Consider the test added in this PR: it is not enough to insert the replacement alloc before any of the access ops, you will hit operand #0 does not dominate this use or end up with an alloc in the for loop. So this PR changes where the alloc is inserted : it is now inserted at the very start of the core's block.
All other changes are here are tiny and non-functional (propagate op location for better diagnostics, don't do map look up twice when only once is needed, use assert instead of error if it shouldn't be reachable)
Consider the test added in this PR: it is not enough to insert the replacement alloc before any of the access ops, you will hit
operand #0 does not dominate this useor end up with an alloc in the for loop. So this PR changes where the alloc is inserted : it is now inserted at the very start of the core's block.All other changes are here are tiny and non-functional (propagate op location for better diagnostics, don't do map look up twice when only once is needed, use assert instead of error if it shouldn't be reachable)