This patch introduces instruction patterns for complex number operations 
in the GCN machine description. These patterns are cmul, cmul_conj, 
vec_addsub, vec_fmaddsub, vec_fmsubadd, cadd90, cadd270, cmla and cmls 
(cmla_conj and cmls_conj were not found to be favorable to implement). 
As a side effect of adding cmls, I also added fms patterns corresponding 
to the existing fma patterns. Tested on CDNA2 GFX90a.

OK to commit?


gcc/ChangeLog:

	* config/gcn/gcn-protos.h (gcn_expand_dpp_swap_pairs_insn)
         (gcn_expand_dpp_distribute_even_insn)
         (gcn_expand_dpp_distribute_odd_insn): Declare.
         * config/gcn/gcn-valu.md (@dpp_swap_pairs<mode>)
         (@dpp_distribute_even<mode>, @dpp_distribute_odd<mode>)
         (cmul<conj_op><mode>3, cml<addsub_as><mode>4, vec_addsub<mode>3)
         (cadd<rot><mode>3, vec_fmaddsub<mode>4, vec_fmsubadd<mode>4)
         (fms<mode>4<exec>, fms<mode>4_negop2<exec>, fms<mode>4)
         (fms<mode>4_negop2): New patterns.
         * config/gcn/gcn.cc (gcn_expand_dpp_swap_pairs_insn)
         (gcn_expand_dpp_distribute_even_insn)
         (gcn_expand_dpp_distribute_odd_insn): New functions.
         * config/gcn/gcn.md: Add entries to unspec enum.

gcc/testsuite/ChangeLog:

	* gcc.target/gcn/complex.c: New test.