GR740-UM-DS, Nov 2017, Version 1.7
483
www.cobham.com/gaisler
GR740
The error caused by the bug is that nPC (%l2 in the trap handler's window) is not generated correctly,
causing execution flow to go wrong after the interrupt handler returns.
Workaround:
Disable the DBPM function on boot (before I-cache is enabled) by clearing bit 25 in
the ASR17 register.
Applicable to:
This issue is only present in silicon revision 0.
43.2.3 LEON4 data cache clock-gating after device reset
After device reset the first processor will be enabled, unless the BREAK signal is used to force it into
power-down mode, and the remaining processors will be clock gated. As part of the processor reset
sequence, the LEON4 will clear (invalidate) its cache. This operation requires at least 128 clock
cycles after reset. The clock gating in the GR740 will clock-gate off the processors before this opera-
tion has completed. Leading to the data cache RAM enable signals being left asserted. The data cache
RAMs are clocked by a running clock in order to enable bus snooping. In applications where CPU1-3
are never enabled, this bug may lead to unnecessary dynamic power consumption by the unused
CPUs. Otherwise, it has no functional impact. Maximum power impact has been measured to 40 mW
(at room temperature).
Workaround:
For the processors that are not used in the application, during the boot process turn on
their clocks for 128 cycles using the clock gating unit's clock override register and then back off
again. For applications using all CPUs, the issue is resolved when CPU1-3 are started up and no
workaround is needed.
Applicable to:
This issue is only present in silicon revision 0.
43.2.4 LEON4 partial WRPSR
The LEON4 partial WRPSR is incorrectly implemented in silicon revision 0. Partial WRPSR may
update the ICC and CWP fields of %psr. The effect of the issue is that WRPSR with the rd field set to
a non-zero will update the ET field, may or may not update the ICC and CWP field, and will leave the
remaining fields untouched.
Workaround:
None
Applicable to:
This issue is only present in silicon revision 0.
43.2.5 Ethernet debug communication link (EDCL) hang after AMBA ERROR
The Ethernet Debug Communication Link (EDCL) may hang when it receives an AMBA ERROR
response. The issue only affects the debug link functionality where the debug link communication
may stop, causing the debug monitor to time out. The rest of the GR740 device remains operational.
Workaround:
No workaround currently exists. The debug link should not be used to test IOMMU
protection functionality, injection of uncorrectable errors in Level-2 cache and external memory, or
other tasks that create AMBA ERROR responses.
Applicable to:
This issue is only present in silicon revision 0.
43.2.6 Memory controller incompatibility in full width mode with UT8SDMQ64M48
The SDRAM controller has a special case for SDRAM geometry selection that leads to the controller
configured for the largest bank size in 64-bit mode requiring SDRAM that has a geometry of 4096
rows x 4096 columns. This prevents full use of Cobham UT8SDMQ64M48 memories to create a 2
GiB external SDRAM area.
Workaround:
None.
Applicable to:
This issue is only present in silicon revision 0. This has been fixed in silicon revision
1 by changing the interpretation of the banks-size field in full-width mode. Now the same
BANKSZ,COLSZ settings should be used for full-width as in half-width mode.