1
0
Commit Graph

110 Commits

Author SHA1 Message Date
Andrew Waterman
34d86ef665 Revert "Avoid data corruption under correctable tag error during flush (#1130)"
This reverts commit 44eb4d12b5.
2017-11-29 16:09:30 -08:00
Andrew Waterman
44eb4d12b5 Avoid data corruption under correctable tag error during flush (#1130)
This esoteric bug manifests if a tag-read error occurs when a FENCE.I is
executed, even if the error was correctable.  Subsequently, an attempt to
flush a dirty line may flush the wrong line's data.
2017-11-29 09:42:00 -08:00
Andrew Waterman
5155eb6059
Don't emit writeback state machine logic for scratchpad (#1127)
Firrtl can't DCE it because it would require analyzing the state machine.
2017-11-22 18:40:02 -06:00
Wesley W. Terpstra
7cfb69e2d5 Queue: silence some warnings 2017-11-14 15:09:09 -08:00
Andrew Waterman
4ebca73d59 Provide option to support AMOs only on I/O, not DTIM/D$ 2017-11-09 17:45:53 -08:00
Andrew Waterman
6176b348dc Invalidate TL error bit in D$ once progress is made 2017-11-07 00:52:18 -08:00
Andrew Waterman
f859da85ff Disable covers that don't apply to DTIM 2017-11-03 15:38:13 -07:00
Andrew Waterman
d6ede818ee DTIM doesn't accept grants 2017-11-03 15:37:48 -07:00
Henry Cook
b64609bfe8 Merge pull request #1039 from freechipsproject/tile-crossing-params
Improvements wrt connecting RocketTiles to SystemBus
2017-10-11 17:12:03 -07:00
Henry Cook
1867a5b226 rocket: only cache when AcquireT is possible 2017-10-10 18:06:58 -07:00
Andrew Waterman
b2bc46471b Conditionalize some covers that are sometimes impossible (#1043) 2017-10-10 17:14:33 -07:00
Henry Cook
9026646459 coreplex: first cut at using RocketCrossingParams 2017-10-10 12:02:04 -07:00
Andrew Waterman
f3825270c1 Add some covers for L1 memory system 2017-10-09 18:33:36 -07:00
Wesley W. Terpstra
bd045a3b95 tilelink: split Acquire into Acquire{Block,Perm} (#1030)
We had planned for a while to add an 'Overwrite' message which obtains
permissions without requiring retrieval of data. This is useful whenever
a master knows it will completely replace the contents of a cache block.

Instead of calling it Overwrite, we decided to split the Acquire type.

If you AcquirePerm, you MUST Release and ProbeAck with Data.
2017-10-05 12:49:49 -07:00
Henry Cook
05112b49a3 Merge branch 'master' into tl-error 2017-09-27 14:50:17 -07:00
Henry Cook
30c8c8c517 Revert "try to give seqmems clearer names"
This reverts commit 8db5bbbae0.

This attempt at clarification instead results in confusing generated verilog like:
`dcache_data_arrays_0 icache_data_arrays_0 (...);`
because of deduplication of identically dimensioned SRAMs...
2017-09-21 18:02:32 -07:00
Henry Cook
a887baa615 rocket: base trait for reporting ecc errors 2017-09-21 14:58:47 -07:00
Andrew Waterman
88c782cc70 Report D$ uncorrectable errors on C channel 2017-09-20 17:15:11 -07:00
Andrew Waterman
6bc20942b5 Don't cache TL error responses; report access exceptions 2017-09-20 17:01:08 -07:00
Andrew Waterman
db57e943f3 Report TL errors into D$ 2017-09-20 00:05:07 -07:00
Andrew Waterman
ed18acaae0 Report D$ errors 2017-09-20 00:05:07 -07:00
Henry Cook
57e8fe0a6b Merge pull request #1000 from freechipsproject/name-seqmems
try to give seqmems clearer names for use with external tools
2017-09-19 17:59:00 -07:00
Henry Cook
8db5bbbae0 try to give seqmems clearer names 2017-09-19 13:41:11 -07:00
Andrew Waterman
d93d7b9fa4 Only merge stores that aren't yet pending
This fixes a deadlock (and possibly memory corruption, though that is
unconfirmed).  The following sequence manifests it, assuming t0
is 32-byte aligned:

    sw t0, 0(t0)
    sw t0, 16(t0)
    lw t1, 4(t0)
    lw t2, 4(t0)
2017-09-17 15:01:07 -07:00
Andrew Waterman
1f606d924f Don't perform in-place correction if there was a recent store (#988)
Since the correction updates the entire word, the WAW hazard detection
logic is not sufficient to prevent overwriting a recent store.  So,
re-read the word after all pending stores have drained.
2017-09-08 16:26:54 -07:00
Andrew Waterman
53dfc5e9be Remove overzealous assertion (#987)
This assertion made sure the D$ controller was able to write the tag RAM
when a cache line was refilled.  However, it is benign if it fails to do
so: the metadata is invalid at this point, so the miss will simply happen
a second time.

This happens when resolving a tag ECC error during hit-under-miss.
2017-09-07 18:17:56 -07:00
Andrew Waterman
3cbc5262ec Don't permit new acquires until the release queue is drained
If the queue is not empty before a dirty miss, C could block D.
I haven't seen this in the wild, but it could happen because of
dirty probe responses backed up in the queue.
2017-08-13 13:18:45 -07:00
Andrew Waterman
0190724492 Actually use the C-channel acquire-before-release queue
oops...
2017-08-13 13:03:35 -07:00
Andrew Waterman
7387f2a93a Don't block D-channel when handling a probe
This is an acquire-before-release regression.
2017-08-12 16:13:24 -07:00
Andrew Waterman
604abd5b07 Only report ECC errors when the RAM was actually read 2017-08-12 15:28:03 -07:00
Andrew Waterman
18fb052fc9 DRY 2017-08-12 15:27:30 -07:00
Andrew Waterman
176110b6d3 Don't trigger ECC writebacks when a release is in flight 2017-08-12 15:23:57 -07:00
Andrew Waterman
809c7e8551 Don't merge stores that manifest WAW hazards
The following sequence would drop the first store when eccBytes=4:

    sb x0, 0(t0)
    nop
    sb x0, 4(t0)
    nop
    sb x0, 1(t0)

Because the first and second store are to different ECC granules, the
hazard check correctly allowed the second one to proceed, but the third
was merged with the second, even though it conflicted with the first.
So, don't allow the third to be merged with the second, since the second
stored to a different ECC granule.
2017-08-08 15:19:05 -07:00
Andrew Waterman
82e13443b2 Merge pull request #937 from freechipsproject/critical-paths
Perform tag error detectoin/correction in same cycle as RAM
2017-08-08 15:03:28 -07:00
Andrew Waterman
7935c61c19 Don't report to the DTIM that data is cacheable
Otherwise, it will attempt to perform AMOs where they're unsupported!
2017-08-08 11:55:04 -07:00
Andrew Waterman
74d309c18e Make I vs. D a static property of TLB, not an input pin
The microarchitecture doesn't really support unified TLBs, so don't fake it.
2017-08-08 11:54:47 -07:00
Andrew Waterman
e92981b0bd DRY 2017-08-08 11:46:38 -07:00
Andrew Waterman
62ccba304c Perform tag error detectoin/correction in same cycle as RAM
The tag RAMs tend to be fast, so take up some of the slack.
This makes s2_nack faster.
2017-08-08 10:21:30 -07:00
Andrew Waterman
402907990c Revert "Remove one gate from D$ ECC check"
This reverts commit 7d94074b05, which
works fine with optimistic behavioral RAMs but not real ones.
2017-08-07 17:33:20 -07:00
Andrew Waterman
7d94074b05 Remove one gate from D$ ECC check
The D$ corrects via writeback, so which word the error was in doesn't
matter, as the entire line is corrected.
2017-08-06 17:36:53 -07:00
Andrew Waterman
83875e3a0c Only flush D$ on FENCE.I if it won't always be probed on I$ miss 2017-08-05 14:22:40 -07:00
Andrew Waterman
991e16de92 Remove probe address mux from TLB response path 2017-08-05 12:57:38 -07:00
Andrew Waterman
2eb239d03f Add option to retime D$ way mux into subsequent pipeline stage 2017-08-01 23:59:20 -07:00
Andrew Waterman
2ecea2ef60 Don't use a pipe queue on D$ TL A-channel
This cuts an I$->D$ path.
2017-08-01 15:17:07 -07:00
Andrew Waterman
5681693ccc Fix a D$ ready-valid signaling regression
I broke this in 66d06460fa.
2017-07-31 18:05:14 -07:00
Henry Cook
11332c1226 dcache: break potential combinatorial loop by making pstore_drain_on_miss more conservative 2017-07-31 14:03:30 -07:00
Andrew Waterman
ac4339a8e7 Pass D$ backpressure to D-channel, rather than asserting 2017-07-29 11:48:36 -07:00
Andrew Waterman
edcd2c696c Avoid needless stall on E-channel back pressure 2017-07-29 11:47:58 -07:00
Andrew Waterman
2e8b02e780 Merge D$ store hits when ECC is enabled
This avoids pipeline flushes due to subword WAW hazards, as with
consecutive byte stores.
2017-07-28 12:56:36 -07:00
Andrew Waterman
838864870e Bypass TLB refill signal to halve L2 TLB hit time
The 4-cycle hit time is 1 cycle too long to avoid a second
pipeline replay, so it was effectively 9 cycles instead of 4.
2017-07-28 12:56:36 -07:00