[PATCH] D25866: [Sema] Support implicit scalar to vector conversions

[Simon Dardis via Phabricator via cfe-commits]

sdardis marked 3 inline comments as done.
sdardis added a comment.

Thanks for getting back to this. I've traced the appearance of the ext_vector type to a piece of code that only produces ext-vector types for comparisons. I'm presuming that's wrong when clang is producing vectors implicitly in the non-OpenCL case. I'll update the diff once I've changed that.



================
Comment at: test/Sema/vector-gcc-compat.c:61
+  //        match.
+  v2i64_r = v2i64_a == 1; // expected-warning {{incompatible vector types assigning to 'v2i64' (vector of 2 'long long' values) from 'long __attribute__((ext_vector_type(2)))' (vector of 2 'long' values)}}
+  v2i64_r = v2i64_a != 1; // expected-warning {{incompatible vector types assigning to 'v2i64' (vector of 2 'long long' values) from 'long __attribute__((ext_vector_type(2)))' (vector of 2 'long' values)}}
----------------
bruno wrote:
> Can you double check where 'long __attribute__((ext_vector_type(2)))' comes from?
> 
> We have regressed in the past year in the way ext-vector interacts with non-ext-vectors, and I don't wanna make it worse until we actually have time to fix that; there's a lot of code out there relying on bitcasts between ext-vectors and non-ext-vectors to bridge between intrinsics headers and ext-vector code.
Sema::CheckVectorCompareOperands calls Sema::GetSignedVectorType which only returns ext-vector types. I presume that is now incorrect if we're producing vectors from literal scalars in the non OpenCL case.


https://reviews.llvm.org/D25866