Requirements
------------
- You need GCC 3.4 or later to compile the s390 port.
- To run valgrind a z10 machine or any later model is recommended.
Older machine models down to and including z990 may work but have
not been tested extensively.
Limitations
-----------
- 31-bit client programs are not supported.
- Hexadecimal floating point is not supported.
- Transactional memory is not supported. The transactional-execution
facility is masked off from HWCAP.
- A full list of unimplemented instructions can be retrieved from
`docs/internals/s390-opcodes.csv', by grepping for "not implemented".
- FP signalling is not accurate. E.g., the "compare and signal"
instructions behave like their non-signalling counterparts.
- On machine models predating z10, cachegrind will assume a z10 cache
architecture. Otherwise, cachegrind will query the hosts cache system
and use those parameters.
- Some gcc versions use mvc to copy 4/8 byte values. This will affect
certain debug messages. For example, memcheck will complain about
4 one-byte reads/writes instead of just a single read/write.
Hardware facilities
-------------------
Valgrind does not require that the host machine has the same hardware
facilities as the machine for which the client program was compiled.
This is convenient. If possible, the JIT compiler will translate the
client instructions according to the facilities available on the host.
This means, though, that probing for hardware facilities by issuing
instructions from that facility and observing whether SIGILL is thrown
may not work. As a consequence, programs that attempt to do so may
behave differently. It is believed that this is a rare use case.
Reading Material
----------------
(1) ELF ABI s390x Supplement
https://github.com/IBM/s390x-abi/releases
(2) z/Architecture Principles of Operation
https://publibfp.dhe.ibm.com/epubs/pdf/a227832d.pdf
(3) Collection of z/Architecture publications
https://linux.mainframe.blog/zarchitecture-principles-of-operation/
