Track RISC-V64 migration

[siraben]

Summary

• Switch the runtime build over to the RISC-V64 guest path and wire the RISC-V emulator, decoder/interpreter, Asbestos scaffolding, syscall dispatch, ELF loading, signals, time/stat layouts, TLS, uname, and proc/cpuinfo handling.
• Add RISC-V helper tools and selftests (rvdecode, rvrun, rvdecode-selftest, rvinterp-selftest).
• Keep rootfs handling aligned with the existing iSH project shape: root filesystems are downloaded/created as build or local artifacts, not committed as an expanded roots/ tree. The local .rv64-rootfs-cache/ and roots/ directories are ignored.
• Add a cross-ISA benchmark harness for i386 vs riscv64 with freestanding C payloads and guest shell/Python/gzip workloads.

Benchmark Figures

Filtered local run:

RUNS=3 WARMUPS=1 \
I386_BUILD=/Users/siraben/Git/ish/build \
RV64_BUILD=/Users/siraben/Git/ish-riscv/build-bench-rv64 \
RV64_FS=/tmp/ish-rv64-testfs \
OUT=/Users/siraben/Git/ish-riscv/e2e_out/bench-compare \
BENCH_FILTER='^(prime_sieve|mandelbrot|file_io|pipe_throughput|gzip_payload)$' \
./benchmarks/hot_paths/compare.sh

benchmark	i386 best (s)	i386 avg (s)	i386 RSD %	riscv64 best (s)	riscv64 avg (s)	riscv64 RSD %	riscv64/i386 best
file_io	0.010	0.010	0.0	0.090	0.097	4.9	9.00x
gzip_payload	0.990	1.003	0.9	28.590	28.883	1.2	28.88x
mandelbrot	0.040	0.047	10.1	5.190	5.257	0.9	129.75x
pipe_throughput	0.010	0.010	0.0	0.090	0.090	0.0	9.00x
prime_sieve	0.020	0.020	0.0	2.120	2.150	1.0	106.00x

Note: this is a smoke run. The i386 C benchmarks are short enough that timer resolution dominates several rows, so these figures are useful for tracking large deltas but not publication-grade performance claims.

Validation

• meson setup build-bench-rv64 . --reconfigure
• ninja -C build-bench-rv64 ish rvdecode rvdecode-selftest rvinterp-selftest
• meson test -C build-bench-rv64 --print-errorlogs
• Compiled freestanding C benchmark payloads for i386 and riscv64 with Homebrew LLVM clang.
• Ran the filtered 3-sample compare shown above.

Tracking Notes

This is a tracking PR against siraben/ish:master for the RISC-V64 migration branch. Follow-up benchmark runs are posted as PR comments so the body stays stable.

[siraben]

Follow-up benchmark after ARM64 fiber routing plus core integer lowering (126e82c7). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.007884s	0.008049s	1.46%	0.014494s	0.014953s	2.30%	1.838x
gzip_payload	0.920000s	0.923333s	0.51%	21.690000s	21.726667s	0.15%	23.576x
mandelbrot	0.042622s	0.043775s	2.18%	3.208335s	3.219829s	0.29%	75.274x
pipe_throughput	0.009748s	0.009911s	1.20%	0.015543s	0.016413s	6.66%	1.594x
prime_sieve	0.016517s	0.016603s	0.52%	0.926400s	0.939972s	1.99%	56.088x

[siraben]

Follow-up benchmark after e74a02f0 (Lower RV64 integer ALU ops to ARM64 gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008350s	0.008489s	1.83%	0.012117s	0.012853s	6.24%	1.451x
gzip_payload	0.940000s	0.946667s	0.50%	16.840000s	16.863333s	0.10%	17.915x
mandelbrot	0.045026s	0.046297s	1.94%	1.432628s	1.439321s	0.33%	31.818x
pipe_throughput	0.010125s	0.010269s	1.49%	0.014651s	0.014929s	2.30%	1.447x
prime_sieve	0.017013s	0.017226s	1.64%	0.955603s	0.958939s	0.43%	56.169x

[siraben]

Follow-up benchmark after 94823519 (Lower RV64 control flow to ARM64 gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008395s	0.008573s	1.62%	0.010521s	0.010940s	2.75%	1.253x
gzip_payload	0.940000s	0.943333s	0.50%	10.550000s	10.556667s	0.04%	11.223x
mandelbrot	0.043764s	0.044724s	2.22%	0.839010s	0.841588s	0.28%	19.171x
pipe_throughput	0.010258s	0.010447s	1.30%	0.012707s	0.012775s	0.69%	1.239x
prime_sieve	0.017631s	0.017795s	0.78%	0.449544s	0.452858s	0.94%	25.497x

[siraben]

Follow-up benchmark after 06bdc445 (Lower RV64 stores to ARM64 TLB gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008280s	0.008493s	2.08%	0.010246s	0.010427s	1.29%	1.237x
gzip_payload	0.970000s	0.990000s	2.18%	7.180000s	7.303333s	1.92%	7.402x
mandelbrot	0.045154s	0.049109s	8.50%	0.839916s	0.842363s	0.23%	18.601x
pipe_throughput	0.009974s	0.010121s	1.05%	0.012116s	0.012436s	1.94%	1.215x
prime_sieve	0.016893s	0.017417s	2.20%	0.198344s	0.201706s	1.52%	11.741x

This checkpoint keeps fast stores enabled. Fast load gadgets are selected in the later load checkpoint.

[siraben]

Follow-up benchmark after 8503879b (Lower RV64 multiply and divide ops to ARM64 gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
mandelbrot	0.044716s	0.046076s	2.67%	0.114958s	0.117289s	1.63%	2.571x
prime_sieve	0.017439s	0.017729s	1.39%	0.193336s	0.195130s	0.78%	11.086x

Compared with the store checkpoint, mandelbrot improved from 0.839916s to 0.114958s best. prime_sieve is roughly unchanged because this benchmark is not multiply/divide-heavy.

[siraben]

Full benchmark set after 8503879b (Lower RV64 multiply and divide ops to ARM64 gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008684s	0.008781s	0.79%	0.009544s	0.009789s	1.78%	1.099x
gzip_payload	0.950000s	0.976667s	2.10%	7.150000s	7.220000s	1.09%	7.526x
mandelbrot	0.044716s	0.046076s	2.67%	0.114958s	0.117289s	1.63%	2.571x
pipe_throughput	0.010972s	0.011203s	1.50%	0.012366s	0.012589s	2.15%	1.127x
prime_sieve	0.017439s	0.017729s	1.39%	0.193336s	0.195130s	0.78%	11.086x

[siraben]

Full benchmark set after ae98c58c (Lower RV64 loads to ARM64 TLB gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008734s	0.009221s	6.00%	0.009273s	0.009407s	1.03%	1.062x
gzip_payload	0.950000s	0.950000s	0.00%	1.630000s	1.643333s	0.76%	1.716x
mandelbrot	0.043968s	0.045213s	2.68%	0.120710s	0.121792s	0.64%	2.745x
pipe_throughput	0.010091s	0.010158s	0.90%	0.010644s	0.011026s	2.80%	1.055x
prime_sieve	0.017066s	0.017416s	1.58%	0.076009s	0.076805s	0.75%	4.454x

The load slow-path bug was INT_NONE == -1 being treated as nonzero failure; after fixing that, the load gadgets are selected for all integer load widths.

[siraben]

Full benchmark set after 86e85033 (Lower RV64 FP memory ops to ARM64 gadgets). Updated after amending the benchmark harness so C benchmarks time themselves with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008546s	0.008644s	0.99%	0.009253s	0.009282s	0.36%	1.083x
gzip_payload	0.930000s	0.943333s	1.32%	1.600000s	1.643333s	2.55%	1.720x
mandelbrot	0.044494s	0.045907s	2.18%	0.120467s	0.121899s	0.84%	2.707x
pipe_throughput	0.010125s	0.010185s	0.82%	0.011163s	0.011645s	3.01%	1.103x
prime_sieve	0.017008s	0.017226s	1.52%	0.075826s	0.076330s	0.65%	4.458x

[siraben]

Full benchmark set after 5a63e240 (Lower RV64 system traps to ARM64 gadgets). C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008745s	0.008975s	2.74%	0.009315s	0.010619s	16.30%	1.065x
gzip_payload	0.940000s	0.966667s	1.95%	1.580000s	1.593333s	0.59%	1.681x
mandelbrot	0.046242s	0.047120s	1.35%	0.118884s	0.124468s	3.44%	2.571x
pipe_throughput	0.010067s	0.010249s	1.56%	0.010798s	0.011505s	4.36%	1.073x
prime_sieve	0.017192s	0.017722s	3.11%	0.077050s	0.077696s	0.77%	4.482x

Validation at this head: rvinterp-selftest passes, and /bin/sh -lc 'echo syscall-ok; uname -m; true' exits 0 and reports riscv64.

[siraben]

Follow-up benchmark after e7aee4eb (Lower RV64 high multiply ops to ARM64 gadgets). C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008027s	0.008192s	1.99%	0.008262s	0.008471s	1.84%	1.029x
gzip_payload	0.930000s	0.933333s	0.51%	1.500000s	1.510000s	0.54%	1.613x
mandelbrot	0.042887s	0.043409s	0.91%	0.107185s	0.108875s	1.50%	2.499x
pipe_throughput	0.010190s	0.010572s	3.68%	0.010444s	0.010483s	0.52%	1.025x
prime_sieve	0.016465s	0.016610s	1.02%	0.073123s	0.073460s	0.37%	4.441x

Validation: rvinterp-selftest passes with explicit mulh, mulhsu, and mulhu coverage; RV64 /bin/sh smoke exits 0.

[siraben]

Follow-up benchmark after b7e36c9e (Lower RV64 fences to ARM64 gadgets). C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008100s	0.008318s	1.98%	0.008364s	0.008575s	2.06%	1.033x
gzip_payload	0.930000s	0.933333s	0.51%	1.510000s	1.516667s	0.31%	1.624x
mandelbrot	0.042662s	0.043384s	2.21%	0.109789s	0.110656s	0.62%	2.573x
pipe_throughput	0.009718s	0.009937s	2.11%	0.010466s	0.011101s	4.16%	1.077x
prime_sieve	0.016289s	0.016850s	3.62%	0.074508s	0.074990s	0.77%	4.574x

This is expected to be mostly neutral on the current suite; it removes fence/fence.i fallback overhead and adds explicit selftest coverage.

[siraben]

Follow-up benchmark after 5c161884 (Lower RV64 CSR ops through ARM64 gadgets). C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008161s	0.008383s	2.41%	0.008836s	0.008876s	0.36%	1.083x
gzip_payload	0.950000s	0.960000s	0.85%	1.500000s	1.533333s	2.22%	1.579x
mandelbrot	0.045695s	0.046493s	1.24%	0.107335s	0.109416s	1.35%	2.349x
pipe_throughput	0.010221s	0.010879s	4.72%	0.010225s	0.010731s	3.62%	1.000x
prime_sieve	0.016651s	0.016949s	1.25%	0.077801s	0.078485s	1.00%	4.672x

Validation: rvinterp-selftest passes with fcsr CSR read/write coverage; RV64 /bin/sh smoke exits 0.

[siraben]

Follow-up benchmark after d3d35888 (Lower RV64 atomics through ARM64 gadgets). C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008168s	0.008357s	1.79%	0.008174s	0.008573s	3.51%	1.001x
gzip_payload	0.940000s	0.950000s	0.86%	1.510000s	1.526667s	0.82%	1.606x
mandelbrot	0.043525s	0.044684s	1.88%	0.102192s	0.104110s	2.28%	2.348x
pipe_throughput	0.009814s	0.010094s	1.97%	0.010010s	0.010270s	2.71%	1.020x
prime_sieve	0.016429s	0.016639s	1.20%	0.074619s	0.076847s	2.34%	4.542x

Validation: rvinterp-selftest passes with amoadd.d coverage; RV64 /bin/sh smoke exits 0.

[siraben]

Follow-up benchmark after 265a7c9e (Lower RV64 FP arithmetic through ARM64 gadgets). This lowers OP-FP and fused multiply-add/subtract via a parameterized helper gadget matching interpreter semantics. C benchmarks are self-timed with clock_gettime; gzip_payload is still shell-timed with /usr/bin/time. RUNS=3 WARMUPS=1.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.008180s	0.008247s	1.03%	0.008356s	0.008496s	1.17%	1.022x
gzip_payload	0.930000s	0.933333s	0.51%	1.510000s	1.520000s	0.54%	1.624x
mandelbrot	0.043210s	0.043817s	0.98%	0.105499s	0.107254s	2.13%	2.442x
pipe_throughput	0.009752s	0.009899s	1.59%	0.010509s	0.010688s	1.67%	1.078x
prime_sieve	0.018427s	0.020644s	14.92%	0.073609s	0.074472s	1.09%	3.995x

Validation: rvinterp-selftest passes its fadd.d path through the helper gadget; RV64 /bin/sh smoke exits 0.

[siraben]

Broad benchmark after 265a7c9e with all current lowering batches through FP arithmetic. C benchmarks are self-timed with clock_gettime; shell/gzip rows use /usr/bin/time. RUNS=3 WARMUPS=1.

RV64 rootfs note: /usr/bin/python3 and /bin/bash are not present in /tmp/ish-rv64-testfs, so RV64 skipped python_startup, python_compute, python_imports, and bash_control; those i386-only rows are not in the comparison table below.

benchmark	i386 best	i386 avg	i386 RSD	riscv64 best	riscv64 avg	riscv64 RSD	best ratio
file_io	0.009550s	0.009714s	1.65%	0.009052s	0.009410s	3.10%	0.948x
fs_metadata	1.110000s	1.116667s	0.84%	1.380000s	1.393333s	0.68%	1.243x
gzip_payload	0.940000s	0.943333s	0.50%	1.580000s	1.606667s	1.92%	1.681x
mandelbrot	0.043309s	0.044596s	3.29%	0.109954s	0.110777s	0.54%	2.539x
pipe_throughput	0.009952s	0.010222s	1.97%	0.011478s	0.011717s	2.08%	1.153x
prime_sieve	0.016805s	0.017439s	2.82%	0.076001s	0.076842s	0.77%	4.523x
shell_control	1.880000s	1.900000s	1.49%	2.680000s	2.696667s	0.63%	1.426x
shell_pipeline	0.600000s	0.610000s	1.34%	0.860000s	0.880000s	2.45%	1.433x
shell_startup	0.000000s	0.000000s	0.00%	0.010000s	0.010000s	0.00%	inf

Coverage status: every decoded RV64 op class now has a lowered path: base integer, branches/jumps, loads/stores, FP memory, M-extension including high multiply, fences, traps/CSR, atomics, OP-FP, and FP fused multiply-add/subtract. Generic rv_exec_one_gadget remains for invalid/unsupported decoded combinations and any future decoder expansion.

[siraben]

Follow-up RV64 lowering/optimization pass using QEMU's RISC-V translator behavior as a reference for direct branch TB chaining and jalr semantics.

Commit pushed: db372fdb (Optimize RV64 block chaining and hot registers)

Changes:

• pinned RV64 ra, sp, gp, tp, t0-t2, s0, s1 in ARM64 callee-saved host registers
• synchronized pinned registers around C helper slow paths
• wired direct jal and conditional branches into asbestos block chaining with RV64 tagged target addresses
• added a return-cache fast path for normal ret (jalr x0, ra, 0) using the existing fiber return-cache frame storage
• added targeted C benchmarks: branch_chaining, call_return, hot_regs

Validation:

• ninja -C build-bench-rv64 ish rvinterp-selftest
• ./build-bench-rv64/rvinterp-selftest
• RV64 shell smoke through Alpine rootfs
• full RUNS=3 WARMUPS=1 ./benchmarks/hot_paths/compare.sh

Full benchmark result, best of 3:

benchmark	i386 best s	rv64 best s	rv64 vs i386
branch_chaining	0.026194	0.055644	2.124x
call_return	0.007880	0.014805	1.879x
file_io	0.008651	0.012534	1.449x
fs_metadata	1.110000	0.870000	0.784x
gzip_payload	0.920000	0.590000	0.641x
hot_regs	0.141409	0.107107	0.757x
mandelbrot	0.044068	0.046357	1.052x
pipe_throughput	0.010004	0.010406	1.040x
prime_sieve	0.016679	0.026040	1.561x
shell_control	1.850000	0.980000	0.530x
shell_pipeline	0.590000	0.340000	0.576x

RV64 still skips Python and bash benchmarks because the current RV64 rootfs does not include /usr/bin/python3 or /bin/bash.

[siraben]

Follow-up fix for the simulator apk add curl crash path.

Commit pushed: 3cbbde5e (Implement RV64 FP conversions and hwprobe).

What changed:

• Implemented RV64 FP conversion cases in the interpreter and JIT helper paths, including the fcvt.s.lu instruction that showed up as SIGILL in Alpine Python (d3 77 37 d0, OP-FP funct5=0x1a, rs2=3).
• Added a conservative riscv_hwprobe syscall implementation for syscall 258. It returns success while marking requested feature keys unsupported, so userspace does not infer unsupported optimized ISA paths from a raw stub.

Validation:

• ninja -C build-bench-rv64 ish rvinterp-selftest
• ./build-bench-rv64/rvinterp-selftest
• Injected Alpine RV64 rootfs with bash and python3 packages: python3 -S, normal python3, and bash all run successfully.
• Full benchmark matrix with Python and bash present on RV64 completed.

Full benchmark result, best of 3 with Python/bash enabled in the RV64 throwaway rootfs:

benchmark	i386 best s	rv64 best s	rv64 vs i386
bash_control	37.560000	27.320000	0.727x
branch_chaining	0.025849	0.053763	2.080x
call_return	0.007694	0.014739	1.916x
file_io	0.008888	0.011203	1.260x
fs_metadata	1.080000	0.790000	0.731x
gzip_payload	0.930000	0.580000	0.624x
hot_regs	0.145530	0.104777	0.720x
mandelbrot	0.043166	0.045507	1.054x
pipe_throughput	0.010212	0.010109	0.990x
prime_sieve	0.016681	0.026182	1.570x
python_compute	9.040000	0.410000	0.045x
python_imports	2.020000	0.710000	0.351x
python_startup	0.170000	0.150000	0.882x
shell_control	1.840000	0.960000	0.522x
shell_pipeline	0.600000	0.370000	0.617x