Jeffnye.dromajo stf update

[jeffnye-gh]

This branch uses latest dromajo f3c3112, the changes move stf start/stop macro detection and trace generation into include/dromajo_template.h, within glue(riscv_cpu_interp, XLEN)(RISCVCPUState *s, int n_cycles)

I have done limited testing with dhrystone and coremark, single cpu, with zstf and stf. stf_diff reports no differences. (clarify: no differences between this and stf's generated by previous impl.)

The main loop body can exit w/ a goto, illegal_instructions/etc. The tracing code makes the assumption that exceptions are not traced. I'm looking at this more.

other minor changes:

• moved some globals into a struct, stf_writer remains global
• I've added a switch --no_stf_priv_check which can disable that check in the tracing enabled loop
• added tohost/fromhost (clarify: struct variables added only, in advance of support)

[klingaard]

Hey @jeffnye-gh, do you think you can finish this PR up and mark it for review? This is good stuff.

[jeffnye-gh]

Hi Knute,

Previous implementation was not as clean as it looked.

I have an updated implementation on a branch in https://github.com/Condor-Performance-Modeling/dromajo, but I have not prepared a new PR.

The remaining problem is a stack pointer (x2) initialization difference at startup in dhrystone on linux. I need to figure out why.

• 1: 0000000000010210 : 000066a2 c.ldsp x13,8(x2) MEM 0000003fffa49ca8: [ 0000000000000000 ]
• 1: 0000000000010210 : 000066a2 c.ldsp x13,8(x2) MEM 0000003fffb53c98: [ 0000000000000000 ]

Day job priorities got in the way, I'll be able to get back to this on the weekend.

jeff

[ghost]

No worries. Just wanted to make sure you weren't waiting on anything from our end.