Mistake in generation of Egg expressions

[JonathanDLTran]

The output after Egg optimization produces a sequence of Egg expressions nodes that do not correspond to any intended pattern in the conversion back to LLVM IR.

Input:
RecExpr { nodes: [Symbol("no-array-name7"), Symbol("1,"), Get([0, 1]), Add([2, 0]), Symbol("no-array-name8"), Symbol("2,"), Get([4, 5]), Symbol("no-temp-name6"), Mul([6, 7]), Symbol("no-array-name9"), Symbol("2,"), Get([9, 10]), Add([11, 8]), Symbol("no-array-name10"), Symbol("3,"), Get([13, 14]), Symbol("no-array-name11"), Symbol("2,"), Get([16, 17]), Mul([15, 18]), Num(0), Vec([3, 12, 19, 20])] }

Output:
[Symbol("no-array-name7"), Symbol("1,"), Get([0, 1]), Symbol("no-array-name9"), Symbol("2,"), Get([3, 4]), Num(0), Num(0), Vec([2, 5, 6, 7]), Num(1), Symbol("no-array-name8"), Symbol("2,"), Get([10, 11]), Symbol("no-array-name10"), Symbol("3,"), Get([13, 14]), Num(1), Vec([9, 12, 15, 16]), Symbol("no-array-name7"), Symbol("no-temp-name6"), Symbol("no-array-name11"), Symbol("2,"), Get([20, 21]), Num(0), Vec([18, 19, 22, 23]), VecMAC([8, 17, 24])]

Error:
thread '<unnamed>' panicked at '"no-array-name7"', src/lib.rs:398:15
Symbol("no-array-name7") should always be paired with an array offset Symbol Node and a Get Node. In this situation, the Symbol("no-array-name7") should not exist by itself.

[sampsyo]

Interesting… maybe I'm just being slow, but I guess I expected that these would be legible as S-expressions instead of this flat symbol list. (Put differently, I'm not exactly sure how to read this RecExpr dump as an expression.)

Is there a chance that this is the output of just formatting a RecExpr directly (i.e., .to_string()), but calling the pretty() method on RecExpr would yield a more legible S-expression?

[JonathanDLTran]

I should have started off by printing the S-expressions, and I realized that the mistake arises in how I convert the LLVM IR to Egg expressions, so I have changed the title of the issue to be less misleading.

The new S-expression input is:
(VecMAC (Vec no-temp-name4 (Get no-array-name4 1,) 0 0) (Vec 1 no-temp-name5 (Get no-array-name5 1,) 1) (Vec no-temp-name4 (Get no-array-name3 1,) (Get no-array-name6 3,) 0))
and the new S-expression output is
(VecMAC (Vec (Get no-array-name7 1,) (Get no-array-name9 2,) 0 0) (Vec 1 (Get no-array-name8 2,) (Get no-array-name10 3,) 1) (Vec no-array-name7 no-temp-name6 (Get no-array-name11 2,) 0))

The corrected reason for the problem is that no-array-7 appears by itself as a node in the input S-expression.

[sampsyo]

Huh, interesting… this might be dumb, but should I be surprised that the symbol no-array-name7 only appears in the output? It seems weird that egg optimization would "invent" a new array instead of just rearranging accesses to existing arrays…

[JonathanDLTran]

This is completely my mistake; I made a copy and paste error, because the correct initial input should be:
(Vec (+ (Get no-array-name7 1,) no-array-name7) (+ (Get no-array-name9 2,) (* (Get no-array-name8 2,) no-temp-name6)) (* (Get no-array-name10 3,) (Get no-array-name11 2,)) 0)

[sampsyo]

Got it. I'm still not 100% clear on why the conclusion is that the LLVM->egg translation is wrong... (or did I misunderstand)? Can you point to the subexpression that was generated incorrectly?

[avanhatt]

Now resolved.