Brandon Potter has uploaded this change for review. (
https://gem5-review.googlesource.com/c/public/gem5/+/22283 )
Change subject: cpu-o3: bugfix for c58cb8c9
......................................................................
cpu-o3: bugfix for c58cb8c9
The partial faults changeset broken some code related to checking
consistency model guarantees (found in X86 benchmarks).
This changeset adds some documentation to the code and obviates
the problem.
Change-Id: Ied9c6b0b1d237538efe4beb2f97ef76248ce2746
---
M src/cpu/o3/lsq_impl.hh
1 file changed, 25 insertions(+), 1 deletion(-)
diff --git a/src/cpu/o3/lsq_impl.hh b/src/cpu/o3/lsq_impl.hh
index c2d5e90..d7dc618 100644
--- a/src/cpu/o3/lsq_impl.hh
+++ b/src/cpu/o3/lsq_impl.hh
@@ -1152,13 +1152,37 @@
return ( (LSQRequest::_requests[0]->getPaddr() & blockMask) ==
blockAddr);
}
+/**
+ * Caches may probe into the load-store queue to enforce memory ordering
+ * guarantees. This method supports probes by providing a mechanism to
compare
+ * snoop messages with requests tracked by the load-store queue.
+ *
+ * Consistency models must enforce ordering constraints. TSO, for instance,
+ * must prevent memory reorderings except stores which are reordered after
+ * loads. The reordering restrictions negatively impact performance by
+ * cutting down on memory level parallelism. However, the core can regain
+ * performance by generating speculative loads. Speculative loads may issue
+ * without affecting correctness if precautions are taken to handle invalid
+ * memory orders. The load queue must squash under memory model violations.
+ * Memory model violations may occur when block ownership is granted to
+ * another core or the block cannot be accurately monitored by the load
queue.
+ */
template<class Impl>
bool
LSQ<Impl>::SplitDataRequest::isCacheBlockHit(Addr blockAddr, Addr
blockMask)
{
bool is_hit = false;
for (auto &r: _requests) {
- if ((r->getPaddr() & blockMask) == blockAddr) {
+ /**
+ * The load-store queue handles partial faults which complicates
this
+ * method. Physical addresses must be compared between requests and
+ * snoops. Some requests will not have a valid physical address,
since
+ * partial faults may have outstanding translations. Therefore, the
+ * existence of a valid request address must be checked before
+ * comparing block hits. We assume no pipeline squash is needed if a
+ * valid request address does not exist.
+ */
+ if (r->hasPaddr() && (r->getPaddr() & blockMask) == blockAddr) {
is_hit = true;
break;
}
--
To view, visit https://gem5-review.googlesource.com/c/public/gem5/+/22283
To unsubscribe, or for help writing mail filters, visit
https://gem5-review.googlesource.com/settings
Gerrit-Project: public/gem5
Gerrit-Branch: master
Gerrit-Change-Id: Ied9c6b0b1d237538efe4beb2f97ef76248ce2746
Gerrit-Change-Number: 22283
Gerrit-PatchSet: 1
Gerrit-Owner: Brandon Potter <brandon.pot...@amd.com>
Gerrit-MessageType: newchange
_______________________________________________
gem5-dev mailing list
gem5-dev@gem5.org
http://m5sim.org/mailman/listinfo/gem5-dev
