[bug] [ir] Fix Block::replace_with segment faults when stmts.size() == 0

[archibate]

Related issue = close #1071

[Click here for the format server]

Seems replace_usage = true not working when new_statements.size() == 0:

taichi/taichi/ir/ir.cpp

Lines 760 to 761 in c193e4b

	if (replace_usages)
	old_statement->replace_with(new_statements.back().get());

... where new_statements.back() == nullptr.

[xumingkuan]

Looks good, but I think we'd better modify replace_with rather than void visit(FrontendEvalStmt *stmt) to prevent other potential bugs. I suggest changing

if (replace_usages)
  old_statement->replace_with(new_statements.back().get());

to

if (replace_usages && !new_statements.empty())
  old_statement->replace_with(new_statements.back().get());

BTW I think it weird when new_statements.size() != 1 && replace_usages == true. Do we need this design in any IR passes?

[archibate]

Hello? @yuanming-hu I know you're busy preparing GAMES201, so can I merge this omo once @xumingkuan approved? There are now 14 PRs waiting for merge, I'll merge some with trival changeset omo with 1 approve from taichi-dev write-members.

[yuanming-hu]

Hello? @yuanming-hu I know you're busy preparing GAMES201, so can I merge this omo once @xumingkuan approved? There are now 14 PRs waiting for merge, I'll merge some with trival changeset omo with 1 approve from taichi-dev write-members.

Sure, having Mingkuan's approval is good enough. Btw, please capitalize the first letter of your PR letter and avoid confusing abbreviations. Thank you!

[xumingkuan]

Cool!

[xumingkuan]

(Actually we can add VecStatement::empty() for convenience.)