[PATCH] D12341: add unpredictable metadata type for control flow

[Sanjay Patel via llvm-commits]

spatel added a comment.

In http://reviews.llvm.org/D12341#236308, @kparzysz wrote:

> LGTM.  I was thinking about having something like MD_prof_hw as the top-level metadata for all HW-specific profiling data, but if any further changes are needed, they can be done later.  However, if this metadata is later extended (for example, to have target/subtarget identification attached to it), would the IR compatibility be a concern?


I think we're protected from IR compatibility problems since it's just metadata now. We can put whatever additional data is needed in those strings without breaking what's in there now. Up in Clang, a new builtin would have to be created if for example, someone wanted to specify that the hint mapped to some specific CPU.

But I think what's in the metadata now has not been decided: Hal suggested that we should not add a feature with no users. Should I remove the 0 to 100 specifier (and related docs/code comments) and make it an empty string as it was in the earlier draft?


http://reviews.llvm.org/D12341