Intel ITANIUM ARCHITECTURE - SOFTWARE DEVELOPERS MANUAL VOLUME 1 REV 2.3 Manual page 353

5.5.5 Deferral of Speculative Load Faults
Speculative and speculative advanced loads can defer fault handling by suppressing the
speculative memory reference, and by setting the deferred exception indicator (NaT bit
or NaTVal) of the load target register. Other effects of the instruction (such as post
increment) are performed. Additionally, software can suppress the memory reference of
speculative and speculative advanced loads independent of any exception.
Deferral is the process of generating a deferred exception indicator and not performing
the exception processing at the time of its detection (and potentially never at all). Once
a deferred exception indicator is generated, it will propagate through all uses until the
speculation is checked by using either a chk.s instruction, a chk.a instruction (for
speculative advanced loads), or a non-speculative use. This causes the appropriate
action to be invoked to deal with the exception.
Three different programming models are supported: no-recovery, recovery and
always-defer. In the no-recovery model, only fatal exceptional conditions are deferred
these are conditions which cannot be resolved without either involving the program's
–
exception-handling code or terminating the program. In the recovery model,
performance may be increased by deferring additional exceptional conditions. The
recovery model is used only if the program provides additional "recovery" code to
re-execute failed speculative computations. When a speculative load is executed with
PSR.ic equal to 1, and ITLB.ed equal to 0, the no-recovery model is in effect. When
PSR.ic is 1 and ITLB.ed is 1, the recovery model is in effect. The always-defer model
is supported for use in system code which has PSR.ic equal to 0. In this model, all
exceptional conditions which can be deferred are deferred. This permits speculation in
environments where faulting would be unrecoverable.
In addition to the deferral of exceptional conditions, speculative loads may be deferred
automatically by hardware based on implementation-dependent criteria, such as the
detection of a cache miss. Such deferral is referred to as spontaneous deferral, and
is done in order to increase performance. Spontaneous deferral is allowed only in the
recovery model.
Table 5-2.
PSR.ic
0
1
1
1
Speculative load exceptions are categorized into three groups:
• Ones which always raise a fault
• Ones which always defer
• Ones which always raise a fault in the no-recovery model, but can defer based on
the speculative deferral control bits in the DCR control register, in the recovery
model.
Volume 2, Part 1: Interruptions
Programming Models
PSR.it ITLB.ed
x x Always defer
0 x No recovery
1 0 No recovery
1 1 Recovery
Model DCR-based Deferral
No
No
No
Yes
Spontaneous Deferral
No
No
No
Yes
2:105