[RattataKing]

Adds the five single-dword GLOBAL LDS loads on CDNA3 and CDNA4 in the generator:

• global_load_lds_ubyte
• global_load_lds_sbyte
• global_load_lds_ushort
• global_load_lds_sshort
• global_load_lds_dword

Add tests:

• GlobalLoadsToLdsPreserveInactiveLanesAndVgprs

Issue ID: #13025

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[kuhar]

We should update the test documentation to match the new execution support:

• emulation/rocjitsu/tests/flat_lds_transfer_test.cpp:372: replace the claim that GLOBAL LDS execution remains unimplemented and belongs to PR #13218. This case checks decoding; execution is now covered by GlobalLoadsToLdsPreserveInactiveLanesAndVgprs and the CDNA4 wide-load tests.
• Optional: emulation/rocjitsu/tests/flat_lds_transfer_test.cpp:271: make the skip reason explicit, for example “This decoder boundary is specific to CDNA3.” That explains why other architectures skip, whereas “Checks the CDNA3 decoder boundary” only describes the test.

[kuhar]

Left one note to consider

[kuhar]

Could we trim the duplicate X3/X4 code now that #14842 is in the base? This PR adds execution for the five single-dword destination forms. The generated CDNA4 X3/X4 bodies already match the base.

• In emulation/rocjitsu/lib/python/amdisa/semantics.py:2177–2178, we can drop the two wide suffixes. Line 2152 handles them first, so editing the later entries would leave both opcodes unchanged.
• In emulation/rocjitsu/tests/flat_lds_transfer_test.cpp:77, the CDNA3 rejection test can reuse global_wide_lds_encoding from line 71. The wide tests at lines 475 and 516 already cover stride, X3 padding, inactive lanes and whole-payload bounds. Folding the exact-fit and one-byte-short cases from line 254 into the existing bounds test would preserve that coverage with one suite to maintain.

This cleanup can follow separately.

LGTM

[peymanr]

PR Pundit review

(Testing PR Pundit Code Review agent. Not all comments may be relevant. This is a one-time automated pass: please address what is useful, no need to request a re-review. Please reply to each point, whether it applies or not: we harvest these replies, and both kinds are enormously helpful for tuning the agent. Reach out to @peymanr for suggestions.)

#		Finding	Where
1	🟠	Narrow GLOBAL LDS regression on subword/sign-extend/M0 would go uncaught: this PR's test uses only aligned addresses and M0=64	emulation/rocjitsu/tests/flat_lds_transfer_test.cpp:157

1. 🟠 Should fix: Narrow GLOBAL LDS regression on subword/sign-extend/M0 would go uncaught: this PR's test uses only aligned addresses and M0=64

emulation/rocjitsu/tests/flat_lds_transfer_test.cpp:157

The new GlobalLoadsToLdsPreserveInactiveLanesAndVgprs test only uses dword-aligned source addresses, payload bytes/shorts with the top bit set (0x81, 0x8081), offsets -4/0/4 and a clean M0=64. So subword addr&3 extraction, zero- vs sign-extension of positive values, and M0 masking are not exercised on the GLOBAL path specifically. The generated GlobalLoadLds* execute bodies are identical to the ScratchLoadLds* ones, which ScratchWidthsAndAddressModesPreserveInactiveLanesAndVgprs covers with per-lane subwords and M0=0xfffc0043, so the risk is low. Adding a subword offset, a positive payload and an unaligned M0 to the GLOBAL test would make it independent of the scratch coverage.

Who is affected: Only test coverage of the emulator's CDNA3/CDNA4 narrow GLOBAL_LOAD_LDS_* path; no runtime behaviour is affected.

Checked against the code by an independent pass.

Related past reviews: #6886 on emulation/rocjitsu/lib/python/amdisa/tests/test_sema_derive.py, #6886 on emulation/rocjitsu/lib/python/amdisa/codegen/_generator.py, #10199 on emulation/rocjitsu/lib/rocjitsu/src/rocjitsu/isa/arch/amdgpu/cdna5/addr_calc.cpp, #10847 on emulation/rocjitsu/lib/rocjitsu/src/rocjitsu/isa/arch/amdgpu/cdna5/addr_calc.cpp

_{Shared from a pilot of PR Pundit, an agentic code reviewer that builds and tests the change and reads the project's past reviews. Reviewed at 4733bb13; findings checked and posted by @peymanr; points already raised on this PR are left out.}

[kuhar]

Could we update the PR description to say this adds the five single-dword GLOBAL LDS loads on CDNA3 and CDNA4? The CDNA4 X3/X4 instructions landed in #14842. This revision extends their existing tests with cases such as exact-fit and one-byte-short LDS allocations, so the reference to Cdna4WideGlobalLoadsUseSixteenByteLdsLaneStride can go too.

LGTM