gpu-compute: stop decoding instructions after stopFetch

[Basemism]

Fetch has two steps: request/cache instruction bytes, then decode buffered bytes into the wavefront instruction buffer. Wavefront::stopFetch() already treated branches, returns, and s_endpgm as boundaries, but decodeInsts() checked it
only when scheduling future fetch requests. Its inner loop continued consuming bytes already buffered after a boundary.

When s_endpgm and following bytes occupied the same fetched region, gem5 decoded those following bytes under the ending kernel's wavefront. Constructing the following instruction immediately mapped its operands against that kernel's
valid reservation and produced a false panic. The instruction would never have executed: s_endpgm erases later buffered instructions and terminates the wave.

Fix:
Honor Wavefront::stopFetch() before decoding either a split instruction or additional instructions from the fetch buffer. This prevents the decoder from continuing to populate the instruction buffer after the wavefront has entered a state that blocks further fetch and decode work.

[Copilot]

🔵 Needs a closer look

The new stopFetch() usage in the while condition can introduce avoidable O(n²) behavior in a hot decode loop due to repeated full scans of the instruction buffer.

Pull request overview

This PR fixes a GPU fetch/decode boundary bug in the shader instruction fetch unit: once a wavefront has buffered a control-flow/end-of-kernel instruction (e.g., s_endpgm), decoding now stops immediately instead of continuing to consume already-fetched bytes and populating the instruction buffer with instructions that should never execute.

Changes:

• Gate FetchBufDesc::decodeInsts() so it does not decode split or subsequent instructions after Wavefront::stopFetch() becomes true.
• Prevent decoding of post-s_endpgm bytes that can incorrectly map operands and trigger false panics.

File summaries

File	Description
src/gpu-compute/fetch_unit.cc	Stops decode from consuming buffered bytes past branch/return/end-of-kernel boundaries by honoring Wavefront::stopFetch() during decode.

Review details

Suppressed comments (1)

src/gpu-compute/fetch_unit.cc:597

• wavefront->stopFetch() is an O(n) scan over instructionBuffer (see Wavefront::stopFetch()), and calling it in the while condition can make decodeInsts() potentially O(n^2) per buffer fill (one full scan per decoded instruction). You can keep the same correctness behavior while avoiding repeated scans by checking stopFetch() once up-front, then breaking based on the newly-decoded instruction being a branch/return/end-of-kernel.

           hasFetchDataToProcess() && !wavefront->stopFetch()) {
        if (splitDecode()) {
            decodeSplitInst();
        } else {
            TheGpuISA::MachInst mach_inst =

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.