EPXA4 [ALTERA]
Excalibur Devices; 神剑设备
| 型号： | EPXA4 |
| 厂家： | 
ALTERA CORPORATION |
| 描述： | Excalibur Devices |
| 文件： | 总34页 (文件大小：453K) |
| 中文： | 中文翻译 | 下载： | 下载PDF数据表文档文件 |
Excalibur EPXA4 Devices
November 2002, ver. 1.2
Errata Sheet
This errata sheet provides updated information about the Excalibur™
EPXA4, revision A (see Figure 1) Devices.
Figure 1. Identify Revision A Devices
Revision Number
The errata fall into two categories:
■
■
Known errata for the EPXA4 device—detailed in this document
Known errata for the ARM922T processor provided by ARM Ltd.—
detailed in Appendix A of this document
The following sections of the device are covered by errata information:
■
■
■
■
■
■
■
■
Expansion bus interface (EBI)
Dual-port SRAM (DPRAM)
AHB bridges
UART
SDRAM
Embedded trace module version 2a
Configuration
Debug module
Contact Altera® for the latest information.
Altera Corporation
1
ES-EPXA4-1.2
Excalibur EPXA4 Devices Errata Sheet
This section provides further information about errata in the EBI.
EBI
1.1 Locked 16-Beat Incrementing Bursts
A locked INCR16 transfer can cause the EBI to read from a peripheral
twice. This can cause erroneous behavior if the peripheral contains read-
sensitive registers.
Work Around
Do not use burst transactions to access read-sensitive peripherals
connected to the EBI.
1.2 EBI Acknowledge Signal
The EBI acknowledge signal, EBI_ACK, functionality results in incorrect
EBI asynchronous mode operation.
Work Around
Do not use the EBI_ACKsignal when interfacing to external memory
devices. Instead, use the programmable wait states, via the EBI_BLOCKn
register, for the individual EBI blocks. Also ensure that the SAbit for each
EBI block is cleared. This effectively puts the EBI in synchronous mode.
EBI synchronous mode can be used to interface with synchronous or
asynchronous memories provided that the appropriate number of wait
states is implemented.
This section provides further information about errata in the DPRAM.
DPRAM
2.1 Port A Clocks Are Not Pre-Balanced
If the DPRAM is configured in either a deep (×8) or a wide (×32) mode, the
dp0_2_portaclkand dp1_3_portaclkclock signals are not pre-
balanced. The Quartus® II fitter is unable to balance locally routed clocks
and generates an error.
Work Around
Promote dp0_2_portaclkor the dp1_3_portaclksignal to a global
signal and recompile the design.
2
Altera Corporation
Excalibur EPXA4 Devices Errata Sheet
2.2 Locking Mechanism is Non-Functional
Using the DPRAM locking mechanism for simultaneous accesses from the
stripe and PLD results in incorrect memory accesses. Simultaneous
accesses to different addresses are not affected.
Work Around
Do not use the DPRAM locking feature. Instead, use a handshake
mechanism such as a semaphore to control PLD and stripe accesses to
common DPRAM addresses.
This section provides further information about errata in the AHB
bridges.
AHB Bridges
3.1 Back-to-back Transactions through the PLD-to-Stripe Bridge Can
Cause Lock-up
If the non-sequential address phase of a new transaction occurs in the
same clock cycle as the final data phase of the previous transaction, the
slave interface of the PLD-to-stripe bridge issues erroneous wait states
causing masters to the PLD-to-stripe bridge in the PLD to lock up.
Incorrect transactions can occur on AHB2.
Work Around
Insert an IDLE address phase between all transactions.
3.2 Corrupted State of the Stripe-to-PLD Bridge after PLD
Reconfiguration under Processor Control
If the MASTER_HCLKsignal is active during PLD reconfiguration under
processor control, the stripe-to-PLD bridge can become corrupted before
the device enters user mode. The bridge cannot respond to AHB2
transactions, which causes the AHB2 bus to lock up. Gating the source of
MASTER_HCLKwith INIT_DONEdoes not prevent the potential lock up.
Work Arounds
To avoid corrupting the bridge during PLD reconfiguration, the clock
driving MASTER_HCLKmust be inactive before the PLD enters user mode.
This can be achieved in either of the following ways:
■
Software work around—compile the design using Quartus® II version
2.1 or higher. These versions of Quartus II route the stripe-to-PLD
bridge signals specifically to avoid the bridge lock up. The routing
Altera Corporation
3
Excalibur EPXA4 Devices Errata Sheet
utilizes three extra logic elements in the device and adds minimal
corresponding delay to the bridge timing. If logic element usage or
bridge timing is critical to your design, you can disable the automatic
Quartus II routing option by adding the following parameter to the
defparamsection of the stripe instantiation file generated by the
MegaWizard®:
lpm_instance.xa_configuration_fix = “FALSE“
If this routing option is disabled, you must use the hardware work
around described below to ensure correct functionality.
1
If you use this recommended software workaround, ensure
that the Remove Redundant Logic Cells option is not
selected for your project.
1
If the Perform WYSIWYG Primitive Resynthesis option is
selected for your project, you may receive warnings that the
stripe signals were not routed correctly. To eliminate the
warnings, re-run the MegaWizard in Quartus II version 2.2.
This creates an additional settings file (alt_exc_stripe.esf) to
ensure that the required logic elements are implemented.
■
Hardware work around—ensure that the external clock that drives
MASTER_HCLKis inactive before the PLD is put into reconfiguration
mode, and is enabled again only after the PLD enters user mode.
INIT_DONEcan be used only to activate the external clock. It should
not be used to disable the clock, for the following reason: when
reconfiguration begins, INIT_DONEdoes not drop to low in time to
prevent the error. You must use another method to disable the
external clock, such as using an external device to control the clock
input.
If a PLL in the PLD portion of the device is used to drive
MASTER_HCLK, you must drive the PLL CLKLK_ENAsignal with
INIT_DONEso that the PLL output does not drive MASTER_HCLK
before the device is in user mode. See AN 115: Using the ClockLock &
ClockBoost PLL Features in APEX Devices for information on enabling
the PLD PLLs with INIT_DONE.
4
Altera Corporation
Excalibur EPXA4 Devices Errata Sheet
This section provides further information about errata in the UART.
4.1 UART State Undefined Following Reset
UART
When the device comes out of reset, the UART may sometimes generate a
modem interrupt and behave as though a character has been received.
Work Around
To avoid generating a modem interrupt, clear all interrupts and flush the
receive and transmit FIFO buffers before interrupts from the UART are
enabled.
This section provides further information about errata in the SDRAM.
SDRAM
5.1 32-bit DDR SDRAM Memories are Not Supported
The SDRAM controller in the ARM-based family (including the EPXA4)
does not support interfacing with 32-bit DDR SDRAM devices that use the
A8 address line for the command sequence. Also 32-bit DDR memories
have one DQSpin; ARM-based devices require four (one per byte).
Memories that use the A10 address line for the command sequence and
have one DQSpin per byte are supported.
5.2 Improper DDR SDRAM Data Accesses for Certain Clock Ratios
Incorrect data may result from DDR SDRAM accesses if the AHB1 or
AHB2 master clock is greater than 4 times faster than the SDRAM clock,
SD_CLK.
Work Around
Ensure that the AHB1/2 clock frequency is less than or equal to 4 times the
SDRAM clock frequency.
EPXA4 devices include the ARM ETM9 version 2a. Please see the
ETM9_Rev_2a_Errata.doc for errata on this version of the ETM9. This
document is available on the ARM Limited website.
Embedded
Trace Module
Version 2a
Version 2a has a different configuration code register value compared to
version 1, which is used in the EPXA10 devices. The ARM Trace Tools
version 1.1 does not support ETM9 version 2a. Support will be included
in a new version of the trace tools planned for Q2 2002.
Altera Corporation
5
Excalibur EPXA4 Devices Errata Sheet
This section provides further information about errata in configuration.
Configuration
7.1 JTAG Configuration Error
JTAG configuration of the EPXA4 does not complete successfully if there
are programmed EPC configuration devices in the JTAG chain and the
EPXA4 is in boot-from-serial mode. The Quartus II Programmer reports
an error stating that the CONF_DONEsignal did not go high.
Work Around
Erase the EPC devices that are in the chain. This allows successful
configuration in boot-from-serial-mode. Or assert the Boot_Flashpin,
which sets the configuration mode to boot-from-flash. In this mode, you
can successfully configure the device via JTAG.
This section provides further information about errata in the debug
module.
Debug Module
8.1 DEBUG_IE_BRKPT and DEBUG_DE_WATCHPT Cause Incorrect
Triggers
The DEBUG_IE_BRKPTand DEBUG_DE_WATCHPTlines do not function. If
asserted, they cause breakpoints or watchpoints to trigger on the wrong
instruction.
Work Around
Do not connect the DEBUG_IE_BRKPTand DEBUG_DE_WATCHPTlines in
your design. The Quartus II software ties them low as appropriate.
6
Altera Corporation
Appendix A
Errata from ARM Ltd.
© 1999-2001 ARM Limited. All rights reserved
This appendix reproduces information supplied by ARM Ltd. on known
errata in the ARM922T™processor and support features. The errata have
little or no impact on the use of the Excalibur device, but could be useful,
on rare occasions, in understanding its interaction with third-party
debugging tools. Many errata detail the work arounds which tool vendors
incorporate into their products, while others inform of changes in the
specification and so are included for completeness.
The information contained herein is the property of ARM Ltd. and is
supplied without liability for errors or omissions. No part may be
reproduced or used except as authorised by contract or other written
permission. The copyright and the foregoing restriction on reproduction
and use extend to all media in which this information may be embodied.
Table 1 describes the errata categories defined by ARM Ltd. and used
throughout this appendix, except for A2.1 “Invalid data trace following
FIFO overflow—Category 1” on page 9, which refers to a previous
definition of category 1 defined by ARM Ltd.
Table 1. Errata Categories
Category
Description
1
2
3
Features which are impossible to work around and severely restrict the use of the device in all or the
majority of applications rendering the device unusable.
Features which contravene the specified behaviour and may limit or severely impair the intended use
of specified features but does not render the device unusable in all or the majority of applications.
Features that were not the originally intended behaviour but should not cause any problems in
applications.
Errata for the ARM-based embedded processor are grouped by module:
■
■
■
A.1—the ARM920T AHB wrapper
A.2—the ETM9 trace module
A.3—the ARM922T processor core
Altera Corporation
7
Errata Sheet - Appendix A
A.1
Errata for the ARM920T AHB Wrapper Module
This section documents errata in the ARM920T AHB wrapper module,
which is used in the Excalibur device.
A.1.1
Linefill Counter—Category 2
The burst counter for a cache linefill is incorrectly pre-loaded if the linefill
follows a waited access. This results in 9 reads being performed on the
AHB bus rather than the required 8. The ARM920T receives the correct
data: the only effects of the extra read are that an extra cycle is taken up on
the bus, and any side-effects are due to the addressed area being read-
sensitive (e.g. a FIFO).
Work Around
The software workaround is to ensure that only address areas which are
not read-sensitive are cached.
A.1.2
Error Response—Category 2
There is a bug in the ERROR response functionality in the wrapper. An
ERROR response should only be accepted by the ARM920T if it is in
response to a non-cacheable or non-bufferable access. The AHB wrapper
only adds wait states to non-bufferable writes when ERROR support is
added, since buffered writes do not need to be held up in order to
propagate the respons. However, the ERROR response from the AHB is
still propagated to the ASB in all cases. Thus a C/B access to the AHB
which results in an ERROR response, followed by a NC/NB* access,
means that the ERROR response appears on the ARM920T ASB interface
as the response to the NC/NB access and so is accepted.
Work Around
The software workaround is to never access address regions which are
capable of generating error responses with C/B accesses.
*NC Non-cacheable
NB Non-bufferable
8
Altera Corporation
Errata Sheet - Appendix A
A.2
Errata for the ETM9 Trace Module
See Table 1 on page 7 and associated text for an explanation of
categorisation.
A.2.1
Invalid data trace following FIFO overflow—Category 1
Note: ARM has updated the errata classifications. This is a category 1
erratum under the previous definition, which was:
‘Features which it is impossible, or very hard, to work around and are
likely to affect use of this device.’
Description
The Embedded Trace Macrocell (ETM) contains a small FIFO which under
some circumstances can overflow, leading to the loss of trace. This is
normal behavior.
However, if the FIFO overflows during an extended wait-state period
while attempting to capture data trace from a block data transfer
instruction, then the data trace can become invalid.
When trace is resumed:
■
■
Data values traced may be invalid until the next indirect branch.
Data addresses traced may be invalid until the next trace gap, or the
next periodic synchronization point.
■
Instruction trace is unaffected.
An example of how exactly the data trace becomes invalid is shown in
Table 2 on page 10. Note that, for simplicity, data address tracing is
ignored.
Altera Corporation
9
Errata Sheet - Appendix A
Table 2. Example of invalid data trace following overflow during wait-state period
Address
Instruction
Data trace
Instruction
Data transfer
Notes
entering flow reported reported by ETM
a
ETM FIFO
by ETM
1004
LDMIA r6, {r0 – r4} [r6]
[r6+1]
1004 with data
r0 <= [r6]
r1 <= [r6+1]
r2 <= [r6+2]
ETM begins to log the data transferred
from the contents of r6 to each registers
in the list in turn, with r6 post-
[r6+2]
incremented each time.
Wait-state begins, stalling data
transfers.
Overflow occurs (due to data trace
continuing to enter the FIFO from the
ETM’s internal pipeline) causing tracing
to be suspended
FIFO drains and overflow clears
Wait-state ends and data transfers
resume
[r6+3]
[r6+4]
Overflow
1004 with data
r0 <= [r6+3]
The instruction is reported once more by
ETM as trace restarts, data trace
resumes and now incorrect data values
are reported.
The required operation was that no data
should enter the FIFO at this point, as
there is no way to indicate which
transfers were traced.
r1 <= [r6+4]
r2 <= [r8]b
r3 <= invalid datab
r4 <= invalid datab
1008
LDR r7, [r8]
[r8]
1008 with data
r7 <= invalid datab
The ETM now also reports an incorrect
data value for register r7
The required operation was that the next
operation after overflow would receive
correct data trace. So, in this example:
r7 <= [r8]
Notes:
(a) Corresponds to data transferred by the core
(b) Data from a future instruction
Note that, in this example, all the data transfers were reported either
before or after the overflow. However, this may not necessarily be the
case, and some transfers before the overflow, and the instruction itself,
may not be reported,
10
Altera Corporation
Errata Sheet - Appendix A
Conditions
Extended wait-state periods can be caused by cache misses in cached
systems, or the use of a slow memory system. Consequently, the problem
is unlikely to occur in uncached systems or in systems where the speed of
the memory is close to that of the processor. Since the clock speed on the
ARM920T and ARM922T processors is slowed to that of the memory
when a cache miss occurs, memory speed is not an issue on these
processors, and consequently this erratum is less likely to occur on
systems with these devices.
The block data transfer instruction must be executing when the overflow
occurs. The instructions which fall into this category are as follows:
■
ARM instructions:
LDM, STM, SWP, SWPB, LDC, STC, LDRD, STRD, MCRR, MRRC.
■
Thumb instructions:
POP, PUSH, LDMIA, STMIA.
The wait-state period must begin before the overflow occurs. It is possible
for an overflow to occur after the wait-state has begun due to trace
continuing to enter the FIFO from the ETM pipeline. The wait-state period
must continue until after the ETM has recovered from the overflow,
having drained its FIFO of all pending trace.
The problem is more likely to occur with a small FIFO than a large FIFO
for two reasons:
■
■
A smaller FIFO is more likely to overflow.
Once an overflow has occurred, a smaller FIFO takes less time to
drain. Consequently, the length of the extended wait-state period
required for the problem to occur is reduced.
As a result, the problem is most likely to occur with the small ETM
configuration, and least likely with the large ETM configuration.
The minimum number of cycles required to drain the FIFO, and therefore
the theoretical minimum length of the wait-state period, is shown in
Table 3 on page 12.
Altera Corporation
11
Errata Sheet - Appendix A
Table 3. Minimum length of the wait-state period for the problem to occur
ETM Configuration FIFO size
Minimum FIFO
depth at overflow
Port size
Cycles to
drain FIFO
Minimum number of
consecutive wait states
Large
45 bytes
18 bytes
9 bytes
37 bytes
10 bytes
1 byte
16-bit
19
20
8-bit
4-bit
16-bit
8-bit
4-bit
8-bit
4-bit
37
74
5
38
75
6
Medium
Small
10
20
1
11
21
2
2
3
While the minimum number of consecutive wait states required for this
erratum to occur is increased by reducing the port size, a wider port size
is still recommended as a more narrow port size will cause overflows to
occur more frequently.
These cycle calculations are based on having eight free bytes in the FIFO
when nine bytes are generated in a cycle. This occurs for the first data
transfer after trace has been turned on, and requires both data value and
data address tracing to be enabled. Most overflows will require a longer
wait-state than this.
The ETM FIFOFULLsignal attempts to preemptively insert processor
wait states to prevent the FIFO from overflowing. Consequently, this
means that this erratum will occur more often when FIFOFULLis
enabled. In this case this erratum interacts with a separate category 3
erratum, “FIFOFULL LOW for one cycle during overflow—Category 3”,
documented on page 25. This causes FIFOFULLto be low for the first
cycle of overflow, during which the ETM will not cause the processor to
stall. As a result, while enabling FIFOFULLcan increase the chance of this
erratum occurring, it does not on its own cause the erratum to occur,
because a wait-state must occur at the same time that FIFOFULLgoes low
for this one cycle.
12
Altera Corporation
Errata Sheet - Appendix A
Implications
While instruction trace remains unaffected, the user is unable to ascertain
whether the data trace is correct.
The erratum must be considered when the following occur together:
■
■
An overflow is reported.
The first instruction traced following the overflow is a block data
transfer instruction.
The first instruction traced following the overflow includes data
trace.
Either:
■
■
–
The address of the instruction traced both before and after
overflow is the same
–
The address of the instruction traced after overflow is the next
instruction address that would have been expected had the
overflow not occurred.
Examples of this case are:
–
The instructions before and after overflow are both LDMIA
instructions at address 1000.
–
The instruction before overflow is a branch at address 1010 that
failed its condition codes, and the instruction after overflow is an
STMDB at address 1014.
–
The instruction before overflow is an LDR to the pc at address
1020 which caused a branch to address 1040, and the instruction
after overflow is an LDCL at address 1040. This case may be
difficult to detect by the user, although it is possible for tools to
be able to detect this because the target address of a branch is
always traced, even if the next instruction is not.
■
The amount of data trace lost depends on the position of the next
indirect branch, and the data tracing mode selected. An indirect
branch is any branch which is not a B, BL or BLX instruction, such as
an LDR to the program counter.
Altera Corporation
13
Errata Sheet - Appendix A
One of the following apply:
The first instruction traced following the overflow is an indirect branch.
The data (addresses and values) traced with this instruction must be
treated as invalid, but data traced for other instructions will be
unaffected. It is always possible for invalid trace to be detected by the
tools when the first instruction is an indirect branch, except in the
case of an SWPB instruction when a 16-bit trace port is in use. If you
are using tools which detect this, such as ARM’s Trace Debug Tools,
the data trace can be treated as valid if the first instruction following
the overflow is an indirect branch and no error has been detected.
The next instruction to have data traced does not occur until after the next
indirect branch.
The data traced with the first instruction must be treated as invalid,
but data traced for future instructions is unaffected. It is not always
possible for the tools to detect the error.
Other instructions have data traced before the next indirect branch, and only data
addresses are being traced.
The data addresses must be treated as invalid for the first instruction,
but will be valid for the instructions which follow.
Another instruction has data traced before the next indirect branch, and either
only data values are being traced, or the first instruction following the overflow
is an MRRC or MCRR.
The data values traced with all the instructions up to the next indirect
branch must be treated as invalid. Data values traced after the next
indirect branch are unaffected by this erratum and will be valid
Another instruction has data traced before the next indirect branch, and data
values and addresses are both being traced.
The data values and addresses traced with all instructions before the
next indirect branch must be treated as invalid. Data addresses traced
for instructions after the next indirect branch, but before the next full
32-bit data address, must also be treated as invalid. The first data
address output after each trace gap is a full 32-bit address, after which
a full data address is forced if one has not been output for 1024 cycles.
14
Altera Corporation
Errata Sheet - Appendix A
While most tools do not report when a full data address has been
output, if a data address differs from the previous data address traced
in bits [31:28], then a full data address would have been output,
allowing full data addresses that have not been periodically forced to
be detected by the user. When a full data address cannot be detected
in this way the user must either find the first data address that is at
least 1024 cycles after the first instruction traced following the
overflow, or treat all data addresses as invalid until the next gap in
the trace.
Workaround
This workaround is for tool vendors only and must be read in conjunction
with the Embedded Trace Macrocell Specification (ARM IHI 0014).
Development tool vendors can implement the above checks automatically
in the trace tools, so that all data given to the user is guaranteed to be
correct. In particular:
■
■
■
The target address of branches is always known, so the comparison
of the addresses before and after overflow can be accurate.
The tools must detect when invalid trace is caused by this erratum,
and not cause instruction trace to be lost as a result.
The tools can detect when the first instruction after overflow is an
indirect branch (traced with a PIPESTAT of BD) and not an SWPB
with a 16-bit trace port, and not treat the data trace as invalid unless
the branch address is earlier than expected. An SWPB cannot be
reliably detected in this way because of the 1-byte gap that is
sometimes left on a 16-bit trace port to align instruction addresses.
Where data addresses are invalid until the next synchronization
point, the tools should treat data addresses as valid following the next
5-byte data address output instead.
■
There is no known way to prevent the erratum from occurring, other than
increasing the amount of filtering performed, particularly by ViewData,
to reduce the chance of an overflow occurring.
Implications of workaround
In some circumstances substantial data trace, and in particular data
addresses, might be lost. Data might be treated as invalid when it is not,
particularly in systems where this erratum never or rarely occurs. As a
result, it is recommended that if the workaround is implemented it is
possible to turn it off.
Altera Corporation
15
Errata Sheet - Appendix A
A.2.2
Execution status unknown prior to an interrupt or prefetch
abort—Category 2
Description
The trace port protocol allows for each instruction traced to be reported
with a corresponding branch address to indicate the address of the next
instruction. If this branch address is not output then the trace tools must
calculate the address of the next instruction. If a branch address is output
then the PIPESTAT pins of the trace port will indicate Branch Executed (BE)
or Branch with Data (BD).
When an interrupt or prefetch abort occurs, the instruction which follows
the last instruction to be executed should be traced (even though it did not
execute) with a branch address to the relevant exception vector. In the case
of interrupts this is often referred to as the interrupted instruction. The
trace tools must recognize that the branch was to an interrupt or prefetch
abort exception vector, and correctly treat that instruction as having not
executed. In this way, the trace can reliably indicate if the last instruction
executed failed its condition codes.
Sometimes the ETM does not trace this extra instruction but instead traces
the last instruction executed as having branched to the exception vector.
16
Altera Corporation
Errata Sheet - Appendix A
Table 4 shows an example of incorrect trace caused by this erratum.
Table 5 shows an example where this erratum does not occur.
Table 4. Example of incorrect trace behavior on interrupt
Actual instructions executed
Correct behavior
Trace generated
Actual behavior
Address
Instruction
Trace generated
Decompressed
trace
Trace starts, address
Trace starts, address
0x00001000
0x00001000
0x00001000
0x00001004
ADD r1, r1, 1
B 0x00001020
Instruction executed
Instruction executed
0x00001000
IRQ!
Instruction executed, branch to Instruction executed, branch to
0x00001020
0x00000018
(0x00001020)
(Not executed due to Instruction executed, branch to
interrupt)
0x00000018a
0x00000018
IRQ hander
IRQ handler traced
IRQ handler traced
0x00000018
.
.
.
.
.
.
.
.
.
.
.
.
SUBS pc, r14, #4 Instruction executed, branch to Instruction executed, branch to
0x00001020
0x00001020
0x00001020
…
…
…
0x00001020
Note:
(a) Extra instruction traced to indicate interrupt
Table 5. Example of correct trace behavior on interrupt
Actual instructions executed
Actual behavior
Decompressed trace
Address
Instruction
Trace generated
Trace starts, address
0x00002000
0x00002000
0x00002004
ADD r1, r1, 1
Instruction executed
0x00002000
BNE 0x00002020
Failed its condition
codes
Instruction executed but failed 0x00002004
its condition codes Failed its condition codes
(0x00002020)
(Not executed due to Instruction executed, branch to IRQ!
interrupt)
0x00000018a
0x00000018
IRQ hander
IRQ handler traced
0x00000018
.
.
.
.
.
.
.
.
.
.
SUBS pc, r14, #4 Instruction executed, branch to
0x00002020
0x00002020
…
…
0x00002020
Note:
(a) Extra instruction traced to indicate interrupt
Altera Corporation
17
Errata Sheet - Appendix A
Conditions
The conditions under which this erratum occurs are not easily predicted.
It never occurs if the last executed instruction failed its condition codes.
Implications
The last instruction executed before an interrupt or prefetch abort might
be missing from the trace. Some tools always report the extra instruction
traced as having executed, in which case the user will instead see an extra
instruction reported before an interrupt or prefetch abort when this
erratum does not occur.
If tracing continues until after the exception handler returns, the user can
use the address of the first instruction executed upon returning from the
exception to determine the actual behavior.
Workaround for tools vendors
Since the execution status is unknown, development tools cannot reliably
report the last instruction executed before an abort or an interrupt.
The development tools must display the last instruction traced with its
execution status shown as ‘unknown’, rather than suppress it as defined
in the trace port protocol.
Implications of workaround
It is not possible to determine which instruction was executed
immediately before an interrupt or prefetch abort, without referring to the
last instruction traced before the interrupt.
18
Altera Corporation
Errata Sheet - Appendix A
A.2.3
Instruction displayed before and after overflow—Category 2
Description
When the FIFO overflows, the last instruction to be traced immediately
before overflow may be repeated on recovery from overflow.
Conditions
The instruction must be executing when the overflow occurs and remain
so until trace is re-enabled following overflow. This generally only arises
if a large number of wait states occur. Consequently it should only occur
in cached systems during a cache miss, or when using a very slow
memory system.
Since the ETM FIFOFULLsignal attempts to preemptively insert
processor wait states to prevent the FIFO from overflowing, having
FIFOFULL enabled can also increase the occurrence of this problem.
Implications
It can appear that an instruction is executed twice when it was only
executed once.
Workaround for tools vendors
The development tools should be modified to discard the instruction
before the overflow, only displaying the instruction after overflow when
the addresses of the instructions before and after overflow match.
Occasionally this would cause an instruction to be discarded
unnecessarily, but a much larger number of instructions would already
have been discarded due to the overflow.
It is important that the instruction before overflow is the one to be
discarded, and not the one following overflow recovery, in case the
erratum is falsely detected. If trace is turned off during overflow
(TraceEnable goes LOW), then enabled again (TraceEnable goes HIGH)
some time after overflow recovery, the first instruction to be traced
following overflow may be significant and must be traced.
Implications of workaround
An extra instruction will occasionally be lost during an overflow
occurring in a loop. This will not be detectable by the user.
Altera Corporation
19
Errata Sheet - Appendix A
A.2.4
Address range cannot include 0xFFFFFFFF—Category 2
Description
There is no way to define a range which includes 0xFFFFFFFF. Ranges
are defined to be exclusive of the upper address, so a range with an upper
address of 0xFFFFFFFFonly includes addresses up to 0xFFFFFFFE.
Conditions
All.
Implications
The ability to monitor accesses to this address is seriously reduced.
Workaround
To monitor accesses to 0xFFFFFFFF, the upper address on the range
must be set to 0xFFFFFFFF, with the size mask set to clear bits 1:0 (bits
4:3 of the access type register must be set to b11). This causes the upper
bound of the address range never to match, and so the range has no upper
limit and accesses to 0xFFFFFFFFare monitored correctly.
Implications of workaround
No implications.
This will become part of the ETM specification.
Extra CPRT data traced—Category 2
Description
A.2.5
The ETM specification states that CPRT data (data transfers between ARM
processor registers and coprocessor registers, which are caused by the
instructions MRC, MCR, MRRC, and MCRR) must be traced if and only if
the MonitorCPRT bit is set (bit 1 of register 0x00, the control register).
Under some circumstances CPRT data can be traced when this bit is not
set.
20
Altera Corporation
Errata Sheet - Appendix A
Conditions
This occurs when both of the following are true:
■
■
ViewData is enabled.
Either data value or data address tracing are enabled (bits 3:2 of
register 0x00, the control register).
Implications
Extra CPRT data is traced. Since the number of such transfers is relatively
small, the amount of extra data generated should not be significant.
As there is no requirement for the tools to correctly handle CPRT data
traced when the MonitorCPRT bit is not set, this may result in unexpected
behavior.
Workaround
If this erratum causes an error in the development tools, the user must
always enable coprocessor register transfer tracing.
The tools must be able to handle CPRT data traced even when no CPRT
data was expected.
If this the extra trace is deemed to be confusing to the user, the tools can
ignore CPRT data when CPRT data tracing has been disabled.
Implications of workaround
A small amount of bandwidth may be wasted.
A.2.6
Stall cycles reported on wrong instruction—Category 2
Description
The timestamps or cycle numbers, if captured, might not truly reflect the number
of cycles taken by each instruction. Stall cycles which cause an instruction to take
longer than the minimum time to execute might be reported on an earlier
instruction.
Altera Corporation
21
Errata Sheet - Appendix A
It is occasionally useful to be able to determine the exact cycles on which
instructions are executed, for example to perform basic performance
analysis or to diagnose problems with the memory system. The
information can be preserved by one of two methods:
■
■
By selecting cycle-accurate mode, the ETM can be configured to cause
trace to be captured on every cycle during a trace region.
By using a trace port analyzer or logic analyzer which is capable of
saving timestamps with each cycle of trace.
Under some circumstances, extra instruction execution cycles are
reported alongside a previous instruction, making it appear that that
instruction took longer to execute instead. While the number of cycles that
a sequence of instructions takes to execute is correctly reported, the
number of cycles taken by individual instructions is not.
This erratum does not affect ETM9 Rev 0a, although ARM strongly
recommends that the latest revision of ETM9 is used in all new designs.
The erratum was introduced as a result of improvements to the
effectiveness of the FIFOFULLmechanism, although it is occurs whether
FIFOFULLis enabled or not.
Table 6 shows an example where a stall caused by an LDR is traced with
the wrong instruction. Note that the ETM reports an instruction when the
following instruction begins execution. Therefore the time taken by an
instruction is the time between that instruction and the previous
instruction in the trace.
Table 6. Example of correct and incorrect trace in the presence of stalls (Part 1 of 2)
Cycle Address Instruction
Actual trace
Expected
trace
Notes
500
501
502
503
504
505
506
507
508
1000
1004
1008
100C
1010
1014
1018
101C
NOP
Instruction enters ETM pipeline
NOP
1004 triggers tracing of 1000, 9 cycles later
NOP
NOP
NOP
NOP
NOP
LDR
(LDR stalled)
Most of ETM pipeline stalled along with processor
pipeline
509
510
1020
1024
NOP
B 1040
IE (instruction
IE for 1000
executed) for 1000
22
Altera Corporation
Errata Sheet - Appendix A
Table 6. Example of correct and incorrect trace in the presence of stalls (Part 2 of 2)
511
512
513
514
515
516
517
(Branch delay) WT (wait)
(Branch delay) IE for 1004
IE for 1004
IE for 1008
IE for 100C
IE for 1010
IE for 1014
IE for 1018
WT
Stall in ETM pipeline observed after only 3 cycles
1040
NOP
IE for 1008
IE for 100C
IE for 1010
IE for 1014
IE for 1018
Stall should be observed in line with other
instructions
518
519
520
IE for 101C
IE for 1020
WT
IE for 101C
IE for 1020
WT
Stalls caused by the processor are correctly
reported
521
522
WT
WT
IE for 1024
IE for 1024
Instruction traced upon completion
Conditions
As shown in Table 6, only external stalls that are caused by deasserting
CLKEN (such as memory stalls) are misreported. All stalls caused by the
processor, including branch delays, register interlocks and coprocessor
busy-waits, are reported correctly. External (CLKEN) stalls are reported 6
cycles early.
Implications
It is hard to use the timestamp information to gain information on stalls
caused by the memory system. While this is beyond the scope of the
original design aims of the ETM, precise stall information has proven to
be extremely useful to some users in debugging system level issues.
Workaround
It is often possible to model the behavior of the core as it would behave if
there were no external stalls to determine which stalls are external and
which are internal. To do this in all situations would require a complete
cycle-accurate model of the core. However, in most cases the number of
cycles required by each instruction can be easily predicted, and is
documented in the Technical Reference Manual for the core.
Altera Corporation
23
Errata Sheet - Appendix A
It is therefore possible to determine where external stall cycles have been
inserted, and to count forward 6 cycles from that point to find where they
should have been inserted, not counting other external stall cycles. This is
a complex procedure which cannot easily be automated, and cannot be
regarded as a complete workaround.
Table 7 shows an example of this technique. Note that, as in Table 6, the
trace reports stall cycles before the corresponding instruction is traced, not
after.
Table 7. Example of how to calculate the correct location of a stall (Part 1 of 2)
Cycle
Address
Instruction
Actual trace
Reconstructed
trace
Notes
600
2000
MOV r1,r0
601
602
2004
LDR r2,var0
(interlock on r2)
This is an internal stall and can
be predicted
603
604
605
2008
200C
ADD r2,r2,#1
B 2018
(Branch delay)
These cycles are also
predictable internal stalls
606
607
608
609
610
(Branch delay)
ADD r2,r2,#1
LDR r3,var1
(LDR stalled)
LDR r4,var2
2018
201C
An external stall
2020
2024
IE (instruction
executed) for 2000
IE for 2000
WT
611
ADD r2,r2,#1
WT (wait)
Only one stall was predicted for 2004.
One must be an external stall.
Counting forward 6 from the first puts
it on 2018, the second on 201C.
Remove it and try to place later.
612
613
614
615
616
617
WT
IE for 2004
IE for 2008
WT
IE for 2004
IE for 2008
WT
WT
WT
IE for 200C
IE for 2018
IE for 200C
The stall could correspond to 2018 in
theory, but in this system only the
memory system can cause external
stalls, and the code is in fast
instruction memory, so an ADD could
not cause an external stall.
24
Altera Corporation
Errata Sheet - Appendix A
Table 7. Example of how to calculate the correct location of a stall (Part 2 of 2)
618
IE for 2018
WT
The stall is far more likely to
correspond to 201C, as this causes a
LDR to on-chip memory. We can
therefore deduce that this LDR
caused precisely 1 stall cycle.
619
620
IE for 201C
IE for 2020
IE for 201C
IE for 2020
The stall could not correspond to this
load, as it is not 6 cycles ahead of a
detected external stall in the trace.
621
IE for 2024
IE for 2024
Note that the total number of cycles
taken for the sequence of instructions
is reported correctly.
Note: If 2018 is also an LDR, rather than an ADD, then it would not be
possible to determine which of the two instructions 2018 or 201C caused
the stall.
A.2.7
FIFOFULL LOW for one cycle during overflow—Category 3
Description
Although FIFOFULL correctly becomes asserted when the space available
in the FIFO is less than the minimum value specified in the FIFOFULL
Level register of the ETM (register 0x0b), it incorrectly becomes de-
asserted for a single cycle when the FIFO overflows.
Conditions
This occurs for one cycle, on the same cycle in which the FIFO overflows.
Implications
One extra cycle of trace can be lost during FIFO overflow. Since this case
only occurs when overflow is already about to occur, the ability of
FIFOFULL to prevent overflow is unaffected.
Workaround
None required. The erratum causes the loss of an extra cycle of trace when
there is an overflow, so the effect is not significant.
Implications of workaround
No implications.
Altera Corporation
25
Errata Sheet - Appendix A
A.2.8
Extra instruction traced prior to debug entry—Category 3
Description
An instruction selected as a breakpoint may be traced prior to debug
entry, even though it was not executed. It is flagged as having failed its
condition codes, regardless of whether it is conditional.
Conditions
This has been observed in devices based on the ARM9TDMI (for example
the ARM9TDMI, ARM920T, and ARM922T cores) only. It has not been
observed in devices based on the ARM9E (for example the ARM9E-S,
ARM966E-S, and ARM946E-S cores).
Implications
It will appear to the user that the breakpoint occurred at the wrong time.
Workaround
No workaround is known.
Implications of workaround
No implications.
Correction
No correction is planned.
26
Altera Corporation
Errata Sheet - Appendix A
A.3
Errata for the ARM922T Processor Core
A.3.1
LDM of user mode registers (ARM9TDMI–8)—Category 2
ARM9 Bug tracking database entry : CPC00_CAM_000013
Summary
Under specific conditions, a LDM to user mode registers will not operate
correctly. These instructions take the form:
LDM{<cond>}<addressing_mode> <Rn>,<registers_without_pc>^
These instructions are only used in system code and not in application
code.
Description
A LDM to user mode registers performs a load to the user mode registers
whilst the processor is in a privileged mode.
Under specific conditions (see below), this instruction will fail to operate
correctly. This results in not all the registers in the register list being
written correctly. It may also cause further failures, dependent on the
construction of the memory system, since the data memory request
signals are driven in an incorrect manner in the failing situation.
Example of failing instruction
LDM sp,{sp,lr}^
Conditions
This errata exists in the following circumstances:
A LDM to user mode registers, where:
1. The base register is register 8 or greater and
2. The base register is the first register (lowest register number) in the
register list and
3. There is more than one register is in the list.
Altera Corporation
27
Errata Sheet - Appendix A
Notes:
In all privileged modes this errata exists if the base register is 8 or greater.
This is not restricted to FIQ mode.
Instructions of the form
LDM{<cond>}<addressing_mode> <Rn>,<registers_and_pc}^
operate correctly since these are not LDM to user mode register
instructions. These instructions perform a return from exception.
Example of failing instructions
LDMIA sp,{sp,lr}^
LDMIA r8,{r8-r10}^
Examples that do NOT fail
LDMIA sp,{r8,sp,lr}^ ; Ok, since first register is not the base register
LDMIA r5,{r5,sp,lr}^ ; Ok, since base register is < r8
LDMIA sp,{sp}^
; Ok, since only one register in the list
LDMIA sp,{sp,lr,pc}^; Ok, since PC in the list and hence is a return from
; exception instruction and not a load of user mode
; registers.
Implications
This errata only applies to hand crafted assembler code, as the ARM
compiler does not generate such instructions. The use of this instruction is
typically limited to a few places in exception handlers, thus limiting the
scope of this erratum.
Work-around
This instruction is only used by hand crafted assembler code. The ARM
compiler will not generate this instruction. To work-around this errata the
LDM should be split into two separate instructions.
28
Altera Corporation
Errata Sheet - Appendix A
e.g. The instruction:
LDMIA sp,{sp,lr}^
should be split into
LDMIA sp,{sp}^
LDMIA sp,{lr}^
and
LDMIA r8,{r8-lr}^
should be split into
LDMIA r8,{r8}^
LDMIA r8,{r9-lr}^
A.3.2
Debug Request coincident with Pipeline Hazards (ARM9TDMI–9)
—Category 2
Summary
The processor may return to normal program execution at an incorrect
point if EDBGRQ or the scan chain created debug request is asserted
whilst it is performing certain tasks.
Conditions
There are three scenarios in which this erratum can occur:
1. Debug Request occurs whilst the processor is waiting for a
coprocessor instruction to be completed by a coprocessor or
2. Debug Request occurs whilst the recognition of a Data Abort is in
progress or
3. Debug Request occurs whilst the recognition of an instruction
causing a watchpoint is in progress.
1
Recognition of a Data Abort is said to be in progress if the last
memory access asserted the DABORTsignal, but the processor
has not yet begun execution at the Data Abort vector.
Altera Corporation
29
Errata Sheet - Appendix A
1
Recognition of a watchpoint is said to be in progress if the last
memory access caused a watchpoint, but debug entry has not yet
completed.
If the above conditions are met then the debug entry mechanism fails to
behave in the defined manner and the device may return from debug and
execute from an incorrect address.
Implications
For each scenario the following implications are expected given the above
conditions:
Busy-waiting Coprocessor Instructions
In this form the debug request supersedes the currently executing
coprocessor instruction, which would normally not occur. As such the PC
is not the correct value as debug entry proceeds. Correspondingly, on
return from debug the standard return address calculation produces an
incorrect address and thus unintended or unpredictable device behaviour
may result.
Data Aborts
In this form the debug request causes correct debug entry and exit.
However, the link register address calculation for the return from the Data
Abort handler is incorrect. Correspondingly, on return from the Data
Abort handler unintended or unpredictable device behaviour may result.
Watchpoints
In this form the debug request causes debug entry first, but the
watchpoint is still pending and has yet to execute the next instruction
before it takes full effect. As debug entry has already occurred this next
instruction will be the first instruction within debug mode to be executed.
After this instruction executes, debug entry occurs for a second time, even
though the device is already in debug, and results in the premature exiting
of debug. Consequently, the return from debug is to an incorrect address
and thus unintended or unpredictable device behaviour may result.
30
Altera Corporation
Errata Sheet - Appendix A
Workarounds
There is no practical workaround for this erratum. This is due to the
difficulty in getting a debugging tool to recognise the symptoms of this
erratum and take the appropriate corrective action. However the
likelihood of failure with this mechanism is extremely low and the impact
of failure is also low as it affects debugging operations only, therefore
there is no plan to revise the design to resolve this erratum.
A.3.3
Data Abort and Watchpoint with Breakpoint Following
(ARM9TDMI–10)—Category 2
Summary
The processor may fail to execute the abort handler if a data abort occurs
on a watchpointed instruction and a breakpoint is in the execution
pipeline.
Conditions
The conditions for this erratum are:
1. An instruction that causes both a Data Abort and a watchpoint to
occur and
2. The execution of the following instruction will cause a breakpoint to
occur
1
There should be no data dependency between the two
instructions such that a pipeline interlock occurs. If there is such
data dependency then the processor behaves correctly.
If the above conditions are met then the Data Abort may be missed.
Implications
In this erratum Data Abort entry is halted by debug entry and this causes
the state indicating a Data Abort to be lost on return from debug. Hence
the Data Abort handler will fail to be invoked. This may result in
unintended or unpredictable device behaviour.
Altera Corporation
31
Errata Sheet - Appendix A
Workarounds
There is no practical workaround for this erratum. This is due to the
difficulty in getting a debugging tool to recognise the symptoms of this
erratum and take the appropriate corrective action. However the
likelihood of failure with this mechanism is extremely low and the impact
of failure is also low as it affects debugging operations only, therefore
there is no plan to revise the design to resolve this erratum.
A.3.4
Watchpoint coincident with Debug Request (ARM9TDMI–11)—
Category 2
Summary
The processor can falsely exit debug state and continue execution if
EDBGRQ or the scan chain created debug request is asserted whilst
debug entry is occurring.
Conditions
A combination of two conditions is required to cause this erratum:
1. Debug request is asserted and
2. An instruction that generates a watchpoint is executing
If the above conditions are met then the debug entry mechanism fails to
behave in the defined manner and the device may return from debug and
execute from the incorrect address.
Implications
In this erratum, the debug request takes affect before the complete
recognition of the watchpoint. This may result in unreliable debug entry,
the watchpoint being missed and premature exit of debug state. Thus
unintended or unpredictable device behaviour may result.
Workarounds
There is no practical workaround for this erratum. This is due to the
difficulty in getting a debugging tool to recognise the symptoms of this
erratum and take the appropriate corrective action. However the
likelihood of failure with this mechanism is extremely low and the impact
of failure is also low as it affects debugging operations only, therefore
there is no plan to revise the design to resolve this erratum.
32
Altera Corporation
Errata Sheet - Appendix A
A.3.5
Register controlled shift data operations where the destination is
the PC (ARM9TDMI–1)—Category 3
ARM9 Bug tracking database entry : CPC00_CAM_000001
Summary
A data operation with a register controlled shift to the PC may calculate
an incorrect result
Description
This fault is exhibited by any data operation involving a register-specified
shift with the Program Counter as the destination. This is in effect a
calculated branch, with an unusual branch address calculation. The
branch target address calculated by the instruction is incorrect.
The ARM C compiler will not generate this instruction. Other compilers
are extremely unlikely to produce this instruction, as no standard high
level languages source code constructs which would map to this
instruction.
It is also extremely unlikely that this instruction has been used in
assembler code, as is explained below.
Conditions
Exists for any instruction that involves a register-specified shift or rotate
operation with the PC as the destination for the resulting data. The fault
does not occur for an immediate specified shift or rotate to the PC.
Implications
Data operations involving a register controlled shift and where the
destination register is the PC have an unpredictable behaviour on an
ARM9TDMI processor core and any processor containing an
ARM9TDMI; for example ARM920T.
The current ARM compiler cannot generate instructions of this class and
so this would only be encountered in hand coded assembler. ARM’s
software staff have considered possible uses for such an instruction in
assembler code, and have only found one theoretical use for this
instruction, which is a strange branch table described below.
For these reasons this erratum is not expected to cause any restrictions or
problems in using the ARM9TDMI.
Altera Corporation
33
Errata Sheet - Appendix A
Examples:
MOV(S) PC, Rm, <SHIFTOP> Rs
ADD(S) PC, Rn, Rm, <SHIFTOP> Rs
The only theoretical use for these instructions which has been suggested
would be for a branch table that was organised in powers of 2.
MOV
r0,#0x100
MOV
r1,#0x4
ADD
PC,r0,r1, LSL r2
0x1004:
Code seq1 : 1 instruction
Code seq2:2 instructions
Code seq3:4 instructions
Code seq4:8 instructions
Code seq5:16 instructions
Code seq6:32 instructions
Code seq7:64 instructions
Code seq8:128 instructions
0x1008 -> 0x100C:
0x1010 -> 0x101C:
0x1020 -> 0x103C:
0x1040 -> 0x107C:
0x1080 -> 0x10FC:
0x1100 -> 0x11FC:
0x1200 -> 0x13FC:
To date ARM knows of no examples of an application for a branch table
organised in such a manner. However, if one was required then an
alternative code sequence could be used.
Copyright © 2002 Altera Corporation. All rights reserved. Altera, The Programmable Solutions Company, the
stylized Altera logo, specific device designations, and all other words and logos that are identified as
trademarks and/or service marks are, unless noted otherwise, the trademarks and service marks of Altera
Corporation in the U.S. and other countries. All other product or service names are the property of their
respective holders. Altera products are protected under numerous U.S. and foreign patents and pending
applications, mask work rights, and copyrights. Altera warrants performance of its
semiconductor products to current specifications in accordance with Altera’s standard
warranty, but reserves the right to make changes to any products and services at any time
without notice. Altera assumes no responsibility or liability arising out of the application
or use of any information, product, or service described herein except as expressly agreed
to in writing by Altera Corporation. Altera customers are advised to obtain the latest
version of device specifications before relying on any published information and before
placing orders for products or services.
101 Innovation Drive
San Jose, CA 95134
(408) 544-7000
http://www.altera.com
Applications Hotline:
(800) 800-EPLD
Literature Services:
lit_req@altera.com
34
Altera Corporation
相关型号：
©2020 ICPDF网 联系我们和版权申明