README.riscv64

Status
~~~~~~

The RISC-V port targets the 64-bit RISC-V architecture and the Linux operating
system. The port has been tested to work on real hardware and under QEMU.

The following ISA base and extensions are currently supported:

| Name         | Description                       | #Instrs | Notes    |
| ------------ | --------------------------------- | ------- | -------- |
| RV64I        | Base instruction set              |   52/52 |          |
| RV64M        | Integer multiplication & division |   12/13 | (1)      |
| RV64A        | Atomic                            |   22/22 | (2)      |
| RV64F        | Single-precision floating-point   |   30/30 | (3)      |
| RV64D        | Double-precision floating-point   |   32/32 |          |
| RV64Zicsr    | Control & status register         |     3/6 | (4), (5) |
| RV64Zifencei | Instruction-fetch fence           |     0/1 | (6)      |
| RV64C        | Compressed                        |   37/37 |          |

Notes:
(1) MULHSU is not recognized.
(2) LR and SC use the VEX "fallback" method which suffers from the ABA problem.
(3) Operations do not check if the input operands are correctly NaN-boxed.
(4) CSRRWI, CSRRSI and CSRRCI are not recognized.
(5) Only registers fflags, frm and fcsr are accepted.
(6) FENCE.I is not recognized.


Implementation tidying-up/TODO notes
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Implement a proper "non-fallback" method for LR and SC instructions.
* Add a check for correct NaN-boxing of 32-bit floating-point operands.
* Optimize instruction selection, in particular make more use of <instr>i
  variants.
* Optimize handling of floating-point exceptions. Avoid helpers and calculate
  exception flags using the same instruction which produced an actual result.
* Review register usage by the codegen.
* Avoid re-use of Intel-constants CFIC_IA_SPREL and CFIC_IA_BPREL. Generalize
  them for all architectures or introduce same CFIC_RISCV64_ variants.
* Get rid of the typedef of vki_modify_ldt_t in include/vki/vki-riscv64-linux.h.
* Review if setup_client_stack() should expose AT_SYSINFO_EHDR to clients.
* Make sure that the final exit sequence in run_a_thread_NORETURN() is not racy
  in regards to accessing the thread state.