Prevent InvalidBitWidth Errors

[pvelesko]

During optimization of loops, LLVM generates non-standard integer types such as i33 or i56.

Tried to find which LLVM pass does this but it seems to happen quite early.

1. Upgrade non-standard sizes
2. Remove redundant trunc instructions

[pjaaskel]

Related: KhronosGroup/SPIRV-LLVM-Translator#1685

[linehill]

Please, run git-clang-format on the changes.

[pvelesko]

@linehill I spent a fair amount of looking already but to confirm: there is no way to instruct LLVM to NOT generate illegal integer types?

[linehill]

@linehill I spent a fair amount of looking already but to confirm: there is no way to instruct LLVM to NOT generate illegal integer types?

Did you look into the pass responsible for spawning the non-standard integers? It might have a command line option for it.

[pvelesko]

Did you look into the pass responsible for spawning the non-standard integers? It might have a command line option for it.

I did but I can't find my notes on which one it was but iirc it was quite early and deals with general loop optimizations so disabling is not desired.

[linehill]

I did but I can't find my notes on which one it was but iirc it was quite early and deals with general loop optimizations so disabling is not desired.

When you find it again, check if it has an option that could stop it to generate non-standard integers - not to disable the whole pass.

[pvelesko]

I did but I can't find my notes on which one it was but iirc it was quite early and deals with general loop optimizations so disabling is not desired.

When you find it again, check if it has an option that could stop it to generate non-standard integers - not to disable the whole pass.

It's IndVarSimplifyPass - none of the options I tried seem to help

[linehill]

Is there a reason to iterate the container in reverse order?

Suggested change

	for (auto It = Replacements.rbegin(); It != Replacements.rend();
	++It) {
	It->Old->eraseFromParent();
	}
	for (auto &R : Replacements)
	R.Old->eraseFromParent();

[pvelesko]

To preserve iterators traversing instructions from leaf to root

[linehill]

But why? I'm thinking you encountered restriction on the Instruction::eraseFromParent() that does not allow instructions with users to be erased and the reverse iteration order works-around it. I guess that explains it.

[linehill]

LGTM, thanks.