From 0dcd3513a633479ce228f24826db4f09e4c0fc8b Mon Sep 17 00:00:00 2001 From: Cliff Click Date: Mon, 16 Nov 2026 14:25:14 -0800 Subject: Phi fix Fix Phi/Add refactoring peephole to canonicalize regardless of peephole order. Treat class fcns as having `self` be the class pointer --- chapter05/README.md | 15 +++++++++++++++ 1 file changed, 15 insertions(+) (limited to 'chapter05/README.md') diff --git a/chapter05/README.md b/chapter05/README.md index 40886af2..0955f36a 100644 --- a/chapter05/README.md +++ b/chapter05/README.md @@ -413,6 +413,21 @@ return a; +### Factoring after strength reduction + +Peephole order should not hide the same arithmetic from Phi factoring. +`Phi(x+x,y+x)` can become `Phi(x,y)+x`, but an earlier peephole may already have +replaced `x+x` with `x*2`. The Phi matcher therefore views `x*2` as two copies of +`x` while factoring mixed Add/Mul arms. It does not expand the graph into an Add +that would immediately fold back to a multiply. If all arms are multiplies, +ordinary multiply factoring still applies. + +[Chapter05Test](src/test/java/com/seaofnodes/simple/Chapter05Test.java) +checks the factored operands before and after strength reduction and with either +arm order. Later chapters extend this same view to `x<<1` when shifts arrive. +The rule stays local to doubling; arbitrary coefficients would require creating +another multiplication, with a less clear benefit. + [^1]: Click, C. (1995). Combining Analyses, Combining Optimizations, 132. -- cgit v1.2.3