AMD ATHLON K User Manual
Page 14
14
AMD Athlon™ Processor Model 4 Revision Guide
23614K—October 2003
Preliminary Information
20
A Speculative SMC Store Followed by an Actual SMC Store May Cause One-Time Stale
Execution
Products Affected. A4, A5, A6, A7, A9
Normal Specified Operation. Self-modifying code sequences should be correctly detected and handled in a
manner consistent with canonical results; stale code should not be executed.
Non-conformance. The following scenario can result in a one-time execution of stale instructions:
1. A speculative store instruction initiates a request (R) to modify a 64-byte cache line with address
A, which currently resides within the L1 instruction cache.
2.
The speculative store instruction is ultimately not executed because of a branch misprediction.
However, the store R is still in flight attempting to bring the line into the data cache in the
modified state.
3.
The instruction cache, which fetches instructions 16 bytes at a time, is redirected by the branch
into the cache line with address A and fetches a portion of the line into the instruction buffer.
4.
R then invalidates the instruction cache line with address A and brings the line into the L1 data
cache, marking it as modified. However, the instruction buffer, which also contains some bytes
from address A, is not invalidated.
5.
The instruction fetch mechanism attempts to read the next 16-byte chunk of code and must issue a
request to bring the 64-byte line back into the instruction cache.
6.
This instruction cache request for address A hits on the modified line now in the L1 cache, and
evicts it from the data cache to the L2.
7.
A second store instruction (S) from the instruction buffer is issued into the execution units. S is a
self-modifying code reference to another instruction that currently exists in the 64-byte cache
block at address A and is also in the instruction buffer.
8.
The execution of S detects that an instruction request to fetch address A is in flight. However, the
store request is given priority. Since it now hits in the L2 and the L2 state is modified, it assumes
that the line cannot be in the instruction cache or the instruction buffer.
Potential Effect on System. The processor will execute stale code instructions.
Suggested Workaround. None. This failure has only been observed in internally generated synthetic code.
Resolution Status. No fix planned.