Optimize interpreter and Opcode::FromCode
#910
Conversation
`Opcode::FromCode` calculated the opcode given a prefix/code pair by using lower_bound over the list of all `OpcodeInfo`s. This was happening for every instruction, which is incredibly slow. Since the interpreter's format is internal only, we can use any encoding we want, so it's simpler and faster to use the `Opcode::Enum` directly without calling `Opcode::FromCode`. `Opcode::FromCode` is also used when reading a binary file, so it should be optimized anyway. Instead of using the `infos_` table, which is indexed by the opcode's `enum_` value, we create a new statically-defined table that maps from prefix-code pair to its enum value. Unfortunately, this can't be done easily in C++ because it does not currently support designated array initializers, so this table is created in a C file instead, `opcode-code-table.c`.
src/opcode-code-table.c
Outdated
| Invalid, | ||
| } WabtOpcodeEnum; | ||
|
|
||
| _Static_assert(Invalid <= 65536, "Too many opcodes"); |
There was a problem hiding this comment.
Maybe static const size_t kNumOpcodes = 65536; in opcode.h
Especially because the extern declaration there needs to match
There was a problem hiding this comment.
It's a bit tricky since I can't include opcode.h here (since it's C++). So I'd have to make a C header and include it there. I was trying to avoid adding a new file, but there's no reason to do so, really.
src/opcode-code-table.c
Outdated
| uint32_t WabtOpcodeCodeTable[65536] = { | ||
| #define WABT_OPCODE(rtype, type1, type2, type3, mem_size, prefix, code, Name, \ | ||
| text) \ | ||
| [(prefix << 8) + code] = Name, |
There was a problem hiding this comment.
This file has to be compiled as C for this to work, right?
Not gonna lie, this is really neat. Meaning both "nifty" and "clean"
There was a problem hiding this comment.
Yep, it's a C99-only feature (until C++20, I think). It works, but I dunno, I thought about doing this before and thought it was kind of gross. :-) But the alternatives are not great either, like building the table at runtime, generating it via script, or writing it out manually.
There was a problem hiding this comment.
Maybe a constexpr static initializer? Should be doable even with C++11 constexprs. Not sure if it's less-gross though
Opcode::FromCodecalculated the opcode given a prefix/code pair byusing lower_bound over the list of all
OpcodeInfos. This was happeningfor every instruction, which is incredibly slow.
Since the interpreter's format is internal only, we can use any encoding
we want, so it's simpler and faster to use the
Opcode::Enumdirectlywithout calling
Opcode::FromCode.Opcode::FromCodeis also used when reading a binary file, so it shouldbe optimized anyway. Instead of using the
infos_table, which isindexed by the opcode's
enum_value, we create a newstatically-defined table that maps from prefix-code pair to its enum
value.
Unfortunately, this can't be done easily in C++ because it does not
currently support designated array initializers, so this table is
created in a C file instead,
opcode-code-table.c.