]> git.ipfire.org Git - thirdparty/gcc.git/commit
middle-end: [PR middle-end/116926] Allow widening optabs for vec-mode -> scalar-mode
authorVictor Do Nascimento <victor.donascimento@arm.com>
Thu, 10 Oct 2024 11:55:04 +0000 (12:55 +0100)
committerVictor Do Nascimento <victor.donascimento@arm.com>
Mon, 14 Oct 2024 09:11:13 +0000 (10:11 +0100)
commita6f4404689f12fa169540d94c0d06532c7152c2f
treebf7869c13bebf22bb873cc60d9c4cb64df655869
parent50e7c51b0a0e9dc1d93f829016ae743b4f2e5070
middle-end: [PR middle-end/116926] Allow widening optabs for vec-mode -> scalar-mode

The recent refactoring of the dot_prod optab to convert-type exposed a
limitation in how `find_widening_optab_handler_and_mode' is currently
implemented, owing to the fact that, while the function expects the

  GET_MODE_CLASS (from_mode) == GET_MODE_CLASS (to_mode)

condition to hold, the c6x backend implements a dot product from V2HI
to SI, which triggers an ICE.

Consequently, this patch adds some logic to allow widening optabs
which accumulate vector elements to a single scalar.

gcc/ChangeLog:

PR middle-end/116926
* optabs-query.cc (find_widening_optab_handler_and_mode): Add
handling of vector -> scalar optab handling.
gcc/optabs-query.cc