search for: speculateselectinstloads

...tually correctly use it like that today. If there aren't many uses, we could also add a (DILocation *) argument to SetInsertPoint to force every client to think about how to properly set the location in their use-case. There's at least one correct use of the current SetInsertPoint() API in speculateSelectInstLoads() :). As you predicted, it expands an existing instruction into a sequence. I'm not sure how many correct uses of the API we have. There are over 300 uses in just {llvm,clang}/lib. I think that prevents us from adding a mandatory argument to SetInsertPoint(): we wouldn't be able to promptl...

...se it like that today. If there aren't many uses, we could also add a (DILocation *) argument to SetInsertPoint to force every client to think about how to properly set the location in their use-case. >> >> There's at least one correct use of the current SetInsertPoint() API in speculateSelectInstLoads() :). As you predicted, it expands an existing instruction into a sequence. >> >> I'm not sure how many correct uses of the API we have. There are over 300 uses in just {llvm,clang}/lib. I think that prevents us from adding a mandatory argument to SetInsertPoint(): we wouldn't...

Hi everybody, I'm investigating some correctness issues with debug line table information in optimized code. I've noticed something problematic in IRBuilder. Setting the insertion point of an IRBuilder sets the builder’s current debug loc to the one attached to the insertion point. IMO this isn't always a correct or convenient thing to do, and it shouldn't be the default behavior.