1
0
Commit Graph

280 Commits

Author SHA1 Message Date
Wesley W. Terpstra
4807ce7ced dcache: put a flow Q to absorb back-pressure without restarting pipeline (#701)
* dcache: put a flow Q to absorb back-pressure without restarting pipeline

When used with a RationalCrossing, pipelined MMIO does not come out cleanly.
The first beat works, but if the second beat gets stalled, the pipeline is
restarted. This is a quick hacky test to absorb the beats. Perhaps a better
fix can be made to achieve the same effect.

* dcache: provision as few stages as possible
2017-04-24 23:28:04 -07:00
Wesley W. Terpstra
9c1d126965 Allow speculative fetch to uncacheable memory if it hits in I$ (#700)
@aswaterman it's in
2017-04-24 19:12:37 -07:00
Andrew Waterman
65928dc6a0 Don't push RAS for "auipc ra, X; jalr ra, ra, Y" 2017-04-24 02:01:15 -07:00
Andrew Waterman
36a7971975 Bypass scoreboard to reduce MMIO latency 2017-04-24 02:01:15 -07:00
Andrew Waterman
f2d4cb8152 Update RAS speculatively from fetch stage 2017-04-24 02:01:15 -07:00
Andrew Waterman
c36c171202 Use correct interrupt priority order 2017-04-24 02:01:15 -07:00
Andrew Waterman
bf861293d9 Add ShiftQueue; use it 2017-04-24 02:01:15 -07:00
Andrew Waterman
d24d8ff84b Don't stall the frontend, making it easier to add more features later 2017-04-24 02:01:15 -07:00
Andrew Waterman
061a0adceb Fetch smaller parcels from the I$ 2017-04-24 02:01:15 -07:00
Megan Wachs
c72b15f2a0 Down with any require() statement that makes me RTFC 2017-04-21 15:44:42 -07:00
Andrew Waterman
67404a665b When not using a cache, LR/SC isn't legal even on cacheable memory 2017-04-20 08:47:03 -07:00
Andrew Waterman
d82a0dc231 Mitigate D$ exception critical path, yet again 2017-04-18 00:47:58 -07:00
Andrew Waterman
c99ce7ce5d Only report D$ exceptions on not-nacked accesses 2017-04-18 00:47:58 -07:00
Andrew Waterman
a956b78dd2 In TLBPermissions, merge across some region types
We only care whether they have side effects or not.
2017-04-18 00:47:58 -07:00
Andrew Waterman
6de6f38894 Pipeline D$ exception response into s2 2017-04-18 00:47:58 -07:00
Andrew Waterman
cc9ec1d51a Send D$ grant acks early; accept release acks early
We now need to block the B-channel for a few cycles after a grant to
allow the processor to get at least one request through, preventing
livelock.
2017-04-18 00:47:58 -07:00
Andrew Waterman
728569c717 Reduce access-exception generation critical path 2017-04-18 00:47:58 -07:00
Andrew Waterman
c366007a0d Tighten PMAs for LR/SC and misaligned accesses
- LR/SC on cacheable memory only (not even scratchpad)
- No misaligned accesses on regions with get/put-effects
2017-04-18 00:47:58 -07:00
Andrew Waterman
74a7838de0 In TLBPermissions, don't merge regions of different types 2017-04-18 00:47:58 -07:00
Andrew Waterman
7871ec82c4 Guarantee probe forward progress during LR storm 2017-04-18 00:47:58 -07:00
Andrew Waterman
debcbca7de Make PMP tolerant to PA size << VA size 2017-04-17 10:28:33 -07:00
Andrew Waterman
a454edaaf7 Treat exceptions as steps for the purposes of single-stepping 2017-04-17 10:28:33 -07:00
Wesley W. Terpstra
2f22fca615 rocket: reverse input edge for better output 2017-04-14 18:09:14 -07:00
Andrew Waterman
fdfcffb0b2 Catch bad physical address MSBs when VA size > PA size 2017-04-14 01:03:11 -07:00
Andrew Waterman
6fbbccca3e Improve Seq indexing QoR 2017-04-14 01:03:11 -07:00
Andrew Waterman
d203c4c654 Check AMO operation legality in TLB 2017-04-14 01:03:11 -07:00
Yunsup Lee
6359ff96e5 Several ScratchpadSlavePort bug fixes (#676)
* only replicate scratch slave d-channel resp when AMO req

* dtim: port can't support put partial mask with holes

* dtim: use \!isRead instead of isAMO

* Fix ScratchpadSlavePort looking at wrong Acquire message

Rename acq to a in the helper method.
Delete isRead and isWrite altogether.
2017-04-13 23:25:51 -07:00
Andrew Waterman
b9e042d2bf Unconditionally write badaddr, possibly to zero
59d33f6b83
2017-04-12 13:35:02 -07:00
Andrew Waterman
470c6711a7 Do some CSE by hand, per @terpstra 2017-04-10 22:38:25 -07:00
Andrew Waterman
a43bf2feae Add vectored interrupt support
4dcaa944ba

I also added a test, which does indeed pass, but I don't want to bump
riscv-tools for that alone:

ba6d88466a
2017-04-08 00:29:45 -07:00
Andrew Waterman
c861c4925e Don't signal access exceptions on invalid PTEs
The PPN should not be interpreted in this case.
2017-04-05 21:46:55 -07:00
Andrew Waterman
2e09253d26 Revive I$ parity option
Pipeline the parity check into the second stage, so that the data
RAM access + parity check do not become the critical path.
2017-04-05 21:46:55 -07:00
Andrew Waterman
43917dd59f Get I$ s1_kill signal off the critical path 2017-04-05 21:46:55 -07:00
Andrew Waterman
744fb2e4b9 Cut imem.resp.ready critical path with a flow queue
This is only necessary for RVC, where the decode latency is much higher.
2017-04-05 21:46:55 -07:00
Andrew Waterman
3e72f9779f Handle single-step with a pipeline stall, not a flush
The pipeline flush approach broke when I changed the pipeline stage
the flush happens from
2017-04-05 19:52:44 -07:00
solomatnikov
127f121ef2 Preserve id_do_fence (#651) 2017-04-05 08:29:45 -07:00
Andrew Waterman
19f0ae64a0 Only set id_reg_fence when AMO/FENCE is actually executed
This is a performance bug, not a correctness bug.  But randomly stalling
because of garbage bits coming out of the I$ should be avoided.

h/t @solomatnikov
2017-04-03 21:13:52 -07:00
Andrew Waterman
410e9cf736 I$ bugfix, to be reworked 2017-03-31 12:17:41 -07:00
Henry Cook
b9550e8523 Merge branch 'master' into name-rams 2017-03-30 17:36:01 -07:00
Andrew Waterman
a8a2ee711c Give I$ RAMs consistent names 2017-03-30 15:50:54 -07:00
Andrew Waterman
2720095b8e Give D$ RAMs consistent names 2017-03-30 15:49:14 -07:00
Andrew Waterman
70e7e90c02 Remove splitMetadata option from L1 caches
This is a property of the specific cache microarchitecture, not actually
an independently tunable knob.
2017-03-30 15:48:55 -07:00
Megan Wachs
9de06f8c83 Merge remote-tracking branch 'origin/master' into debug_v013_pr 2017-03-30 08:01:11 -07:00
Andrew Waterman
fd39eadcd6 New PMP encoding 2017-03-30 00:36:23 -07:00
Wesley W. Terpstra
2f2b472098 rocket: split the interrupt controller into its own node 2017-03-30 00:36:23 -07:00
Andrew Waterman
3546c8d133 If any PMPs are supported, all CSRs exist 2017-03-30 00:36:23 -07:00
Andrew Waterman
8f73a58d90 Report access exception, not page fault, if page-table walk fails 2017-03-30 00:36:23 -07:00
Andrew Waterman
25232070ec Don't redundantly set resp_ae in PTW 2017-03-30 00:36:23 -07:00
Henry Cook
d3bc99e253 get local interrupts out of the tile 2017-03-30 00:36:23 -07:00
solomatnikov
0b9fc94421 Assertion for back-to-back uncached and cached ops (#631) 2017-03-29 23:07:17 -07:00
Megan Wachs
d8033b20fc Merge remote-tracking branch 'origin/master' into debug_v013_pr 2017-03-29 14:58:04 -07:00
Andrew Waterman
44fb3be7d0 Fix MMIO/cache refill concurrency bug in DCache
There's a structural hazard on s2_req, so disallow cache refill initiation
while any MMIO loads are in flight.
2017-03-28 17:16:29 -07:00
Andrew Waterman
4215f480ef Write instruction to badaddr on illegal instruction traps 2017-03-28 00:56:14 -07:00
Megan Wachs
bb64c92906 csr: Bring functionality in line with v13 spec. ebreak does not cause exception in Debug Mode, it just starts at Debug ROM again. 2017-03-27 21:21:48 -07:00
Andrew Waterman
05cbdced78 Work around zero-entry vec issue in Chisel 2017-03-27 17:57:26 -07:00
Andrew Waterman
d42d8aaea7 Make SEIP writable 2017-03-27 16:37:09 -07:00
Andrew Waterman
c7c357e716 Add local interrupts to core (but not yet to coreplex) 2017-03-27 16:37:09 -07:00
Andrew Waterman
069858a20c rocket: separate page faults from physical memory access exceptions 2017-03-27 16:37:09 -07:00
Andrew Waterman
ea0714bfcb rocket: hard-wire UXL/SXL fields to 0
a2a3346e73
2017-03-27 16:37:09 -07:00
Wesley W. Terpstra
75eba294ec DCache: Release from the correct ID as well 2017-03-27 16:30:46 -07:00
Wesley W. Terpstra
4959771c97 Revert "For D$, use source 0 through N-1 for MMIO, not 1 through N"
This reverts commit 0538dc77ce.
2017-03-27 16:30:46 -07:00
Wesley W. Terpstra
fa7ead6357 Revert "Use Reg(Vec) instead of Seq(Reg) for DCache MMIO"
This reverts commit fb6498f2c3.
2017-03-27 16:30:46 -07:00
Megan Wachs
08c4f7cea6 RocketTile: Create a wrapper for SyncRocketTile as well (#616)
* RocketTile: Create a wrapper for SyncRocketTile as well

There is no guarantee that debugInterrupt is synchronous
to tlClk, even though it is true in the current implementation.
It will not be true in future implementations, as decoupling
this allows the debugInterrupt to be asserted across tlClk
gating/reset scenarios.

Therefore, even for SyncRocketTile, the debug interrupt needs to be
synchronized to coreClk, and for RationalRocketTile, 1 cycle
of synchronization is not sufficient.

Even though other interrupts may be synchronized, we just
synchronize them all to simplify the code at the expense of
a few cycles latency.

It could still be nice to use a parameter vs hard coding "3".

* RocketTile: Actually use the SyncRocketTile wrapper to get properly synchronized resets.
2017-03-27 02:45:37 -07:00
Andrew Waterman
5d1165c850 Express PMP mask generator using a carry chain
This allows it to be optimized like an adder, improving QoR when it
is on the critical path.
2017-03-26 14:20:16 -07:00
Henry Cook
996a31364a rocket: remove hard-coded paddrBits (#610)
Fall back on global variable but check that it is compatible with memory as seen from rocket's tilelink master port.
2017-03-24 22:30:18 -07:00
Andrew Waterman
17b1ee3037 Default to 8 PMPs; support 0 PMPs 2017-03-24 16:39:52 -07:00
Andrew Waterman
97006ab396 Don't modulate PMP privilege on passsthrough when !usingVM 2017-03-24 16:39:52 -07:00
Andrew Waterman
3f0d2fe826 Instantiate PTW unconditionally
This keeps the PMP datapaths intact.  The PTW itself will get optimized
away for the !usingVM case.
2017-03-24 16:39:52 -07:00
Andrew Waterman
30415215b8 Don't check for exceptions on ScratchpadSlavePort accesses 2017-03-24 16:39:52 -07:00
Andrew Waterman
ccd5bc9a91 Improve QoR of PMP homogeneity checker 2017-03-24 16:39:52 -07:00
Andrew Waterman
e9cadf29d2 Improve DCache MMIO QoR
No need to store the cmd field.  From the perspective of the cache, all
MMIO responses that have data can be treated the same as loads.
2017-03-24 16:39:52 -07:00
Andrew Waterman
fb6498f2c3 Use Reg(Vec) instead of Seq(Reg) for DCache MMIO 2017-03-24 16:39:52 -07:00
Andrew Waterman
0538dc77ce For D$, use source 0 through N-1 for MMIO, not 1 through N
This makes the code a bit cleaner.
2017-03-24 16:39:52 -07:00
Andrew Waterman
3951e57789 Force each TLB entry into its own clock-gate group
This ameliorates a PMP critical path.

I can't figure out how to do this without asUInt/asTypeOf.
2017-03-24 16:39:52 -07:00
Andrew Waterman
8d7f1d777e Fix an embarrassing typo in the PMPHeterogeneityChecker 2017-03-24 16:39:52 -07:00
Andrew Waterman
10c39cb8d6 Disable mprv in D-mode 2017-03-24 16:39:52 -07:00
Andrew Waterman
d3bda9574c Put page homogeneity checker in PMP
Avoids redundancy between ITLB and DTLB
2017-03-24 16:39:52 -07:00
Andrew Waterman
9e05200e51 Don't require that PMP ranges be aligned to access size
e.g., if a range permits access to 0x0-0xb, allow 8-byte accesses 0x0-0x7.
2017-03-24 16:39:52 -07:00
Andrew Waterman
29e67279ba add comments 2017-03-24 16:39:52 -07:00
Andrew Waterman
4c8be13a4d Improve homogeneity circuit QoR 2017-03-24 16:39:52 -07:00
Andrew Waterman
59d6afa132 mideleg/medeleg not present without less-privileged traps 2017-03-24 16:39:52 -07:00
Andrew Waterman
38808f55d5 Share PMP mask gen between I$ and D$ 2017-03-24 16:39:52 -07:00
Andrew Waterman
86d84959cf More WIP on PMP 2017-03-24 16:39:52 -07:00
Andrew Waterman
2888779422 Flush pipeline from WB stage, not MEM
Fixes sptbr write -> instruction translation hazard.
2017-03-24 16:39:52 -07:00
Andrew Waterman
44ca3b60ab Retime PTW response valid bits
It's not just to save the gate delay; it also reduces wire delay by
allowing the flops to be closer to their respective TLBs.
2017-03-24 16:39:52 -07:00
Andrew Waterman
a03556220c Default TLB size = 32
@davidbiancolin
2017-03-24 16:39:52 -07:00
Andrew Waterman
1875407316 Get TLB permission checks off D$ clock gating critical path 2017-03-24 16:39:52 -07:00
Andrew Waterman
a4164348b4 Expose MXR to S-mode 2017-03-24 16:39:52 -07:00
Andrew Waterman
0380aed329 PUM -> SUM 2017-03-24 16:39:52 -07:00
Andrew Waterman
29414f3a23 Simplify interrupt-stack discipline
f2ed45b179
2017-03-24 16:39:52 -07:00
Andrew Waterman
723352c3e2 Mitigate some more PMP critical paths 2017-03-24 16:39:52 -07:00
Andrew Waterman
7484f27ed3 Don't gate exception-cause pipeline registers separately
They are too narrow to justify gating separately from the other pipeline
registers (and one of the clock gates was on the PMP critical path).
2017-03-24 16:39:52 -07:00
Andrew Waterman
487b8db5ef Address some PMP critical paths 2017-03-24 16:39:52 -07:00
Andrew Waterman
03fb334c4c Take mprv calculation off critical path 2017-03-24 16:39:52 -07:00
Andrew Waterman
f0796f0509 Pass correct access size information to PMP checker 2017-03-24 16:39:52 -07:00
Andrew Waterman
a6874c03f7 Remove DecoupledTLB 2017-03-24 16:39:52 -07:00
Andrew Waterman
78f9f6b9ef When SFENCE.VMA has rs2 != x0, don't flush global mappings 2017-03-24 16:39:52 -07:00
Andrew Waterman
1b950128e1 PTW should always use S-mode privilege
If an exception occurs while a page-table walk is coincidentally in
progress (e.g., an illegal instruction executes during data TLB refill),
then the processor might enter M-mode.  However, the PTW's accesses
should proceed without M privilege, to avoid bypassing PMPs.

Note, the same argument doesn't apply to the nonblocking cache's replay
queues, because those accesses have already been checked against the PMPs.
The cache correctly ignores access exceptions reported on replays,
provided no exceptions were reported on the initial access.
2017-03-24 16:39:52 -07:00
Andrew Waterman
aace526857 WIP on PMP 2017-03-24 16:39:52 -07:00
Andrew Waterman
b1b405404d Set PRV=M when entering debug mode
Debug mode mostly behaves like M-mode, so this approach avoids having
to check the debug bit in most permission checks.
2017-03-24 16:39:52 -07:00