perf(compiler): optimize vector arithmetic and numeric loop type inference#2237
perf(compiler): optimize vector arithmetic and numeric loop type inference#2237ActualMasterOogway wants to merge 3 commits intoluau-lang:masterfrom
Conversation
- Relax commutative swap optimization in compileExprBinary to include LBC_TYPE_VECTOR. - Implement visit(AstStatFor) in TypeMapVisitor to correctly infer numeric type for loop variables. Fixes luau-lang#2235
| bytecode.emitABC(getBinaryOpArith(expr->op, /* k= */ true), target, rr, uint8_t(lc)); | ||
|
|
||
| hintTemporaryExprRegType(expr->right, rr, LBC_TYPE_NUMBER, /* instLength */ 1); | ||
| hintTemporaryExprRegType(expr->right, rr, *ty, /* instLength */ 1); |
There was a problem hiding this comment.
| hintTemporaryExprRegType(expr->right, rr, *ty, /* instLength */ 1); | |
| hintTemporaryExprRegType(expr->right, rr, LBC_TYPE_NUMBER, /* instLength */ 1); |
Hint system is used to provide hints on types that are not the expected operands of an instruction.
By saying that 'vector' is expected for an add/mul, hint will not be generated and this will cause jit performance to regress on a wrong type speculation.
For example in this case:
local function foo(t: {a: vector})
return 2 * t.a
endWith this change, the fast-path is generated for number*number, causing a fallback to be taken at runtime.
Since this was a regression not caught by existing tests, this example should be added as a regression test to IrLowering.test.cpp
vegorov-rbx
left a comment
vegorov-rbx
left a comment
There was a problem hiding this comment.
PR says that it is making an improvement with an additional optimization.
While that is true, not a single test case has been added to cover this improvement and show that it is working.
New tests should be placed in Compiler.test.cpp
vegorov-rbx
left a comment
vegorov-rbx
left a comment
There was a problem hiding this comment.
These changes have to be flagged to be accepted: https://github.com/luau-lang/luau/blob/master/CONTRIBUTING.md#feature-flags
|
Thank you for providing an improvement. With a few changes it will get to a form we can merge. |
2 participants
This PR improves the compiler's ability to optimize arithmetic expressions, specifically targeting vector operations and numeric
forloops. By improving type inference, we can now apply better optimizations in more places.Closes #2235