[all-commits] [llvm/llvm-project] e496d3: [InstCombine] Drop one-use check in matchesSquareS...

[konoongg via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: e496d3c1f3e60b27d8b016c0b36ecb89653ae38f
      https://github.com/llvm/llvm-project/commit/e496d3c1f3e60b27d8b016c0b36ecb89653ae38f
  Author: konoongg <120199924+konoongg at users.noreply.github.com>
  Date:   2026-10-01 (Thu, 01 Oct 2026)

  Changed paths:
    M llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
    M llvm/test/Transforms/InstCombine/add.ll
    M llvm/test/Transforms/InstCombine/fadd.ll

  Log Message:
  -----------
  [InstCombine] Drop one-use check in matchesSquareSum for (a*a) (#228020)

The pattern `(a * a) + (((a * 2) + b) * b)` is equivalent to `(a + b) *
(a + b)`. The current implementation requires the `a * a` subexpression
to have a single use, which prevents the fold when `a * a` is used
elsewhere. This check is unnecessary: we can still perform the fold and
keep the original `a * a` instruction for other uses.

For example, with an extra use of `a * a`:

  %1 = mul i32 %a, %a
  %2 = mul i32 %a, 2
  %3 = add i32 %2, %b
  %4 = mul i32 %3, %b
  %5 = add i32 %4, %1
  %6 = mul i32 %1, %1

After this change, InstCombine produces:

  %1 = add i32 %a, %b
  %2 = mul i32 %1, %1
  %3 = mul i32 %a, %a
  %4 = mul i32 %3, %3

This reduces the instruction count from 6 to 4, even though `a * a` is
now computed both explicitly (for the extra use) and implicitly as part
of `(a + b) * (a + b)`.

The same applies to the floating-point (fadd) variant. Both are
updated together.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications