Keep a finished divide in Execute when a flush arrives during StallM by davidharrishmc · Pull Request #1925 · openhwfoundation/cvw

davidharrishmc · 2026-10-05T15:55:46Z

Merge together with #1941, which contains this PR's commits. The stall injector added here also exposes a bug #1941 fixes: during an injected stall, trap CSRs (mepc/mcause, sepc/scause/stval) can be written for a trap that is never taken. With this PR alone, lockstep can report such a mismatch at a test's final ecall when the injected stall lands there. Merging #1941 merges both; if this PR is merged first, merge #1941 immediately after.

A divide that the BTB predicts as a jump raises FlushE in its DONE cycle. If StallM held it in Execute, the divider FSMs went IDLE and the divide restarted with stale forwarded operands: 100/7 gave 0, and an fdiv could flip sign. The FSMs now reset only when the Execute pipeline register clears (FlushE & ~StallM, which equals FlushE & ~StallE because a flush clears the divider's own Execute stall), which also makes a coverage exclusion obsolete.

This is reachable only through ExternalStall (the RVVI packetizer in the FPGA builds). The PR therefore adds an opt-in testbench stall injector, triggered by the custom-use HINT slti x0, x0, imm, and two tests:

divflushstall on rv64gc, which divides on the FPU (fdivsqrtfsm.sv)
divflushstall32 on rv32gc, which uses the integer divider (div.sv)

rv64gc never exercises div.sv, and cvw had no RV32 coverage tests, so this adds tests/coverage32/, built by make coverage and run on rv32gc in regression-wally's nightly lockstep run; divflushstall32 reports a mismatch through tohost. It is a separate directory because the nightly lockstep run executes every ELF under tests/coverage/ on rv64gc. Part of #1921.

🤖 Generated with Claude Code

A divide mispredicted as a jump raises FlushE in its DONE cycle; when StallM held it in Execute the divider FSM went IDLE and restarted with stale operands. Reset the FSMs only when Execute advances; add an opt-in testbench ExternalStall injector and the divflushstall test. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

…v.sv on rv32gc Use FlushE & ~StallE, matching flopenrc, which clears only when enabled. rv64gc divides on the FPU, so add divflushstall32 for rv32gc to exercise div.sv. RV32 self-checking tests live in tests/coverage32 (built by 'make coverage', run by the self-checked coverage32gc suite) so the nightly lockstep run over tests/coverage on rv64gc does not pick them up. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

StallE equals StallM whenever FlushE is asserted: FlushECause masks the divider's own StallECause, and LatestUnstalledE implies ~StallE. FlushE & ~StallM is therefore the same condition, and it needs no new StallE port through fpu, fdivsqrt, mdu, the core and testbench_fp. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

rosethompson · 2026-10-06T19:50:40Z

+                                 {StallLength, StallDelay} <= InstrM[31:20];
+      else if (StallDelay != 0)  StallDelay <= StallDelay - 1;
+      else if (StallLength != 0) StallLength <= StallLength - 1;
+    assign RVVIStall = (StallDelay == 0) & (StallLength != 0);


Since we are repurposing the RVVIStall, would it make sense to rename this ExternalStall? It won't break my hardware RVVI because it uses it's own tb and top level.

Good idea, done in b01c3bd. The testbench net is now ExternalStall, since it comes from either the RVVI wrapper or the stall injector. The wrapper's port keeps the name RVVIStall. #1930 has the same change merged in.

The net now has two drivers, the RVVI wrapper and the stall injector, so name it after the core port it drives. rvvitbwrapper keeps its RVVIStall port. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

Brings in the ExternalStall rename of the testbench net. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

…iv-flush-stall Port the PR to main's ELF-based test flow: - testbench/tests.vh is gone, so drop the coverage32gc list and its testbench.sv case; keep the ExternalStall injector and the RVVIStall -> ExternalStall wiring on main's testbench. - tests/coverage/divflushstall.S runs from tests/coverage like the other coverage tests and reports through tohost via WALLY-selfcheck.h. - tests/coverage32 is built by the root Makefile's coverage target (which regression-wally's buildTests uses), and regression-wally --nightly runs it on rv32gc in lockstep next to the rv32gc ACT tests, as it runs tests/coverage on rv64gc. - divflushstall32.S reports failure to tohost ((2 << 1) | 1) instead of always storing 1, so a mismatch fails the run under main's convention. - coverage-exclusions-rv64gc.do keeps main's /core scopes and drops the fdivsqrtfsm StallM exclusions that divflushstall now covers. Signed-off-by: David Harris <David_Harris@hmc.edu> Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…nhwfoundation#1895) into dh/wfi-wait dh/div-flush-stall carries upstream/main through 20f2af7 (riscof retired, openhwfoundation#1895), so this also brings dh/wfi-wait up to date with main. Resolution: - testbench/tests.vh is gone; the new wfi/trap tests run from tests/coverage like the other coverage tests and report through tohost via main's WALLY-selfcheck.h. - lockstepwaivers keeps main's entries and adds wfiTW.elf, wfitimeoutint.elf and wfitimeoutnext.elf (matched by ELF file name); tests/coverage32 runs on rv32gc in nightly lockstep as in openhwfoundation#1925. - csrc.sv keeps the taken-trap HPM events (TrapM & ~StallW) with main's division-cycles comment and exclusion in place of coverage off/on. - coverage-exclusions-rv64gc.do takes main's version (/core scopes; the D$ SetValidEN/ClearValidEN exclusions whose comment this branch edited are gone on main) and drops the stale fdivsqrtfsm StallM exclusions that c7cbd78 had brought back although the exclusion tag is gone. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: David Harris <David_Harris@hmc.edu>

This was referenced Oct 5, 2026

Record trap CSRs and trap HPM events only when the trap is taken (~StallW) #1930

Closed

Pipeline review: deferred-exception, WFI, divider, trap-CSR and MMU bugs with directed tests #1921

Open

davidharrishmc and others added 2 commits October 6, 2026 06:42

rosethompson reviewed Oct 6, 2026

View reviewed changes

davidharrishmc marked this pull request as draft October 6, 2026 22:26

davidharrishmc marked this pull request as ready for review October 6, 2026 22:27

This was referenced Oct 7, 2026

Wait for wfi in one place; record traps only when taken #1941

Open

Implement the Zawrs extension (rva23) #1942

Open

Navigation Menu

Repository navigation

Keep a finished divide in Execute when a flush arrives during StallM - #1925

Keep a finished divide in Execute when a flush arrives during StallM#1925
davidharrishmc wants to merge 5 commits into
openhwfoundation:mainfrom
davidharrishmc:dh/div-flush-stall

davidharrishmc commented Oct 5, 2026 •

edited

Loading

Uh oh!

rosethompson Oct 6, 2026

Uh oh!

davidharrishmc Oct 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

davidharrishmc commented Oct 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rosethompson Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

davidharrishmc Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

davidharrishmc commented Oct 5, 2026 •

edited

Loading