGPR allocation granularity.

There may need to be some fixup between disassembly and reassembly to account for the remaining non-reassembleable bits produced by llvm-objdump, but they should be pretty minor for a trivial kernel, and I would expect you could handle them with just sed which seems to be available to LIT tests.

1254

Just as with the VGPR count, we cannot use the symbol to define this, and we can compute a (non-unique) input which produces the same output.

1264

This shift isn't necessary, you just need to check for the presence of any set bits.

I also noticed when checking the types here that for some reason we declare the enum for these masks/shifts as int32_t, which makes one have to think about both integer promotion rules and then possibly which bitwise-operations are valid for signed integers. I think here the signed mask is promoted to unsigned, and you get what you want, but it may be good to go fix the definition of the masks separately.

1269

I think a short-lived well-defined preprocessor macro could make the patch shorter and easier to read. For example:

#define PRINT_TRIVIAL_FIELD(DIRECTIVE, MASK)  do { \
  KdStream << Indent << DIRECTIVE " " << ((FourByteBuffer & amdhsa:: ## MASK) >> amdhsa:: ## MASK ## _SHIFT) << '\n';
  } while (0);
PRINT_TRIVIAL_FIELD(".amdhsa_float_round_mode_32", COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32)
PRINT_TRIVIAL_FIELD(".amdhsa_float_round_mode_16_64", COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64)
// ...
#undef HANDLE_TRIVIAL_FIELD

I think in part this is because the definition of the masks themselves is in terms of preprocessor macros.

1290

Unnecessary shift

1301

Unnecessary shift, same for all other cases below when checking that the bits under a mask are 0.

1345

I don't know the general conventions here, but I don't think I have seen a comment for the end of a function elsewhere in LLVM. I do know that it is required for namespaces, so maybe it is permitted for long functions?

1387

All of these comments seem redundant to me, especially when the condition is simplified to just:

if (Buffer & MASK) return Fail;

At the very least, repeating the actual bit indices here when they are a part of the mask definition seems verbose.

1472

I don't understand the intent with carefully maintaining Size for the failure case. Aren't we certain by this point that this should be a kernel descriptor, and so the correct thing to do when we fail is to disassemble everything as .byte directives (i.e. set Size to the max value)? Why would we prefer to return a partial failure and have the disassembler start working on the remaining bytes as if they were instructions?

That would also shorten the patch and make it obvious that the Size tracking is correct.

1484

Rather than have these comments, which are still just filled with magic numbers, could we define these offsets more explicitly somewhere, as e.g. amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET? For example in llvm/include/llvm/Support/AMDHSAKernelDescriptor.h alongside the other definitions needed by the compiler? We could then also update the static_asserts there to use those definitions, so we aren't relying on inspection to know the same offsets are used everywhere. I.e. the following:

165 static_assert(
  1     sizeof(kernel_descriptor_t) == 64,
  2     "invalid size for kernel_descriptor_t");
  3 static_assert(
  4     offsetof(kernel_descriptor_t, group_segment_fixed_size) == 0,
  5     "invalid offset for group_segment_fixed_size");
  6 static_assert(
  7     offsetof(kernel_descriptor_t, private_segment_fixed_size) == 4,
  8     "invalid offset for private_segment_fixed_size");
  9 static_assert(
 10     offsetof(kernel_descriptor_t, reserved0) == 8,
 11     "invalid offset for reserved0");
...

Becomes:

165 static_assert(
  1     sizeof(kernel_descriptor_t) == 64,
  2     "invalid size for kernel_descriptor_t");
  3 static_assert(
  4     offsetof(kernel_descriptor_t, group_segment_fixed_size) == GROUP_SEGMENT_FIXED_SIZE_OFFSET,
  5     "invalid offset for group_segment_fixed_size");
  6 static_assert(
  7     offsetof(kernel_descriptor_t, private_segment_fixed_size) == PRIVATE_SEGMENT_FIXED_SIZE_OFFSET,
  8     "invalid offset for private_segment_fixed_size");
  9 static_assert(
 10     offsetof(kernel_descriptor_t, reserved0) == RESERVED0_OFFSET,
 11     "invalid offset for reserved0");
...

1630

The != 0 here is redundant.

1635

Could you move the call to .drop_back(3) out into the calling function, so it appears next to the check for .endswith(StringRef(".kd"))?

1665

I'm still not sure what we landed on for the semantics of SoftFail here?

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h

83

Could you either rename this to satisfy the linter, or else explicitly suppress the lint with a comment like:

// NOLINTNEXTLINE(readability-identifier-naming)

llvm/test/tools/llvm-objdump/ELF/AMDGPU/code-object-v3.ll

48 ↗

(On Diff #274016)

I think we need more tests to ensure:

• The edge cases of the granularity calculation are correct
• That our promise of round-trip is fulfilled for at least some representative cases
• That some of the failure modes are handled how we expect

[rochauha]

Updating the code bit by bit based on comments by @scott.linder.

[rochauha]

• Used macro to shorten printing directives
• Replaced "\n" with '\n'
• Removed unnecessary bit shifts for bits that must be 0
• Removed extra comments for bits that must be 0

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1345

I'm not sure. I added those comments because these functions were getting quite long.

1630

I know, but I thought that it is more readable this way.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1665

It should be Success / Fail based on what the bytes are for code object v2. But there's nothing we are 'doing' at the moment for v2, I returned SoftFail.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1245

Right now we can't really re-assemble in the lit-test. This needs to be tested 'informally' by:

• Manually writing a small test case. Make a copy of it too.
• Assembling it into the binary : Binary-1.
• Disassembling it.
• Replace the original kernel descriptor with the disassembled kernel descriptor in the copy.
• Assemble the copy : Binary-2.
• Compare Binary-1 and Binary-2.

[rochauha]

• Fixed code-object-v3 lit test failure introduced in the previous change to this patch.
• More changes as per inline comments.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1472

My initial take was to decode as byte directives to the point of failure indicated by Size. Then going back to the normal flow of disassembling as instructions.

But I get what you want to say in this comment.

Updated the code based on this comment.

1635

Done.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h

83

Done.

[rochauha]

Size = 64 regardless of success or failure.

[rochauha]

• Compute .amdhsa_next_free_vgpr based on inverse of what the assembler does to compute GRANULATED_WORKITEM_VGPR_COUNT.
• Some changes to accomodate differences between GFX9 and GFX10
• Updated test case for GFX10 as well

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1245

Went this route to check whether re-assembled binaries match or not. Turns out that both binaries match, in size (overall size as well as size of sections) and also in terms of all the disassembled content. But a diff object1 object2 says that binary files differ.

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1245

I'm not sure I follow what you are describing; my thought was to start with just an asm source file containing only the kernel descriptor directive in the default section, and compare the output of the following (with, e.g. diff, as you mention):

• Assemble it to an object file with llvm-mc
• Assemble it to an object file with llvm-mc | disassemble the kernel descriptor symbol | trim any human-readable prologue | assemble it to an object file with llvm-mc

As a trivial example, diff doesn't find any difference for the following example:

$ printf '.amdhsa_kernel my_kernel\n.amdhsa_next_free_vgpr 0\n.amdhsa_next_free_sgpr 0\n.end_amdhsa_kernel' >a.s
$ release/bin/llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj a.s >a.o
$ diff a.o \
       <(release/bin/llvm-objdump --triple=amdgcn-amd-amdhsa --mcpu=gfx908 --disassemble-symbols=my_kernel.kd a.o \
                 | tail -n +8 \
                 | release/bin/llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj)

I don't think you need to use FileCheck for these tests at all, you can just rely on ending the RUN pipeline with diff, which seems to be supported by lit. You can then just copy-paste the test and edit fields in the input to validate edge cases for things like the SGPR/VGPR allocation directives.

I think more comprehensive testing, including for other sections and executables/DSOs, would be good eventually but for now we should at least have some tests that explicitly confirm the KD disassembly round-trips.

1345

I would lean towards omitting these, especially with the functions becoming shorter. For example, decodeCOMPUTE_PGM_RSRC2() is now <50 lines long at the entire definition now fits on one screen for me. It seems like there are other examples of this in the codebase, though, so I'm OK with it for the longer functions.

1349

Can you expand this comment a little and move it to a Doxygen comment for the function?

1624

Need to handle the "default" case here:

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp:1510:1: warning: control may reach end of non-void function [-Wreturn-type]

1630

Fair enough, in a type-safe language it would be required anyway, so it seems reasonable.

1665

If SoftFail isn't applicable I don't think we should return it, even if it is just because we haven't implemented something yet. It existing doesn't mean it needs to be used, I think it has a very narrow definition that doesn't apply here.

Maybe just emit a diagnostic and return Fail so we get the "decode as .byte" behavior? What exactly happens now with the current patch as-is?

[rochauha]

• Handled default statement to silence the warning.
• Expanded comments for decodeCOMPUTE_PGM_RSRC1 and decodeCOMPUTE_PGM_RSRC2.
• Removed extra comment at the end of functions.
• Changed SoftFail to Success for code object v2.
• Replaced the old test case with a small assembly file.

[rochauha]

• Return MCDisassembler::Fail for code object v2.
• Add missing full stops in doxygen comments.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1345

Done.

1349

Done.

1624

Done.

1665

Done.

[jhenderson]

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor.s

1 ↗

(On Diff #276063)

No need for the <. llvm-mc is quite capable of taking inputs on the command-line as positional arguments.

2 ↗

(On Diff #276063)

This line is too long. Please break it up into individual lines:
; RUN: llvm-objdump ... | \
; RUN: tail -n +8 | llvm-mc ...

5–8 ↗

(On Diff #276063)

This test is also quite small. Does it actually cover every code path?

10 ↗

(On Diff #276063)

Is this a FIXME/TODO? If so, please add "FIXME" or "TODO".

[rochauha]

Changes as per review by @jhenderson.

[rochauha]

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor.s

2 ↗

(On Diff #276063)

Done.

5–8 ↗

(On Diff #276063)

These values must be always specified by the user. Values of some bytes are computed using the values passed here. Consequently there are some cases where we need to test for getting the exact same bytes in the re-assembled binary, even if the disassembled values slightly deviate from the original values.

Other bytes/bits hold the exact values specified using the assembler directives. Also they take default values if nothing is specified by the user. So I guess we don't need to test that for re-assembly.

10 ↗

(On Diff #276063)

Done.

[rochauha]

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor.s

1 ↗

(On Diff #276063)

Done.

[rochauha]

Switched to disassembling as numerical value rather than .amdgcn.next_free_sgpr.

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1243–1250

I think this can just be replaced with:

NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) * getVGPRAllocGranule(STI);

Or we could add another function called getNumVGPRs(const MCSubtargetInfo *STI, unsigned NumVGPRBlocks, Optional<bool> EnableWavefrontSize32) and put the definition directly next to getNumVGPRBlocks(const MCSubtargetInfo *STI, unsigned NumVGPRs, Optional<bool> EnableWavefrontSize32) so any future changes affecting one also affect the other. I would lean towards this, and documenting that they are the inverse of one another.

1278–1285

Same as above, I think a new getNumSGPRs to complement getNumSGPRBlocks would make this easier to read.

For the GFX10 case we could either leave the check here, or have the new function return Optional to indicate when there is an error.

1290

What is "GS" and why is this commented out?

1484

I would still like to see this done, the magic numbers here could lead to a few problems down the line, and presently they just make the code harder to read.

1572–1576

Can this just be return decodeKernelDescriptor(...);?

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor.s

6 ↗

(On Diff #279490)

Can you implement more tests? I don't know if it is feasible to include them all in the same .s file, as you have to work around the output not being re-assembleable in general, but even just copy-pasting and editing is fine with me. Just having some more test cases to cover at least a reasonable sample of the different branches, failure modes, etc.

[rochauha]

• Got rid of raw number and added an enum for offsets.
• Added new tests.
• Updated the nop-data.ll test so that it passes with this patch.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1243–1250

I think this route is good for now :
NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) * getVGPRAllocGranule(STI);

Similarly for SGPRs.

I'd like to add the new functions via a separate patch. This patch is already quite big in terms of size.

1278–1285

As I mentioned in my other reply, this patch is already quite big. So it'd be better to have a separate patch for the new functions.

1290

This was was for printing Granulated wave front SGPR count.

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor.s

6 ↗

(On Diff #279490)

I have added new tests in separate files.

[scott.linder]

LGTM, thank you!

[jhenderson]

I'm probably not in a position to review the majority of this further. However, I do have big reservations about the testing - there are a high number of possible code paths, but I only see 5 tests, which clearly can't cover all these code paths.

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kernel-descriptor-1.s

1 ↗

(On Diff #281180)

It would be helpful to add a small comment at the start of each of these tests explaining what the are specifically testing.

[scott.linder]

In D80713#2186791, @jhenderson wrote:

I'm probably not in a position to review the majority of this further. However, I do have big reservations about the testing - there are a high number of possible code paths, but I only see 5 tests, which clearly can't cover all these code paths.

I think some of the difficulty in writing tests for the failure cases is not having a tool to produce them. I think one would have to generate the code object, and then hand-edit it in a hex-editor, copy it into the source tree (i.e. under Inputs/) maybe with some accompanying comment about the nature of the original source and the edits made?

It would be nice to at least cover some of the failure cases, even if they are more awkward to tests. There could also be some similar tests for e.g. GFX7,GFX10 and the edge cases around the {S,V}GPR calculation for each.

[jhenderson]

In D80713#2187653, @scott.linder wrote:

In D80713#2186791, @jhenderson wrote:

I'm probably not in a position to review the majority of this further. However, I do have big reservations about the testing - there are a high number of possible code paths, but I only see 5 tests, which clearly can't cover all these code paths.

I think some of the difficulty in writing tests for the failure cases is not having a tool to produce them. I think one would have to generate the code object, and then hand-edit it in a hex-editor, copy it into the source tree (i.e. under Inputs/) maybe with some accompanying comment about the nature of the original source and the edits made?

It would be nice to at least cover some of the failure cases, even if they are more awkward to tests. There could also be some similar tests for e.g. GFX7,GFX10 and the edge cases around the {S,V}GPR calculation for each.

It seems to me that the format of the data structure is well-understood (otherwise you wouldn't be able to write code to disassemble it). In similar situations, e.g. the DWARF .debug_line parsing, we didn't rely on the built-in .file/.loc directives to generate our line table, and instead wrote it out by hand using .byte/.half/.long/.quad. It's not the prettiest of things of course, but it's better than code paths that never get exercised in tests. You could also use yaml2obj to do a generate the section with a raw hex blob, and/or write an array of bytes in a gtest unit test. The latter situation might be particularly useful because it would allow you to use the same "base" array, and modify individual bytes in individual test cases to check the behaviour for each.

[rochauha]

• Updated tests and added a failure test.
• Checking error when cursor holds Error::success() while handling failure cases.
• Removed Size from subroutines' signatures as we aren't tracking the size.

[rochauha]

Can I consider this patch to be at NFC stage?

[jhenderson]

Hi @rochauha,

I'm not sure this should have been pushed - nobody has reviewed your latest update, and I've had concerns about the testing prior to that, as stated. There is also a "Needs Revision" marker still outstanding, and I usually take that to mean that this shouldn't be pushed until the relevant reviewer is satisfied, regardless of what others have said (note that Phabricator highlighted that this patch landded in a "Needs Review" state).

Please could you revert, as there are potential assertion failures in my reading of this code, at least in the event you hit malformed input, in addition to the other issues raised.

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1220–1225

This is not how to handle malformed input. This will result in an assertion in debug builds, which is equivalent to a crash, without any useful context to draw on, because you've not reported the error. More below.

1408

Rather than all these checkError calls, I'd expect to see a check of the Cursor followed by a return of MCDisassembler::Fail to indicate there was a problem.

1542–1548

If this loop terminates due to C being invalid, you don't want to fall out the bottom and return Success, I think. You'd want to check C after loop termination and return Fail. Alternatively, if you return Fail from the decodeKernelDescriptorDirective you can do cantFail(C.takeError()); after the loop.

Ideally, we'd actually report the error from C in the event of failure, but currently there's no way of communicating that back up to the caller.

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-failure.s

2

Please add a comment to the top of this test explaining what this test is actually testing.

4

Rather than writing text to another input file at run time in this way, you can use the new split-file tool to have all the input inline below, and split it up using the tool into multiple files.

12

Please don't mix comment markers within the same file. You use '//' here, but ';' everywhere else.

Additionally, new LLVM binutils tests tend to use double comment markers to indicate true comments as opposed to RUN/CHECK lines (i.e. ';;' in this context).

35–36

Please remove the additional blank lines at the end of file.

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-sgpr.s

2

Please follow the comments from kd-failure.s in all the tests (where applicable) too.

[rochauha]

Changes based on comments by @jhenderson

• Used llvm::cantFail(C.takeError()) to handle error.
• Removed blank lines at the end of test files.
• Streamlined beginning of comments.
• Used split-file tool to separate SGPR and VGPR test cases.

[rochauha]

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-failure.s

4

In case of kd-failure.s, concatenation is necessary because the disassembled output will be .bytes, and the symbol information is needed to be get the same binary again. I feel that "printf'ing" to a file and concatenating with disassembled text is a simpler compared to split file and concatenating.

However, I have used the split-file tool to separate the S/VGPR test cases.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1542–1548

Thanks for the pointer! I think cantFail seems to be all that is needed, because the cursor was holding Error::success in some cases, which needs to be 'checked' before moving further.

Since the kernel descriptor is well defined, all cases where failure needs to be handled are handled using MCDisassembler::Fail.

[jhenderson]

@kzhuravl, do you have any additional comments to add at all, since you rejected the original patch?

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1394

I think for self-documentation purposes, it would be helpful to assert that Bytes.size() == 64 here. I see that it is verified in the calling function but a) that's not obvious when looking at this function in isolation, and b) in the future, we don't want other places calling this code without that check, so the assert provides a backstop of sorts.

1542–1548

Ah, I missed that the input can't be truncated, so C can't get into a failure state itself. Thanks! (See also my comment about an assert above).

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-failure.s

2

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-sgpr.s

22

Nit: missing full stop.

[rochauha]

Changes based on comments by @jhenderson.

[jhenderson]

No more comments from me, but please give @kzhuravl a chance to respond before pushing this patch.

[kzhuravl]

LGTM, thank you!

[rochauha]

3 of the test cases - kd-sgpr.s, kd-vgpr.s, kd-zeroed-gfx10.s fail with the PowerPC buildbot (http://lab.llvm.org:8011/builders/clang-ppc64be-linux/builds/53608). From what I understand, these tests should only run if AMDGPU target is built. The lit.local.cfg file specifies that.

[jhenderson]

In D80713#2261550, @rochauha wrote:

3 of the test cases - kd-sgpr.s, kd-vgpr.s, kd-zeroed-gfx10.s fail with the PowerPC buildbot (http://lab.llvm.org:8011/builders/clang-ppc64be-linux/builds/53608). From what I understand, these tests should only run if AMDGPU target is built. The lit.local.cfg file specifies that.

The PowerPC build bot just means a build bot that is running on a PowerPC host. It actually targets most targets. See the following excerpt from the CMake.

-- Targeting AArch64
-- Targeting AMDGPU
-- Targeting ARM
-- Targeting AVR
-- Targeting BPF
-- Targeting Hexagon
-- Targeting Lanai
-- Targeting Mips
-- Targeting MSP430
-- Targeting NVPTX
-- Targeting PowerPC
-- Targeting RISCV
-- Targeting Sparc
-- Targeting SystemZ
-- Targeting WebAssembly
-- Targeting X86
-- Targeting XCore

I've run into this sort of problem before. The issue is almost certainly either a) incorrect assumption about host system endianness, meaning that you've incorrectly/inadvertently assumed the host is little endian, or b) assumed a 64-bit system somewhere. a) is almost certainly the issue here, based on both the test output and build bot name (ppc64*be*). I've skimmed the patch, but can't obviously see where the code is going wrong, but the test output for kd-zeroed-gfx10.s suggests it's around bytes 48-52. There may be other issues elsewhere though, since 0 renders the same regardless of size and endianness.

[nemanjai]

In D80713#2262751, @jhenderson wrote:

...

I've run into this sort of problem before. The issue is almost certainly either a) incorrect assumption about host system endianness, meaning that you've incorrectly/inadvertently assumed the host is little endian, or b) assumed a 64-bit system somewhere. a) is almost certainly the issue here, based on both the test output and build bot name (ppc64*be*). I've skimmed the patch, but can't obviously see where the code is going wrong, but the test output for kd-zeroed-gfx10.s suggests it's around bytes 48-52. There may be other issues elsewhere though, since 0 renders the same regardless of size and endianness.

Yes, considering the failures appear to be exactly the same on PowerPC Big Endian and SystemZ (which is also Big Endian), I would assume that there is a host endianness assumption here (I haven't looked at the code). Please pull this patch until a fix is ready. There are at least the 4 bots that are red because of this for quite a few consecutive builds. Pull the patch so we can get the bots back to green and we can work on a fix afterwards.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

I think this is where little endian is being 'hardcoded'. However, since AMDGPU relocatable objects are meant to be little endian, I don't understand why they are big endian on pp64.

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

I think this is actually a bug with the encoding code. Like you say, we should be host-endianness agnostic when encoding the kernel descriptor, but it seems we aren't.

I think I vaguely remember this coming up when we implemented it, but I don't remember why we didn't do this to start. I think it is just an oversight.

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

I think https://reviews.llvm.org/D88858 should be the fix, need to confirm if big-endian testers will run it automatically.

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

I tried to determine if the pre-checkin builders include any big-endian archs, but gave up and just committed it. I'll keep an eye on the builders and see if it needs to be revered. After that you can proceed with this patch again.

[rochauha]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

Thanks!

[scott.linder]

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

1395

An update, I missed one test in the initial commit, but I followed up in bf5c1d92d92ef8cee2adbfa17ecca20a8f65dc0e and now the big-endian testers seem to be happy. I think you can try reapplying this.

[rochauha]

Previous reversions were due to test cases failing on big endian hosts.
This was due to multibyte values being laid out as-is from host memory.
Now https://reviews.llvm.org/D88858 addresses the above, and has landed.

[rochauha]

Re-applied patch.

llvm/include/llvm/Support/AMDHSAKernelDescriptor.h

struct kernel_descriptor_t {
   uint8_t reserved1[20];
   uint32_t compute_pgm_rsrc3; // GFX10+
   uint32_t compute_pgm_rsrc1;
   uint32_t compute_pgm_rsrc2;
   uint16_t kernel_code_properties;
   uint8_t reserved2[6];
 };
 
+enum : uint32_t {
+  GROUP_SEGMENT_FIXED_SIZE_OFFSET = 0,
+  PRIVATE_SEGMENT_FIXED_SIZE_OFFSET = 4,
+  RESERVED0_OFFSET = 8,
+  KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET = 16,
+  RESERVED1_OFFSET = 24,
+  COMPUTE_PGM_RSRC3_OFFSET = 44,
+  COMPUTE_PGM_RSRC1_OFFSET = 48,
+  COMPUTE_PGM_RSRC2_OFFSET = 52,
+  KERNEL_CODE_PROPERTIES_OFFSET = 56,
+  RESERVED2_OFFSET = 58,
+};
+
 static_assert(
     sizeof(kernel_descriptor_t) == 64,
     "invalid size for kernel_descriptor_t");
-static_assert(
-    offsetof(kernel_descriptor_t, group_segment_fixed_size) == 0,
+static_assert(offsetof(kernel_descriptor_t, group_segment_fixed_size) ==
+                  GROUP_SEGMENT_FIXED_SIZE_OFFSET,
               "invalid offset for group_segment_fixed_size");
-static_assert(
-    offsetof(kernel_descriptor_t, private_segment_fixed_size) == 4,
+static_assert(offsetof(kernel_descriptor_t, private_segment_fixed_size) ==
+                  PRIVATE_SEGMENT_FIXED_SIZE_OFFSET,
               "invalid offset for private_segment_fixed_size");
-static_assert(
-    offsetof(kernel_descriptor_t, reserved0) == 8,
+static_assert(offsetof(kernel_descriptor_t, reserved0) == RESERVED0_OFFSET,
               "invalid offset for reserved0");
-static_assert(
-    offsetof(kernel_descriptor_t, kernel_code_entry_byte_offset) == 16,
+static_assert(offsetof(kernel_descriptor_t, kernel_code_entry_byte_offset) ==
+                  KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET,
               "invalid offset for kernel_code_entry_byte_offset");
-static_assert(
-    offsetof(kernel_descriptor_t, reserved1) == 24,
+static_assert(offsetof(kernel_descriptor_t, reserved1) == RESERVED1_OFFSET,
               "invalid offset for reserved1");
-static_assert(
-    offsetof(kernel_descriptor_t, compute_pgm_rsrc3) == 44,
+static_assert(offsetof(kernel_descriptor_t, compute_pgm_rsrc3) ==
+                  COMPUTE_PGM_RSRC3_OFFSET,
               "invalid offset for compute_pgm_rsrc3");
-static_assert(
-    offsetof(kernel_descriptor_t, compute_pgm_rsrc1) == 48,
+static_assert(offsetof(kernel_descriptor_t, compute_pgm_rsrc1) ==
+                  COMPUTE_PGM_RSRC1_OFFSET,
               "invalid offset for compute_pgm_rsrc1");
-static_assert(
-    offsetof(kernel_descriptor_t, compute_pgm_rsrc2) == 52,
+static_assert(offsetof(kernel_descriptor_t, compute_pgm_rsrc2) ==
+                  COMPUTE_PGM_RSRC2_OFFSET,
               "invalid offset for compute_pgm_rsrc2");
-static_assert(
-    offsetof(kernel_descriptor_t, kernel_code_properties) == 56,
+static_assert(offsetof(kernel_descriptor_t, kernel_code_properties) ==
+                  KERNEL_CODE_PROPERTIES_OFFSET,
               "invalid offset for kernel_code_properties");
-static_assert(
-    offsetof(kernel_descriptor_t, reserved2) == 58,
+static_assert(offsetof(kernel_descriptor_t, reserved2) == RESERVED2_OFFSET,
               "invalid offset for reserved2");
 
 } // end namespace amdhsa
 } // end namespace llvm
 
 #endif // LLVM_SUPPORT_AMDHSAKERNELDESCRIPTOR_H

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h

 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_AMDGPU_DISASSEMBLER_AMDGPUDISASSEMBLER_H
 #define LLVM_LIB_TARGET_AMDGPU_DISASSEMBLER_AMDGPUDISASSEMBLER_H
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/MC/MCContext.h"
-#include "llvm/MC/MCInstrInfo.h"
 #include "llvm/MC/MCDisassembler/MCDisassembler.h"
 #include "llvm/MC/MCDisassembler/MCRelocationInfo.h"
 #include "llvm/MC/MCDisassembler/MCSymbolizer.h"
+#include "llvm/MC/MCInstrInfo.h"
+#include "llvm/Support/DataExtractor.h"
 
 #include <algorithm>
 #include <cstdint>
 #include <memory>
 
 namespace llvm {
 
 class MCInst;
public:
   MCOperand createRegOperand(unsigned RegClassID, unsigned Val) const;
   MCOperand createSRegOperand(unsigned SRegClassID, unsigned Val) const;
 
   MCOperand errOperand(unsigned V, const Twine& ErrMsg) const;
 
   DecodeStatus tryDecodeInst(const uint8_t* Table, MCInst &MI, uint64_t Inst,
                              uint64_t Address) const;
 
+  Optional<DecodeStatus> onSymbolStart(SymbolInfoTy &Symbol, uint64_t &Size,
+                                       ArrayRef<uint8_t> Bytes,
+                                       uint64_t Address,
+                                       raw_ostream &CStream) const override;
+
+  DecodeStatus decodeKernelDescriptor(StringRef KdName, ArrayRef<uint8_t> Bytes,
+                                      uint64_t KdAddress) const;
+
+  DecodeStatus
+  decodeKernelDescriptorDirective(DataExtractor::Cursor &Cursor,
+                                  ArrayRef<uint8_t> Bytes,
+                                  raw_string_ostream &KdStream) const;
+
+  /// Decode as directives that handle COMPUTE_PGM_RSRC1.
+  /// \param FourByteBuffer - Bytes holding contents of COMPUTE_PGM_RSRC1.
+  /// \param KdStream       - Stream to write the disassembled directives to.
+  // NOLINTNEXTLINE(readability-identifier-naming)
+  DecodeStatus decodeCOMPUTE_PGM_RSRC1(uint32_t FourByteBuffer,
+                                       raw_string_ostream &KdStream) const;
+
+  /// Decode as directives that handle COMPUTE_PGM_RSRC2.
+  /// \param FourByteBuffer - Bytes holding contents of COMPUTE_PGM_RSRC2.
+  /// \param KdStream       - Stream to write the disassembled directives to.
+  // NOLINTNEXTLINE(readability-identifier-naming)
+  DecodeStatus decodeCOMPUTE_PGM_RSRC2(uint32_t FourByteBuffer,
+                                       raw_string_ostream &KdStream) const;
+
   DecodeStatus convertSDWAInst(MCInst &MI) const;
   DecodeStatus convertDPP8Inst(MCInst &MI) const;
   DecodeStatus convertMIMGInst(MCInst &MI) const;
 
   MCOperand decodeOperand_VGPR_32(unsigned Val) const;
   MCOperand decodeOperand_VRegOrLds_32(unsigned Val) const;
 
   MCOperand decodeOperand_VS_32(unsigned Val) const;

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCDisassembler/MCDisassembler.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCFixedLenDisassembler.h"
 #include "llvm/MC/MCInst.h"
 #include "llvm/MC/MCSubtargetInfo.h"
+#include "llvm/Support/AMDHSAKernelDescriptor.h"
 #include "llvm/Support/Endian.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/TargetRegistry.h"
 #include "llvm/Support/raw_ostream.h"
 #include <algorithm>
 #include <cassert>
 #include <cstddef>
bool AMDGPUDisassembler::isGFX9() const {
   return STI.getFeatureBits()[AMDGPU::FeatureGFX9];
 }
 
 bool AMDGPUDisassembler::isGFX10() const {
   return STI.getFeatureBits()[AMDGPU::FeatureGFX10];
 }
 
 //===----------------------------------------------------------------------===//
+// AMDGPU specific symbol handling
+//===----------------------------------------------------------------------===//
+#define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
+  do {                                                                         \
+    KdStream << Indent << DIRECTIVE " "                                        \
+             << ((FourByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';           \
+  } while (0)
+
+// NOLINTNEXTLINE(readability-identifier-naming)
+MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC1(
+    uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
+  using namespace amdhsa;
+  StringRef Indent = "\t";
+
+  // We cannot accurately backward compute #VGPRs used from
+  // GRANULATED_WORKITEM_VGPR_COUNT. But we are concerned with getting the same
+  // value of GRANULATED_WORKITEM_VGPR_COUNT in the reassembled binary. So we
+  // simply calculate the inverse of what the assembler does.
+
+  uint32_t GranulatedWorkitemVGPRCount =
+      (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT) >>
+      COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT_SHIFT;
+
+  uint32_t NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) *
+                          AMDGPU::IsaInfo::getVGPREncodingGranule(&STI);
+
+  KdStream << Indent << ".amdhsa_next_free_vgpr " << NextFreeVGPR << '\n';
+
+  // We cannot backward compute values used to calculate
+  // GRANULATED_WAVEFRONT_SGPR_COUNT. Hence the original values for following
+  // directives can't be computed:
+  // .amdhsa_reserve_vcc
+  // .amdhsa_reserve_flat_scratch
+  // .amdhsa_reserve_xnack_mask
+  // They take their respective default values if not specified in the assembly.
+  //
+  // GRANULATED_WAVEFRONT_SGPR_COUNT
+  //    = f(NEXT_FREE_SGPR + VCC + FLAT_SCRATCH + XNACK_MASK)
+  //
+  // We compute the inverse as though all directives apart from NEXT_FREE_SGPR
+  // are set to 0. So while disassembling we consider that:
+  //
+  // GRANULATED_WAVEFRONT_SGPR_COUNT
+  //    = f(NEXT_FREE_SGPR + 0 + 0 + 0)
+  //
+  // The disassembler cannot recover the original values of those 3 directives.
+
+  uint32_t GranulatedWavefrontSGPRCount =
+      (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT) >>
+      COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT_SHIFT;
+
+  if (isGFX10() && GranulatedWavefrontSGPRCount)
+    return MCDisassembler::Fail;
+
+  uint32_t NextFreeSGPR = (GranulatedWavefrontSGPRCount + 1) *
+                          AMDGPU::IsaInfo::getSGPREncodingGranule(&STI);
+
+  KdStream << Indent << ".amdhsa_reserve_vcc " << 0 << '\n';
+  KdStream << Indent << ".amdhsa_reserve_flat_scratch " << 0 << '\n';
+  KdStream << Indent << ".amdhsa_reserve_xnack_mask " << 0 << '\n';
+  KdStream << Indent << ".amdhsa_next_free_sgpr " << NextFreeSGPR << "\n";
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIORITY)
+    return MCDisassembler::Fail;
+
+  PRINT_DIRECTIVE(".amdhsa_float_round_mode_32",
+                  COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
+  PRINT_DIRECTIVE(".amdhsa_float_round_mode_16_64",
+                  COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
+  PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_32",
+                  COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
+  PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_16_64",
+                  COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIV)
+    return MCDisassembler::Fail;
+
+  PRINT_DIRECTIVE(".amdhsa_dx10_clamp", COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_DEBUG_MODE)
+    return MCDisassembler::Fail;
+
+  PRINT_DIRECTIVE(".amdhsa_ieee_mode", COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_BULKY)
+    return MCDisassembler::Fail;
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_CDBG_USER)
+    return MCDisassembler::Fail;
+
+  PRINT_DIRECTIVE(".amdhsa_fp16_overflow", COMPUTE_PGM_RSRC1_FP16_OVFL);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC1_RESERVED0)
+    return MCDisassembler::Fail;
+
+  if (isGFX10()) {
+    PRINT_DIRECTIVE(".amdhsa_workgroup_processor_mode",
+                    COMPUTE_PGM_RSRC1_WGP_MODE);
+    PRINT_DIRECTIVE(".amdhsa_memory_ordered", COMPUTE_PGM_RSRC1_MEM_ORDERED);
+    PRINT_DIRECTIVE(".amdhsa_forward_progress", COMPUTE_PGM_RSRC1_FWD_PROGRESS);
+  }
+  return MCDisassembler::Success;
+}
+
+// NOLINTNEXTLINE(readability-identifier-naming)
+MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC2(
+    uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
+  using namespace amdhsa;
+  StringRef Indent = "\t";
+  PRINT_DIRECTIVE(
+      ".amdhsa_system_sgpr_private_segment_wavefront_offset",
+      COMPUTE_PGM_RSRC2_ENABLE_SGPR_PRIVATE_SEGMENT_WAVEFRONT_OFFSET);
+  PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_x",
+                  COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
+  PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_y",
+                  COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
+  PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_z",
+                  COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
+  PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_info",
+                  COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
+  PRINT_DIRECTIVE(".amdhsa_system_vgpr_workitem_id",
+                  COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_ADDRESS_WATCH)
+    return MCDisassembler::Fail;
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_MEMORY)
+    return MCDisassembler::Fail;
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC2_GRANULATED_LDS_SIZE)
+    return MCDisassembler::Fail;
+
+  PRINT_DIRECTIVE(
+      ".amdhsa_exception_fp_ieee_invalid_op",
+      COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
+  PRINT_DIRECTIVE(".amdhsa_exception_fp_denorm_src",
+                  COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
+  PRINT_DIRECTIVE(
+      ".amdhsa_exception_fp_ieee_div_zero",
+      COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
+  PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_overflow",
+                  COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
+  PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_underflow",
+                  COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
+  PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_inexact",
+                  COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
+  PRINT_DIRECTIVE(".amdhsa_exception_int_div_zero",
+                  COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
+
+  if (FourByteBuffer & COMPUTE_PGM_RSRC2_RESERVED0)
+    return MCDisassembler::Fail;
+
+  return MCDisassembler::Success;
+}
+
+#undef PRINT_DIRECTIVE
+
+MCDisassembler::DecodeStatus
+AMDGPUDisassembler::decodeKernelDescriptorDirective(
+    DataExtractor::Cursor &Cursor, ArrayRef<uint8_t> Bytes,
+    raw_string_ostream &KdStream) const {
+#define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
+  do {                                                                         \
+    KdStream << Indent << DIRECTIVE " "                                        \
+             << ((TwoByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';            \
+  } while (0)
+
+  uint16_t TwoByteBuffer = 0;
+  uint32_t FourByteBuffer = 0;
+  uint64_t EightByteBuffer = 0;
+
+  StringRef ReservedBytes;
+  StringRef Indent = "\t";
+
+  DataExtractor DE(Bytes, /*IsLittleEndian=*/true, /*AddressSize=*/8);
+
+  switch (Cursor.tell()) {
+  case amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET:
+    FourByteBuffer = DE.getU32(Cursor);
+    KdStream << Indent << ".amdhsa_group_segment_fixed_size " << FourByteBuffer
+             << '\n';
+    return MCDisassembler::Success;
+
+  case amdhsa::PRIVATE_SEGMENT_FIXED_SIZE_OFFSET:
+    FourByteBuffer = DE.getU32(Cursor);
+    KdStream << Indent << ".amdhsa_private_segment_fixed_size "
+             << FourByteBuffer << '\n';
+    return MCDisassembler::Success;
+
+  case amdhsa::RESERVED0_OFFSET:
+    // 8 reserved bytes, must be 0.
+    EightByteBuffer = DE.getU64(Cursor);
+    if (EightByteBuffer) {
+      return MCDisassembler::Fail;
+    }
+    return MCDisassembler::Success;
+
+  case amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET:
+    // KERNEL_CODE_ENTRY_BYTE_OFFSET
+    // So far no directive controls this for Code Object V3, so simply skip for
+    // disassembly.
+    DE.skip(Cursor, 8);
+    return MCDisassembler::Success;
+
+  case amdhsa::RESERVED1_OFFSET:
+    // 20 reserved bytes, must be 0.
+    ReservedBytes = DE.getBytes(Cursor, 20);
+    for (int I = 0; I < 20; ++I) {
+      if (ReservedBytes[I] != 0) {
+        return MCDisassembler::Fail;
+      }
+    }
+    return MCDisassembler::Success;
+
+  case amdhsa::COMPUTE_PGM_RSRC3_OFFSET:
+    // COMPUTE_PGM_RSRC3
+    //  - Only set for GFX10, GFX6-9 have this to be 0.
+    //  - Currently no directives directly control this.
+    FourByteBuffer = DE.getU32(Cursor);
+    if (!isGFX10() && FourByteBuffer) {
+      return MCDisassembler::Fail;
+    }
+    return MCDisassembler::Success;
+
+  case amdhsa::COMPUTE_PGM_RSRC1_OFFSET:
+    FourByteBuffer = DE.getU32(Cursor);
+    if (decodeCOMPUTE_PGM_RSRC1(FourByteBuffer, KdStream) ==
+        MCDisassembler::Fail) {
+      return MCDisassembler::Fail;
+    }
+    return MCDisassembler::Success;
+
+  case amdhsa::COMPUTE_PGM_RSRC2_OFFSET:
+    FourByteBuffer = DE.getU32(Cursor);
+    if (decodeCOMPUTE_PGM_RSRC2(FourByteBuffer, KdStream) ==
+        MCDisassembler::Fail) {
+      return MCDisassembler::Fail;
+    }
+    return MCDisassembler::Success;
+
+  case amdhsa::KERNEL_CODE_PROPERTIES_OFFSET:
+    using namespace amdhsa;
+    TwoByteBuffer = DE.getU16(Cursor);
+
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_buffer",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_ptr",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_queue_ptr",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_segment_ptr",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_id",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_flat_scratch_init",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
+    PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size",
+                    KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
+
+    if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0)
+      return MCDisassembler::Fail;
+
+    // Reserved for GFX9
+    if (isGFX9() &&
+        (TwoByteBuffer & KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32)) {
+      return MCDisassembler::Fail;
+    } else if (isGFX10()) {
+      PRINT_DIRECTIVE(".amdhsa_wavefront_size32",
+                      KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
+    }
+
+    if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED1)
+      return MCDisassembler::Fail;
+
+    return MCDisassembler::Success;
+
+  case amdhsa::RESERVED2_OFFSET:
+    // 6 bytes from here are reserved, must be 0.
+    ReservedBytes = DE.getBytes(Cursor, 6);
+    for (int I = 0; I < 6; ++I) {
+      if (ReservedBytes[I] != 0)
+        return MCDisassembler::Fail;
+    }
+    return MCDisassembler::Success;
+
+  default:
+    llvm_unreachable("Unhandled index. Case statements cover everything.");
+    return MCDisassembler::Fail;
+  }
+#undef PRINT_DIRECTIVE
+}
+
+MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeKernelDescriptor(
+    StringRef KdName, ArrayRef<uint8_t> Bytes, uint64_t KdAddress) const {
+  // CP microcode requires the kernel descriptor to be 64 aligned.
+  if (Bytes.size() != 64 || KdAddress % 64 != 0)
+    return MCDisassembler::Fail;
+
+  std::string Kd;
+  raw_string_ostream KdStream(Kd);
+  KdStream << ".amdhsa_kernel " << KdName << '\n';
+
+  DataExtractor::Cursor C(0);
+  while (C && C.tell() < Bytes.size()) {
+    MCDisassembler::DecodeStatus Status =
+        decodeKernelDescriptorDirective(C, Bytes, KdStream);
+
+    cantFail(C.takeError());
+
+    if (Status == MCDisassembler::Fail)
+      return MCDisassembler::Fail;
+  }
+  KdStream << ".end_amdhsa_kernel\n";
+  outs() << KdStream.str();
+  return MCDisassembler::Success;
+}
+
+Optional<MCDisassembler::DecodeStatus>
+AMDGPUDisassembler::onSymbolStart(SymbolInfoTy &Symbol, uint64_t &Size,
+                                  ArrayRef<uint8_t> Bytes, uint64_t Address,
+                                  raw_ostream &CStream) const {
+  // Right now only kernel descriptor needs to be handled.
+  // We ignore all other symbols for target specific handling.
+  // TODO:
+  // Fix the spurious symbol issue for AMDGPU kernels. Exists for both Code
+  // Object V2 and V3 when symbols are marked protected.
+
+  // amd_kernel_code_t for Code Object V2.
+  if (Symbol.Type == ELF::STT_AMDGPU_HSA_KERNEL) {
+    Size = 256;
+    return MCDisassembler::Fail;
+  }
+
+  // Code Object V3 kernel descriptors.
+  StringRef Name = Symbol.Name;
+  if (Symbol.Type == ELF::STT_OBJECT && Name.endswith(StringRef(".kd"))) {
+    Size = 64; // Size = 64 regardless of success or failure.
+    return decodeKernelDescriptor(Name.drop_back(3), Bytes, Address);
+  }
+  return None;
+}
+
+//===----------------------------------------------------------------------===//
 // AMDGPUSymbolizer
 //===----------------------------------------------------------------------===//
 
 // Try to find symbol name for specified label
 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(MCInst &Inst,
                                 raw_ostream &/*cStream*/, int64_t Value,
                                 uint64_t /*Address*/, bool IsBranch,
                                 uint64_t /*Offset*/, uint64_t /*InstSize*/) {
 
   if (!IsBranch) {
     return false;
   }
 
   auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
   if (!Symbols)
     return false;
 
   auto Result = std::find_if(Symbols->begin(), Symbols->end(),
                              [Value](const SymbolInfoTy& Val) {
                                 return Val.Addr == static_cast<uint64_t>(Value)
                                     && Val.Type == ELF::STT_NOTYPE;
                              });
   if (Result != Symbols->end()) {
     auto *Sym = Ctx.getOrCreateSymbol(Result->Name);
     const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
     Inst.addOperand(MCOperand::createExpr(Add));
     return true;
   }
   return false;
 }
 
 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
                                                        int64_t Value,
                                                        uint64_t Address) {
   llvm_unreachable("unimplemented");
 }
 
 //===----------------------------------------------------------------------===//
 // Initialization
 //===----------------------------------------------------------------------===//
 
 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
                               LLVMOpInfoCallback /*GetOpInfo*/,
                               LLVMSymbolLookupCallback /*SymbolLookUp*/,
                               void *DisInfo,
                               MCContext *Ctx,
                               std::unique_ptr<MCRelocationInfo> &&RelInfo) {
   return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
 }
 
 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
                                                 const MCSubtargetInfo &STI,
                                                 MCContext &Ctx) {
   return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo());
 }
 
 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUDisassembler() {
   TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
                                          createAMDGPUDisassembler);
   TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
                                        createAMDGPUSymbolizer);
 }

llvm/test/CodeGen/AMDGPU/nop-data.ll

 ; RUN: llc -mtriple=amdgcn--amdhsa -mattr=-code-object-v3 -mcpu=fiji -filetype=obj < %s | llvm-objdump -d - --mcpu=fiji | FileCheck %s
 
 ; CHECK: <kernel0>:
-; CHECK-NEXT: s_endpgm
+; CHECK: s_endpgm
 define amdgpu_kernel void @kernel0() align 256 {
 entry:
   ret void
 }
 
 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0

 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0
 ; CHECK-NEXT: s_nop 0  // 0000000001FC: BF800000
 
 ; CHECK-EMPTY:
 ; CHECK-NEXT: <kernel1>:
-; CHECK-NEXT: s_endpgm
+; CHECK: s_endpgm
 define amdgpu_kernel void @kernel1(i32 addrspace(1)* addrspace(4)* %ptr.out) align 256 {
 entry:
   ret void
 }

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-failure.s

+;; Failure test. We create a malformed kernel descriptor (KD) by manually
+;; setting the bytes. Because in actuality, one can't create a malformed KD
+;; using the assembler directives.
+
+; RUN: llvm-mc %s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t.o
+
+; RUN: printf ".type  my_kernel.kd, @object \nmy_kernel.kd:\n.size my_kernel.kd, 64\n" > %t1.sym_info
+; RUN: llvm-objdump --disassemble-symbols=my_kernel.kd %t.o \
+; RUN: | tail -n +9 > %t1.sym_content
+; RUN: cat %t1.sym_info %t1.sym_content > %t1.s
+
+; RUN: llvm-mc %t1.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t-re-assemble.o
+; RUN: diff %t.o %t-re-assemble.o
+
+;; Test failure by setting one of the reserved bytes to non-zero value.
+
+.type my_kernel.kd, @object
+.size my_kernel.kd, 64
+my_kernel.kd:
+  .long 0x00000000           ;; group_segment_fixed_size
+  .long 0x00000000           ;; private_segment_fixed_size
+  .quad 0x00FF000000000000   ;; reserved bytes.
+  .quad 0x0000000000000000   ;; kernel_code_entry_byte_offset, any value works.
+
+  ;; 20 reserved bytes.
+  .quad 0x0000000000000000
+  .quad 0x0000000000000000
+  .long 0x00000000
+
+  .long 0x00000000           ;; compute_PGM_RSRC3
+  .long 0x00000000           ;; compute_PGM_RSRC1
+  .long 0x00000000           ;; compute_PGM_RSRC2
+  .short 0x0000              ;; additional fields.
+
+  ;; 6 reserved bytes.
+  .long 0x0000000
+  .short 0x0000

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-sgpr.s

+;; Test disassembly for GRANULATED_WAVEFRONT_SGPR_COUNT in the kernel descriptor.
+
+; RUN: split-file %s %t.dir
+
+; RUN: llvm-mc %t.dir/1.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_1.kd %t1 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1-re-assemble
+; RUN: diff %t1 %t1-re-assemble
+
+; RUN: llvm-mc %t.dir/2.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_2.kd %t2 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2-re-assemble
+; RUN: diff %t2 %t2-re-assemble
+
+; RUN: llvm-mc %t.dir/3.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t3
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_3.kd %t3 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t3-re-assemble
+; RUN: diff %t3 %t3-re-assemble
+
+
+;--- 1.s
+;; Only set next_free_sgpr
+.amdhsa_kernel my_kernel_1
+  .amdhsa_next_free_vgpr 0
+  .amdhsa_next_free_sgpr 42
+  .amdhsa_reserve_flat_scratch 0
+  .amdhsa_reserve_xnack_mask 0
+  .amdhsa_reserve_vcc 0
+.end_amdhsa_kernel
+
+;--- 2.s
+;; Only set other directives.
+.amdhsa_kernel my_kernel_2
+  .amdhsa_next_free_vgpr 0
+  .amdhsa_next_free_sgpr 0
+  .amdhsa_reserve_flat_scratch 1
+  .amdhsa_reserve_xnack_mask 1
+  .amdhsa_reserve_vcc 1
+.end_amdhsa_kernel
+
+;--- 3.s
+;; Set all affecting directives.
+.amdhsa_kernel my_kernel_3
+  .amdhsa_next_free_vgpr 0
+  .amdhsa_next_free_sgpr 35
+  .amdhsa_reserve_flat_scratch 1
+  .amdhsa_reserve_xnack_mask 1
+  .amdhsa_reserve_vcc 1
+.end_amdhsa_kernel

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-vgpr.s

+;; Test disassembly for GRANULATED_WORKITEM_VGPR_COUNT in the kernel descriptor.
+
+; RUN: split-file %s %t.dir
+
+; RUN: llvm-mc %t.dir/1.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_1.kd %t1 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1-re-assemble
+; RUN: diff %t1 %t1-re-assemble
+
+; RUN: llvm-mc %t.dir/2.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_2.kd %t2 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2-re-assemble
+; RUN: diff %t2 %t2-re-assemble
+
+; RUN: llvm-mc %t.dir/3.s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t3
+; RUN: llvm-objdump --disassemble-symbols=my_kernel_3.kd %t3 | tail -n +8 \
+; RUN: | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t3-re-assemble
+; RUN: diff %t3 %t3-re-assemble
+
+;--- 1.s
+.amdhsa_kernel my_kernel_1
+  .amdhsa_next_free_vgpr 23
+  .amdhsa_next_free_sgpr 0
+.end_amdhsa_kernel
+
+;--- 2.s
+.amdhsa_kernel my_kernel_2
+  .amdhsa_next_free_vgpr 14
+  .amdhsa_next_free_sgpr 0
+.end_amdhsa_kernel
+
+;--- 3.s
+.amdhsa_kernel my_kernel_3
+  .amdhsa_next_free_vgpr 32
+  .amdhsa_next_free_sgpr 0
+.end_amdhsa_kernel

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-zeroed-gfx10.s

+;; Entirely zeroed kernel descriptor (for GFX9).
+
+; RUN: llvm-mc %s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx1010 -filetype=obj -o %t
+; RUN: llvm-objdump -s -j .text %t | FileCheck --check-prefix=OBJDUMP %s
+
+;; TODO:
+;; This file and kd-zeroed-raw.s should produce the same output for the kernel
+;; descriptor - a block of 64 zeroed bytes. But looks like the assembler sets
+;; the FWD_PROGRESS bit in COMPUTE_PGM_RSRC1 to 1 even when the directive
+;; mentions 0 (see line 36).
+
+;; Check the raw bytes right now.
+
+; OBJDUMP:      0000 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0010 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0020 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0030 01000000 00000000 00000000 00000000
+
+.amdhsa_kernel my_kernel
+  .amdhsa_group_segment_fixed_size 0
+  .amdhsa_private_segment_fixed_size 0
+  .amdhsa_next_free_vgpr 8
+  .amdhsa_reserve_vcc 0
+  .amdhsa_reserve_flat_scratch 0
+  .amdhsa_reserve_xnack_mask 0
+  .amdhsa_next_free_sgpr 8
+  .amdhsa_float_round_mode_32 0
+  .amdhsa_float_round_mode_16_64 0
+  .amdhsa_float_denorm_mode_32 0
+  .amdhsa_float_denorm_mode_16_64 0
+  .amdhsa_dx10_clamp 0
+  .amdhsa_ieee_mode 0
+  .amdhsa_fp16_overflow 0
+  .amdhsa_workgroup_processor_mode 0
+  .amdhsa_memory_ordered 0
+  .amdhsa_forward_progress 0
+  .amdhsa_system_sgpr_private_segment_wavefront_offset 0
+  .amdhsa_system_sgpr_workgroup_id_x 0
+  .amdhsa_system_sgpr_workgroup_id_y 0
+  .amdhsa_system_sgpr_workgroup_id_z 0
+  .amdhsa_system_sgpr_workgroup_info 0
+  .amdhsa_system_vgpr_workitem_id 0
+  .amdhsa_exception_fp_ieee_invalid_op 0
+  .amdhsa_exception_fp_denorm_src 0
+  .amdhsa_exception_fp_ieee_div_zero 0
+  .amdhsa_exception_fp_ieee_overflow 0
+  .amdhsa_exception_fp_ieee_underflow 0
+  .amdhsa_exception_fp_ieee_inexact 0
+  .amdhsa_exception_int_div_zero 0
+  .amdhsa_user_sgpr_private_segment_buffer 0
+  .amdhsa_user_sgpr_dispatch_ptr 0
+  .amdhsa_user_sgpr_queue_ptr 0
+  .amdhsa_user_sgpr_kernarg_segment_ptr 0
+  .amdhsa_user_sgpr_dispatch_id 0
+  .amdhsa_user_sgpr_flat_scratch_init 0
+  .amdhsa_user_sgpr_private_segment_size 0
+  .amdhsa_wavefront_size32 0
+.end_amdhsa_kernel

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-zeroed-gfx9.s

+;; Entirely zeroed kernel descriptor (for GFX9).
+
+; RUN: llvm-mc %s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1
+; RUN: llvm-objdump --arch-name=amdgcn --mcpu=gfx908 --disassemble-symbols=my_kernel.kd %t1 \
+; RUN: | tail -n +8 | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2
+; RUN: diff %t1 %t2
+
+; RUN: llvm-objdump -s -j .text %t1 | FileCheck --check-prefix=OBJDUMP %s
+
+; OBJDUMP:      0000 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0010 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0020 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0030 00000000 00000000 00000000 00000000
+
+;; This file and kd-zeroed-raw.s produce the same output for the kernel
+;; descriptor - a block of 64 zeroed bytes.
+
+.amdhsa_kernel my_kernel
+  .amdhsa_group_segment_fixed_size 0
+  .amdhsa_private_segment_fixed_size 0
+  .amdhsa_next_free_vgpr 0
+  .amdhsa_reserve_vcc 0
+  .amdhsa_reserve_flat_scratch 0
+  .amdhsa_reserve_xnack_mask 0
+  .amdhsa_next_free_sgpr 0
+  .amdhsa_float_round_mode_32 0
+  .amdhsa_float_round_mode_16_64 0
+  .amdhsa_float_denorm_mode_32 0
+  .amdhsa_float_denorm_mode_16_64 0
+  .amdhsa_dx10_clamp 0
+  .amdhsa_ieee_mode 0
+  .amdhsa_fp16_overflow 0
+  .amdhsa_system_sgpr_private_segment_wavefront_offset 0
+  .amdhsa_system_sgpr_workgroup_id_x 0
+  .amdhsa_system_sgpr_workgroup_id_y 0
+  .amdhsa_system_sgpr_workgroup_id_z 0
+  .amdhsa_system_sgpr_workgroup_info 0
+  .amdhsa_system_vgpr_workitem_id 0
+  .amdhsa_exception_fp_ieee_invalid_op 0
+  .amdhsa_exception_fp_denorm_src 0
+  .amdhsa_exception_fp_ieee_div_zero 0
+  .amdhsa_exception_fp_ieee_overflow 0
+  .amdhsa_exception_fp_ieee_underflow 0
+  .amdhsa_exception_fp_ieee_inexact 0
+  .amdhsa_exception_int_div_zero 0
+  .amdhsa_user_sgpr_private_segment_buffer 0
+  .amdhsa_user_sgpr_dispatch_ptr 0
+  .amdhsa_user_sgpr_queue_ptr 0
+  .amdhsa_user_sgpr_kernarg_segment_ptr 0
+  .amdhsa_user_sgpr_dispatch_id 0
+  .amdhsa_user_sgpr_flat_scratch_init 0
+  .amdhsa_user_sgpr_private_segment_size 0
+.end_amdhsa_kernel

llvm/test/tools/llvm-objdump/ELF/AMDGPU/kd-zeroed-raw.s

+; RUN: llvm-mc %s -mattr=+code-object-v3 --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t1
+; RUN: llvm-objdump --disassemble-symbols=my_kernel.kd %t1 \
+; RUN: | tail -n +8 | llvm-mc --triple=amdgcn-amd-amdhsa -mcpu=gfx908 -filetype=obj -o %t2
+; RUN: llvm-objdump -s -j .text %t2 | FileCheck --check-prefix=OBJDUMP %s
+
+;; Not running lit-test over gfx10 (see kd-zeroed-gfx10.s for details).
+;; kd-zeroed-raw.s and kd-zeroed-*.s should produce the same output for the
+;; kernel descriptor - a block of 64 zeroed bytes.
+
+;; The disassembly will produce the contents of kd-zeroed-*.s which on being
+;; assembled contains additional relocation info. A diff over the entire object
+;; will fail in this case. So we check by looking the bytes in .text.
+
+; OBJDUMP:      0000 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0010 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0020 00000000 00000000 00000000 00000000
+; OBJDUMP-NEXT: 0030 00000000 00000000 00000000 00000000
+
+;; The entire object is zeroed out.
+
+.type my_kernel.kd, @object
+.size my_kernel.kd, 64
+my_kernel.kd:
+  .long 0x00000000           ;; group_segment_fixed_size
+  .long 0x00000000           ;; private_segment_fixed_size
+  .quad 0x0000000000000000   ;; reserved bytes.
+  .quad 0x0000000000000000   ;; kernel_code_entry_byte_offset, any value works.
+
+  ;; 20 reserved bytes.
+  .quad 0x0000000000000000
+  .quad 0x0000000000000000
+  .long 0x00000000
+
+  .long 0x00000000           ;; compute_PGM_RSRC3
+  .long 0x00000000           ;; compute_PGM_RSRC1
+  .long 0x00000000           ;; compute_PGM_RSRC2
+  .short 0x0000              ;; additional fields.
+
+  ;; 6 reserved bytes.
+  .long 0x0000000
+  .short 0x0000

llvm/tools/llvm-objdump/llvm-objdump.cpp

for (unsigned SI = 0, SE = Symbols.size(); SI != SE; ++SI) {
 
       if (!PrintedSection) {
         PrintedSection = true;
         outs() << "\nDisassembly of section ";
         if (!SegmentName.empty())
           outs() << SegmentName << ",";
         outs() << SectionName << ":\n";
       }
 
-      if (Obj->isELF() && Obj->getArch() == Triple::amdgcn) {
-        if (Symbols[SI].Type == ELF::STT_AMDGPU_HSA_KERNEL) {
-          // skip amd_kernel_code_t at the begining of kernel symbol (256 bytes)
-          Start += 256;
-        }
-        if (SI == SE - 1 ||
-            Symbols[SI + 1].Type == ELF::STT_AMDGPU_HSA_KERNEL) {
-          // cut trailing zeroes at the end of kernel
-          // cut up to 256 bytes
-          const uint64_t EndAlign = 256;
-          const auto Limit = End - (std::min)(EndAlign, End - Start);
-          while (End > Limit &&
-            *reinterpret_cast<const support::ulittle32_t*>(&Bytes[End - 4]) == 0)
-            End -= 4;
-        }
-      }
-
       outs() << '\n';
       if (!NoLeadingAddr)
         outs() << format(Is64Bits ? "%016" PRIx64 " " : "%08" PRIx64 " ",
                          SectionAddr + Start + VMAAdjustment);
       if (Obj->isXCOFF() && SymbolDescription) {
         outs() << getXCOFFSymbolDescription(Symbols[SI], SymbolName) << ":\n";
       } else
         outs() << '<' << SymbolName << ">:\n";