MetaphIR Matcher

https://xuhongxu.com/metaphir/

Check How Often an IR Pattern Occurs in Real Code — In Your Browser

When you write or review an InstCombine fold, one question is whether the pattern it matches shows up in real programs at all. Answering it usually means writing a matcher, building LLVM, and running it over a corpus of IR.

MetaphIR is a small pattern language for LLVM IR, and the MetaphIR Matcher runs it in the browser against an indexed snapshot of real-world IR from llvm-opt-benchmark-nightly.

Write a pattern, hit Match, and see every function in the corpus that contains it.

Announcement on the LLVM Discourse: A web-based matcher for checking how often an IR pattern occurs in real code

The pattern language

A pattern is LLVM IR with meta variables (?x) and constraints (lines starting with ;;). It matches functions that contain these instructions.

; (X u/ 2) + (X u% 2), in either order
%div = udiv i?n ?x, 2
%rem = urem i?n ?x, 2
%sum = add:c i?n %div, %rem
;; ugt(?n, 1) && isKnownNonZero(?x)

• Meta variables — ?x binds a value, a type, an integer width, a predicate, a flag, a callee, and so on. A meta variable used twice binds one thing: mul i32 ?x, ?x squares a value.
• Meta flags — :c matches the operands in either order, :s requires a single use, and flags on operands such as :knownPowerOfTwo say what the operand must be.
• Constraints — conditions on the bindings, such as noOverflow("umul", ?y, ?z).

For example, the fold (X udiv Y) udiv Z --> X udiv (Y * Z) from #214541 is written in InstCombine as

match(Op0, m_OneUse(m_UDiv(m_Value(X), m_Value(Y)))) &&
    willNotOverflowUnsignedMul(Y, Op1, I)

and in MetaphIR as

%d = udiv i?n ?x, ?y
%r = udiv i?n %d:s, ?z
;; noOverflow("umul", ?y, ?z)

The Help button on the site walks through the syntax and semantics with examples, and the example dropdown loads patterns of folds from recent LLVM pull requests.

What you get

• Matches — every matching function, with its module and the bindings of the meta variables. You can search them by module, function, or binding.
• IR view — click a match to see its function with the matched instructions highlighted.
• Statistics — how many functions matched out of how many were scanned. A broad search pauses every 100 matches, so you can stop early or continue to the end.
• Share — a link that opens the matcher with your pattern.

Notes

• The index is downloaded on demand: about 100–600 MB for a broad query. The cache is incremental, so the first queries are slow and later ones are faster.
• MetaphIR is an early research prototype, and its syntax is still changing. Feedback is welcome on the Discourse thread.