ImperasDV PBMT misaligned Mismatch

[rosethompson]

PBMT configured misaligned addresses do not agree between ImperasDV and Wally.

wsim rv64gc tests/coverage/tlbMisaligned.elf --lockstep

 Info (IDV) Instruction executed prior to mismatch '0x8001503c(main+3c): 929e     add     x5,x5,x7'
# Error (IDV) PC mismatch (HartId:0, PC:0x000000008001503e main+3e):
# Error (IDV) Mismatch 0>
# Error (IDV)   . dut:0x0000000080000050 
# Error (IDV)   . ref:0x000000008001503e main+3e

[rosethompson]

@eroom1966 I think this might come down to how the spec is interpreted.

Zicclsm Misaligned loads and stores to main memory regions with both the cacheability and coherence PMAs must be supported.

Since PMBT is used to disable cacheablity I believe it should not support misaligned access even though the memory physically does support misaligned.

[rosethompson]

My attempt to generate the debug database resulted in a segfault.

I added this to the vlog command.
+define+IDV_INCLUDE_TRACE2BIN

and set this variable before calling questa.
IMPERAS_RVVI_DEBUG=1

# Platform = "linux_x86_64"
# Signature = ce21318267b9d07e1c15c754173193d2
# --> START OF USERCODE
# 0    0x00007ffff2825114: 'rvviDebugCaptureClock + 0x0000000000000054' in '/opt/riscv/ImperasDV-OpenHW/lib/Linux64/ImperasLib/imperas.com/verification/riscv/1.0/model.so'
# <-- END OF USERCODE
# 1    0x00007feff05a3718: '/opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv:24'
# 2    0x00007feff416c2c5: '/opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv:193'
# 3    0x00000000023e5c82: '<unknown (@0x23e5c82)>'
# 4    0x00000000004df304: '<unknown (@0x4df304)>'
# 5    0x000000000074da13: '<unknown (@0x74da13)>'
# 6    0x0000000000ca62ed: '<unknown (@0xca62ed)>'
# 7    0x0000000000cab610: '<unknown (@0xcab610)>'
# 8    0x0000000000cacf8e: '<unknown (@0xcacf8e)>'
# 9    0x0000000000f9c58d: '<unknown (@0xf9c58d)>'
# 10   0x000000000287b08d: '<unknown (@0x287b08d)>'
# 11   0x000000000287f4e6: '<unknown (@0x287f4e6)>'
# 12   0x0000000002880bd1: '<unknown (@0x2880bd1)>'
# 13   0x0000000002880f36: '<unknown (@0x2880f36)>'
# 14   0x000000000122f669: '<unknown (@0x122f669)>'
# 15   0x0000000002920e3f: '<unknown (@0x2920e3f)>'
# 16   0x00000000029749b7: '<unknown (@0x29749b7)>'
# 17   0x00000000029378a7: '<unknown (@0x29378a7)>'
# 18   0x0000000002937b89: '<unknown (@0x2937b89)>'
# 19   0x0000000002708d7d: '<unknown (@0x2708d7d)>'
# 20   0x0000000000c6e53c: '<unknown (@0xc6e53c)>'
# End of Stack Trace


# ** Fatal: (SIGSEGV) Bad handle or reference.
#    Time: 0 ns  Iteration: 1  Process: /testbench/rvvi/#ALWAYS#114 File: /opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv
# Fatal error at /opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv line 24
# 
# HDL call sequence:
# Stopped at /opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv 24  
# called from  /opt/riscv/ImperasDV-OpenHW/ImpProprietary/source/host/idv/trace2bin.sv 193  
# 
# End time: 08:55:28 on Oct 14,2024, Elapsed time: 0:00:01
# Errors: 1, Warnings: 0

[rosethompson]

tlbMisaligned.tar.gz

[eroom1966]

There is a configuration option, to provide the behavior you are looking for
Zicclsm Set parameter unaligned to T

[davidharrishmc]

Lee, I think there may be a misunderstanding. We already have --override cpu/unaligned=T # Zicclsm (should be true) in the configuration. That correctly supports misaligned accesses to cachable main memory, without regard to PBMT. The issue we are concerned about is that when unaligned=T, ImperasDV also allows misaligned accesses to pages marked as uncachable in PBMT. Wally implements misaligned access via the cache, so it cannot support uncachable misaligned accesses. We believe an additional configuration option is necessary to control misaligned accesses based on PBMT, when Zicclsm is implemented (unaigned=T).

…

On Dec 6, 2024, at 1:32 AM, Lee Moore ***@***.***> wrote: There is a configuration option, to provide the behavior you are looking for Zicclsm Set parameter unaligned to T — Reply to this email directly, view it on GitHub <#976 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AR4AA36ASFZMDLBOVZQ23VD2EFVMTAVCNFSM6AAAAABPEI4N2SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMRSGY2DGMRRGQ>. You are receiving this because you are subscribed to this thread.

[rosethompson]

Fixed with --cpu/aligned_uncached_PBMT=F. Thank you Lee!