[Opt] Improve the binary ops simplify pass

[xumingkuan]

Related issue = #1226 #656

• Do this optimization when fast_math=False and the operands are (return value is) of integral types.
• Add (a - b) + c -> a - (b - c), (a - b) - c -> a - (b + c), (a / b) * c -> a / (b / c), (a / b) / c -> a / (b * c).
• Disable (a / b) * c -> a / (b / c), (a * b) / c -> a * (b / c) when the data type is integral.
• Let this pass return true when two operands are swapped but modifier.modify_ir() returns false.

Benchmark (#stmts): unchanged.

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1646 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1646   +/-   ##
=======================================
  Coverage   44.40%   44.40%           
=======================================
  Files          41       41           
  Lines        5876     5876           
  Branches     1022     1022           
=======================================
  Hits         2609     2609           
  Misses       3118     3118           
  Partials      149      149           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 29071b3...39488c5. Read the comment docs.

[TH3CHARLie]

It's a nice improvement to the IR and optimization system.

• Add (a - b) + c -> a - (b - c), (a - b) - c -> a - (b + c), (a / b) * c -> a / (b / c), (a / b) / c -> a / (b * c).

For the (a / b) * c -> a / (b / c) case, it can not be safely swapped due to the integer division roundoff. Consider a = 10, b = 3 and c = 3, the former produces 9 while the swapped one produces 10.

[xumingkuan]

It's a nice improvement to the IR and optimization system.

• Add (a - b) + c -> a - (b - c), (a - b) - c -> a - (b + c), (a / b) * c -> a / (b / c), (a / b) / c -> a / (b * c).

For the (a / b) * c -> a / (b / c) case, it can not be safely swapped due to the integer division roundoff. Consider a = 10, b = 3 and c = 3, the former produces 9 while the swapped one produces 10.

Oh yes. I disabled this optimization when the return type is integral. And IIRC our typecheck guarantees that binary operations have the same type among operands and the return value.

[TH3CHARLie]

LGTM, just update the PR description about disabling integral division.