https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82416
--- Comment #1 from Martin Jambor <jamborm at gcc dot gnu.org> --- Created attachment 42308 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42308&action=edit Fix This is a fix I have come up with so far. The key change is the simple flip from true to false in hsa_function_representation::reg_for_gimple_ssa, everything else is there to compensate the big impact that has. Basically, until the time we really output them, registers keep the type of the data that they hold and we extend them whenever we want perform on them any computation that can only be done in at least 32 bits in HSA, which is almost all of them - and then we convert them back to a shorter type if necessary. The only exception are complex numbers because combine/expand also only exist in 32bits which in theory might break like the testcase does when someone does something similar with a complex char... but hopefully gimple does not perform such optimization on complex numbers. If it turns out it does, I guess we'll refuse to generate HSA for them. Pekka, can you try if this works for you as well? I plan to commit this early next week after a tiny bit more testing. 2017-10-04 Martin Jambor <mjam...@suse.cz> PR hsa/82416 gcc/ * hsa-common.h (hsa_op_with_type): New method extend_int_to_32bit. * hsa-gen.c (hsa_extend_inttype_to_32bit): New function. (hsa_type_for_scalar_tree_type): Use it. Always force min32int for COMPLEX types. (hsa_fixup_mov_insn_type): New function. (hsa_op_with_type::get_in_type): Use it. (hsa_build_append_simple_mov): Likewise. Allow sub-32bit immediates in an assert. (hsa_op_with_type::extend_int_to_32bit): New method. (gen_hsa_insns_for_bitfield): Fixup instruction and intermediary types. Convert to dest type if necessary. (gen_hsa_insns_for_bitfield_load): Fixup load type if necessary. (reg_for_gimple_ssa): Pass false as min32int to hsa_type_for_scalar_tree_type. (gen_hsa_addr): Fixup type when creating addresable temporary. (gen_hsa_cmp_insn_from_gimple): Extend operands if necessary. (gen_hsa_unary_operation): Extend operands and convert to dest type if necessary. Call hsa_fixup_mov_insn_type. (gen_hsa_binary_operation): Changed operand types to hsa_op_with_type, extend operands and convert to dest type if necessary. (gen_hsa_insns_for_operation_assignment): Extend operands and convert to dest type if necessary. (set_output_in_type): Call hsa_fixup_mov_insn_type. Just ude dest if conversion nt necessary and size matches. (gen_hsa_insns_for_load): Call hsa_fixup_mov_insn_type, convert to dest type if necessary. (gen_hsa_insns_for_store): Call hsa_fixup_mov_insn_type. (gen_hsa_insns_for_switch_stmt): Likewise. Also extend operands if necessary. (gen_hsa_clrsb): Likewise. (gen_hsa_ffs): Likewise. (gen_hsa_divmod): Extend operands and convert to dest type if necessary. (gen_hsa_atomic_for_builtin): Change type of op to hsa_op_with_type. libgomp/ * testsuite/libgomp.hsa.c/pr82416.c: New test.
