riscv_dis_opts_batch_1

Disassembler: Core improvements and optimizations (batch 1)

• Branch: riscv-dis-opts-batch-1
• Tracking PR: #28 (view Pull Request and Diff)
• Mailing List:
  • Not yet

Conflicts With

• Disassembler: Fix address printer
  It also touches the core disassembler.
• Pending: Disassembler: Implement "generic subsets" (PATCH v1 only)
  It also changes the core disassembler. Note that the latest version depends on this patchset.
• Pending (RFC): Disassembler: Check shift amount against XLEN (Idea 1)
  It also changes the core disassembler.

Based On

• Disassembler: Use faster hash table
• Disassembler: Minor optimizations (batch 1)
• Disassembler: Cache instruction class support
• Disassembler: Use xlen on ADDIW address sequence

Feature Summary

To improve the core disassembler (both for performance and feature), this patchset now implements several optimizations for further performance on objdump and GDB.

It will be a prerequisite of upcoming core disassembler changes:

• Disassembler: RV32E support
• Disassembler: Add overridable priv-spec and arch options
• Disassembler: Implement "generic subsets"

Changes in this Patchset:

• Various minor fixes (originally in... (1) (2))
• Split match/print steps on the disassembler (originally in "generic subsets")
• One time CSR hash table initialization (previously in minor optimizations batch 1)
• Minimize disassembler state initialization (general/arch-related) (previously in minor optimizations batch 1)
• Use faster hash table on disassembling
• Cache instruction support

Benchmarking System

This benchmark is performed on:

• Ubuntu 22.04 LTS
• AMD Ryzen 5 PRO 5650G processor

In the parallel run, I ran 6 parallel jobs with -j6 (corresponding 6 cores; although the processor has 12 hardware threads, -j12 just slowed the benchmark).

Aggregate Performance Improvements

On disassembling linked RISC-V ELF programs using objdump, performance improvements achieved by this patchset is usually about 25-33%. Isn't it nice to have?

Note that a few percent may be compensated by future non-performance changes. In fact, some performance changes for objdump -d hurts GDB performance a bit because of extra initialization (which is caused per instruction on the current GDB design).

This is relative to the latest master (commit 8e883b5e112).

objdump -d (ELF)

Program	Improvements	Notes
Busybox 1.35.1 (RV64GC)	29.0-32.8%
OpenSBI 1.1 (generic fw_*.elf)	42.3-42.5%
Linux kernel 5.18 (vmlinux)	27.3-29.2%
Linux kernel 5.18 (vmlinux.o)	2.5-14.8%	Not finally linked
glibc (libc.so.6)	26.7-30.5%

Due to extra overhead handling many debugging information and symbols, disassembling *.o will be relatively slower.

objdump -d (ELF-based archive)

Program	Improvements
glibc (libc.a)	6.5-7.2%
newlib (libc.a)	7.0-10.6%

Due to extra archive file handling, disassembler performance improvements are not large enough.

objdump -D (binary)

Program	Improvements
Linux kernel 5.18 (vmlinux)	73.3-101.1%
Random files (/dev/urandom)	91.9-113.9%
1M (1048576) CSR instructions	374.1%

gdb: disas of near all code region

Program	Improvements
Linux kernel 5.18 (vmlinux) with debug info	30.8%
Linux kernel 5.18 (vmlinux) without debug info	91.6%
OpenSBI 1.1 (generic fw_*.elf)	80.0-80.5%
1M (1048576) CSR instructions (ELF)	220.1%

Batch: objdump -d on Linux distribution

Serial Run: All /usr/bin Programs

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	563	25.2%
Debian unstable (as of 2022-07-20)	269	25.4%

Parallel Run: Top 100 in Size (including data-only ELFs)

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	100	13.0%
Debian unstable (as of 2022-07-20)	100	5.0%

Parallel Run: All (including data-only ELFs)

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	7666	13.4%
Debian unstable (as of 2022-07-20)	946	6.8%

Batch: objdump -D (as binary) on Linux distribution

Serial Run: All /usr/bin Programs

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	563	50.0%
Debian unstable (as of 2022-07-20)	269	42.1%

Parallel Run: Top 100 in Size (including data-only ELFs)

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	100	70.5%
Debian unstable (as of 2022-07-20)	100	50.0%

Parallel Run: All (including data-only ELFs)

System	N	Improvements
Ubuntu 22.04 LTS (image for HiFive Unmatched)	7666	56.2%
Debian unstable (as of 2022-07-20)	946	43.6%