Compaq EV67 User Manual
D–14
PALcode Restrictions and Guidelines
Alpha 21264/EV67 Hardware Reference Manual
Restriction 24: HW_RET/STALL After HW_MTPR IC_FLUSH, IC_FLUSH_ASM,
xxx
HW_ST/C -> R0
Bxx R0, try_again
STQ ; Force next ST/C to fail if no preceding LDxL
HW_RET
D.20 Restriction 24: HW_RET/STALL After HW_MTPR IC_FLUSH,
IC_FLUSH_ASM, CLEAR_MAP
There must be a HW_RET/STALL after a HW_MTPR IC_FLUSH, IC_FLUSH_ASM, or
CLEAR_MAP. The Icache flush associated with these instructions will not occur until
the HW_RET/STALL occurs and all outstanding Istream fetches have been completed.
CLEAR_MAP. The Icache flush associated with these instructions will not occur until
the HW_RET/STALL occurs and all outstanding Istream fetches have been completed.
Also, there must be a guarantee that the HW_MTPR IC_FLUSH or HW_MTPR
IC_FLUSH_ASM will not be retired simultaneously with the HW_RET/STALL. This
can be ensured by inserting a conditional branch between the two (BNE R31, 0 cannot
be mispredicted in PALmode), or by ensuring at least 10 instructions between the
MTPR instruction and the HW_RET/STALL containing at least one instruction in each
quad aligned group with a valid destination. Finally, the HW_RET/STALL that is used
for CLEAR_MAP cannot trigger a cache flush. That is, if both a CLEAR_MAP and
IC_FLUSH are desired, there must be two HW_RET/STALLs, one following each
HW_MTPR.
IC_FLUSH_ASM will not be retired simultaneously with the HW_RET/STALL. This
can be ensured by inserting a conditional branch between the two (BNE R31, 0 cannot
be mispredicted in PALmode), or by ensuring at least 10 instructions between the
MTPR instruction and the HW_RET/STALL containing at least one instruction in each
quad aligned group with a valid destination. Finally, the HW_RET/STALL that is used
for CLEAR_MAP cannot trigger a cache flush. That is, if both a CLEAR_MAP and
IC_FLUSH are desired, there must be two HW_RET/STALLs, one following each
HW_MTPR.
D.21 Restriction 25: HW_MTPR ITB_IA After Reset
An HW_MTPR ITB_IA is required in the reset PALcode to initialize the ITB. It is also
required that PALcode not be exited, even via a mispredicted path until this
HW_MTPR ITB_IA has been retired. PALmode can change temporarily after fetching
a HW_RET, regardless of the STALL qualifier, down a mispredicted path leading to use
of the ITB before it is actually initialized.
required that PALcode not be exited, even via a mispredicted path until this
HW_MTPR ITB_IA has been retired. PALmode can change temporarily after fetching
a HW_RET, regardless of the STALL qualifier, down a mispredicted path leading to use
of the ITB before it is actually initialized.
Unexpected instruction fetch and execution can occur following misprediction of any
memory format control instruction (JMP, JSR, RET, JSR_CO, or HW_JMP, HW_JSR,
HW_RET, HW_JSR_CO regardless of the STALL qualifier), or after any mispredicted
conditional branch instruction. If the unexpected instruction flow contains a HW_RET
instruction, PALmode may be exited prematurely.
memory format control instruction (JMP, JSR, RET, JSR_CO, or HW_JMP, HW_JSR,
HW_RET, HW_JSR_CO regardless of the STALL qualifier), or after any mispredicted
conditional branch instruction. If the unexpected instruction flow contains a HW_RET
instruction, PALmode may be exited prematurely.
One way to ensure that PALmode is not exited is to place the HW_MTPR ITB_IA at
least 80 instructions before any possible HW_RET instruction can be encountered via
any fetch path. Since memory format control instructions can mispredict to any cache
location, they should also be avoided within these 80 instructions.
least 80 instructions before any possible HW_RET instruction can be encountered via
any fetch path. Since memory format control instructions can mispredict to any cache
location, they should also be avoided within these 80 instructions.
D.22 Guideline 26: Conditional Branches in PALcode
To avoid pollution of the branch predictors and improve overall branch prediction accu-
racy, conditional branch instructions in PALcode will be predicted to not be taken. The
only exception to this rule are conditional branches within the first cache fetch (up to
four instructions) of all PALcode flows except CALL_PAL flows. Conditional branches
should be avoided in this window.
racy, conditional branch instructions in PALcode will be predicted to not be taken. The
only exception to this rule are conditional branches within the first cache fetch (up to
four instructions) of all PALcode flows except CALL_PAL flows. Conditional branches
should be avoided in this window.