Skip to content

AVX512FP16: Support vector shuffle builtins

Hongyu Wang requested to merge users/intel/hongyuw1/fp16_shuffle into users/intel/fp16_suffle/master

Support HFmode vector shuffle by creating HImode subreg when expanding permutation expr.

gcc/ChangeLog:

* config/i386/i386-expand.c (ix86_expand_vec_perm): Convert
HFmode input operand to HImode.
(ix86_vectorize_vec_perm_const): Likewise.
(ix86_expand_vector_init): Allow HFmode for one_operand_shuffle.
* config/i386/sse.md (*avx512bw_permvar_truncv16siv16hi_1_hf):
New define_insn.
(*avx512f_permvar_truncv8siv8hi_1_hf):
Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/i386/avx512fp16-builtin_shuffle-1.c: New test.
* gcc.target/i386/avx512fp16-pr101846.c: Ditto.
* gcc.target/i386/avx512fp16-pr94680.c: Ditto.

Merge request reports