On Thu, 19 Mar 2020, Richard Sandiford wrote: > In this PR we had a 512-bit VECTOR_TYPE whose mode is XImode > (an integer mode used for four 128-bit vectors). When trying > to expand a zero constant for it, we hit code in expand_expr_real_1 > that tries to use the associated integer type instead. The code used > type_for_mode (XImode, 1) to get this integer type. > > However, the c-family implementation of type_for_mode checks for > any registered built-in type that matches the mode and has the > right signedness. This meant that it could return a built-in > vector type when given an integer mode (particularly if, as here, > the vector type isn't supported by the current subtarget and so > TYPE_MODE != TYPE_MODE_RAW). The expand code would then cycle > endlessly trying to use this "new" type instead of the original > vector type. > > The search loop is probably too lax in other ways -- e.g. it could > return records that just happen to have the right mode -- but this > seems like a safe, incremental improvement. > > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
OK. -- Joseph S. Myers jos...@codesourcery.com