[VM] 2nd Performance optimization

[brew0722]

No description provided.

[brew0722]

First, after the first analysis of the singlepass code, we confirmed the possibility of optimizing the code size of emit_memory_op.

And analyzed that the static memory bound check patch is also valid for singlepass.

• Minimum 27% ~ maximum 208% performance improvement compared to the highest performance before application (418txs → 532~870txs)
• 286KB reducing JIT code cache size.
• 1405KB(1.37MB) reducing CacheGen size.(related ExceptionTable)

assembly generation step

• emit (mov reg_addr_base, mem_base)
  #IF IS_DYNAMIC
  • emit (mov reg_offset_bound, mem_bound)
  • emit (add reg_offset_bound, reg_addr_base) == reg_offset_bound changed to reg_addr_bound
  • emit (mov reg_offset_target, target_addr)
  • emit (add reg_offset_target, memarg.offset)
  • emit (add reg_offset_target, memarg.value_size) == reg_offset_target changed to reg_offset_target_bound
  • emit (add reg_offset_target_bound, reg_addr_base) == reg_offset_target_bound changed to reg_addr_target_bound
  • emit (cmp reg_addr_target_bound, reg_addr_bound)
  • emit (ja label_trap)
• emit (mov reg_offset_target, target_addr)
• emit (add reg_offset_target, memarg.offset)
• emit (add reg_offset_target, mem_base) == reg_offset_target changed to reg_addr_target
  #IF CHECK_ALIGNMEJNT
  • emit (mov ...)
  • emit (and ...)
  • emit (jne label_tgrap)
• call post generation callback

As a result of analysis, 7333 dynamic emits were performed in erc-20.
So, in the case of erc-20, the bound check inline assembly occupies 7333*40bytes = 293,320Bytes (286KB).
As a result, by not executing the bound check, the amount of code generated is reduced, and the speed is increased.

[brew0722]

First, the first analysis is completed. Further analysis and actual poc will be do out later.

[brew0722]

Unlike the previous meeting, plans related to VM apply are not in a hurry, so I will proceed with this work again.

[brew0722]

init_local_remove_redurrent_repeat_mov.patch.zip

2nd patch. Assembly code that initializes the local stack is inefficient.
Generates a 1:1 mov assembly instruction per stack slot used.
For example, if the local stack range is ebp-8 to ebp-336, initializing 41 slots. Since a single mov instruction takes up 12 bytes, it takes up a total of 492 bytes.

This can reduce to code size of all stack size initialization with a fixed 24 bytes using the rep stosq command.

result, for erc20 reduced by 8KB.

[brew0722]

#48

3rd patch. We analyzed the inefficiency of ExceptionTable structure and changed it to Offset Range.
Patch result, For erc20, the module cache file was successfully reduced by 612KB.

[brew0722]

#51

4th patch. copying wasm memory data from cosmwasm_vm::read_region, copy was implemented as assign for 0..n The cost of copying is severe.

Patching to use std::copy for raw pointers, the performance range has been fixed, and maximum performance has been improved.(min64%, max 7%)

1st performance measurement result - 876 ~ 913txs
Secondary performance measurement result - 868 ~ 963txs

[brew0722]

Closes:
wasmerio/wasmer#2012
wasmerio/wasmer#2030
CosmWasm/cosmwasm#730