From 5326695a6d3585f5c49633813909342098d1714c Mon Sep 17 00:00:00 2001 From: Andrew Stubbs Date: Thu, 17 Jan 2019 12:33:19 +0000 Subject: [PATCH] GCN back-end code This patch contains the major part of the GCN back-end. The machine description has been broken out to avoid the mailing list size limit. The back-end contains various bits that support OpenACC and OpenMP, but the middle-end and libgomp patches are missing, as is mkoffload. I include them here because they're harmless and carving up the files seems like unnecessary effort. The remaining offload support will be posted at a later date. The gcn-run.c is a separate tool that can run a GCN program on a GPU using the ROCm drivers and HSA runtime libraries. 2019-01-17 Andrew Stubbs Kwok Cheung Yeung Julian Brown Tom de Vries Jan Hubicka Martin Jambor gcc/ * common/config/gcn/gcn-common.c: New file. * config/gcn/driver-gcn.c: New file. * config/gcn/gcn-builtins.def: New file. * config/gcn/gcn-hsa.h: New file. * config/gcn/gcn-modes.def: New file. * config/gcn/gcn-opts.h: New file. * config/gcn/gcn-passes.def: New file. * config/gcn/gcn-protos.h: New file. * config/gcn/gcn-run.c: New file. * config/gcn/gcn-tree.c: New file. * config/gcn/gcn.c: New file. * config/gcn/gcn.h: New file. * config/gcn/gcn.opt: New file. * config/gcn/t-gcn-hsa: New file. Co-Authored-By: Jan Hubicka Co-Authored-By: Julian Brown Co-Authored-By: Kwok Cheung Yeung Co-Authored-By: Martin Jambor Co-Authored-By: Tom de Vries From-SVN: r268023 --- gcc/ChangeLog | 22 + gcc/common/config/gcn/gcn-common.c | 38 + gcc/config/gcn/driver-gcn.c | 32 + gcc/config/gcn/gcn-builtins.def | 116 + gcc/config/gcn/gcn-hsa.h | 115 + gcc/config/gcn/gcn-modes.def | 41 + gcc/config/gcn/gcn-opts.h | 36 + gcc/config/gcn/gcn-passes.def | 19 + gcc/config/gcn/gcn-protos.h | 146 + gcc/config/gcn/gcn-run.c | 850 ++++ gcc/config/gcn/gcn-tree.c | 721 ++++ gcc/config/gcn/gcn.c | 6123 ++++++++++++++++++++++++++++ gcc/config/gcn/gcn.h | 662 +++ gcc/config/gcn/gcn.opt | 78 + gcc/config/gcn/t-gcn-hsa | 52 + 15 files changed, 9051 insertions(+) create mode 100644 gcc/common/config/gcn/gcn-common.c create mode 100644 gcc/config/gcn/driver-gcn.c create mode 100644 gcc/config/gcn/gcn-builtins.def create mode 100644 gcc/config/gcn/gcn-hsa.h create mode 100644 gcc/config/gcn/gcn-modes.def create mode 100644 gcc/config/gcn/gcn-opts.h create mode 100644 gcc/config/gcn/gcn-passes.def create mode 100644 gcc/config/gcn/gcn-protos.h create mode 100644 gcc/config/gcn/gcn-run.c create mode 100644 gcc/config/gcn/gcn-tree.c create mode 100644 gcc/config/gcn/gcn.c create mode 100644 gcc/config/gcn/gcn.h create mode 100644 gcc/config/gcn/gcn.opt create mode 100644 gcc/config/gcn/t-gcn-hsa diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 12489cdf838..2f1a36f64ef 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,25 @@ +2019-01-17 Andrew Stubbs + Kwok Cheung Yeung + Julian Brown + Tom de Vries + Jan Hubicka + Martin Jambor + + * common/config/gcn/gcn-common.c: New file. + * config/gcn/driver-gcn.c: New file. + * config/gcn/gcn-builtins.def: New file. + * config/gcn/gcn-hsa.h: New file. + * config/gcn/gcn-modes.def: New file. + * config/gcn/gcn-opts.h: New file. + * config/gcn/gcn-passes.def: New file. + * config/gcn/gcn-protos.h: New file. + * config/gcn/gcn-run.c: New file. + * config/gcn/gcn-tree.c: New file. + * config/gcn/gcn.c: New file. + * config/gcn/gcn.h: New file. + * config/gcn/gcn.opt: New file. + * config/gcn/t-gcn-hsa: New file. + 2019-01-17 Andrew Stubbs Kwok Cheung Yeung Julian Brown diff --git a/gcc/common/config/gcn/gcn-common.c b/gcc/common/config/gcn/gcn-common.c new file mode 100644 index 00000000000..f34953dc04b --- /dev/null +++ b/gcc/common/config/gcn/gcn-common.c @@ -0,0 +1,38 @@ +/* Common hooks for GCN + Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +#include "config.h" +#include "system.h" +#include "coretypes.h" +#include "tm.h" +#include "common/common-target.h" +#include "common/common-target-def.h" +#include "opts.h" +#include "flags.h" +#include "params.h" + +/* Set default optimization options. */ +static const struct default_options gcn_option_optimization_table[] = + { + { OPT_LEVELS_1_PLUS, OPT_fomit_frame_pointer, NULL, 1 }, + { OPT_LEVELS_NONE, 0, NULL, 0 } + }; + +#undef TARGET_OPTION_OPTIMIZATION_TABLE +#define TARGET_OPTION_OPTIMIZATION_TABLE gcn_option_optimization_table + +struct gcc_targetm_common targetm_common = TARGETM_COMMON_INITIALIZER; diff --git a/gcc/config/gcn/driver-gcn.c b/gcc/config/gcn/driver-gcn.c new file mode 100644 index 00000000000..a5181bc3995 --- /dev/null +++ b/gcc/config/gcn/driver-gcn.c @@ -0,0 +1,32 @@ +/* Subroutines for the gcc driver. + Copyright (C) 2018-2019 Free Software Foundation, Inc. + +This file is part of GCC. + +GCC is free software; you can redistribute it and/or modify +it under the terms of the GNU General Public License as published by +the Free Software Foundation; either version 3, or (at your option) +any later version. + +GCC is distributed in the hope that it will be useful, +but WITHOUT ANY WARRANTY; without even the implied warranty of +MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +GNU General Public License for more details. + +You should have received a copy of the GNU General Public License +along with GCC; see the file COPYING3. If not see +. */ + +#include "config.h" +#include "system.h" +#include "coretypes.h" +#include "tm.h" + +const char * +last_arg_spec_function (int argc, const char **argv) +{ + if (argc == 0) + return NULL; + + return argv[argc-1]; +} diff --git a/gcc/config/gcn/gcn-builtins.def b/gcc/config/gcn/gcn-builtins.def new file mode 100644 index 00000000000..47897607abf --- /dev/null +++ b/gcc/config/gcn/gcn-builtins.def @@ -0,0 +1,116 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +/* The first argument to these macros is the return type of the builtin, + the rest are arguments of the builtin. */ +#define _A1(a) {a, GCN_BTI_END_OF_PARAMS} +#define _A2(a,b) {a, b, GCN_BTI_END_OF_PARAMS} +#define _A3(a,b,c) {a, b, c, GCN_BTI_END_OF_PARAMS} +#define _A4(a,b,c,d) {a, b, c, d, GCN_BTI_END_OF_PARAMS} +#define _A5(a,b,c,d,e) {a, b, c, d, e, GCN_BTI_END_OF_PARAMS} + +DEF_BUILTIN (FLAT_LOAD_INT32, 1 /*CODE_FOR_flat_load_v64si*/, + "flat_load_int32", B_INSN, + _A3 (GCN_BTI_V64SI, GCN_BTI_EXEC, GCN_BTI_V64SI), + gcn_expand_builtin_1) + +DEF_BUILTIN (FLAT_LOAD_PTR_INT32, 2 /*CODE_FOR_flat_load_ptr_v64si */, + "flat_load_ptr_int32", B_INSN, + _A4 (GCN_BTI_V64SI, GCN_BTI_EXEC, GCN_BTI_SIPTR, GCN_BTI_V64SI), + gcn_expand_builtin_1) + +DEF_BUILTIN (FLAT_STORE_PTR_INT32, 3 /*CODE_FOR_flat_store_ptr_v64si */, + "flat_store_ptr_int32", B_INSN, + _A5 (GCN_BTI_VOID, GCN_BTI_EXEC, GCN_BTI_SIPTR, GCN_BTI_V64SI, + GCN_BTI_V64SI), + gcn_expand_builtin_1) + +DEF_BUILTIN (FLAT_LOAD_PTR_FLOAT, 2 /*CODE_FOR_flat_load_ptr_v64sf */, + "flat_load_ptr_float", B_INSN, + _A4 (GCN_BTI_V64SF, GCN_BTI_EXEC, GCN_BTI_SFPTR, GCN_BTI_V64SI), + gcn_expand_builtin_1) + +DEF_BUILTIN (FLAT_STORE_PTR_FLOAT, 3 /*CODE_FOR_flat_store_ptr_v64sf */, + "flat_store_ptr_float", B_INSN, + _A5 (GCN_BTI_VOID, GCN_BTI_EXEC, GCN_BTI_SFPTR, GCN_BTI_V64SI, + GCN_BTI_V64SF), + gcn_expand_builtin_1) + +DEF_BUILTIN (SQRTVF, 3 /*CODE_FOR_sqrtvf */, + "sqrtvf", B_INSN, + _A2 (GCN_BTI_V64SF, GCN_BTI_V64SF), + gcn_expand_builtin_1) + +DEF_BUILTIN (SQRTF, 3 /*CODE_FOR_sqrtf */, + "sqrtf", B_INSN, + _A2 (GCN_BTI_SF, GCN_BTI_SF), + gcn_expand_builtin_1) + +DEF_BUILTIN (CMP_SWAP, -1, + "cmp_swap", B_INSN, + _A4 (GCN_BTI_UINT, GCN_BTI_VOIDPTR, GCN_BTI_UINT, GCN_BTI_UINT), + gcn_expand_builtin_1) + +DEF_BUILTIN (CMP_SWAPLL, -1, + "cmp_swapll", B_INSN, + _A4 (GCN_BTI_LLUINT, + GCN_BTI_VOIDPTR, GCN_BTI_LLUINT, GCN_BTI_LLUINT), + gcn_expand_builtin_1) + +/* DEF_BUILTIN_BINOP_INT_FP creates many variants of a builtin function for a + given operation. The first argument will give base to the identifier of a + particular builtin, the second will be used to form the name of the patter + used to expand it to and the third will be used to create the user-visible + builtin identifier. */ + +DEF_BUILTIN_BINOP_INT_FP (ADD, add, "add") +DEF_BUILTIN_BINOP_INT_FP (SUB, sub, "sub") + +DEF_BUILTIN_BINOP_INT_FP (AND, and, "and") +DEF_BUILTIN_BINOP_INT_FP (IOR, ior, "or") +DEF_BUILTIN_BINOP_INT_FP (XOR, xor, "xor") + +/* OpenMP. */ + +DEF_BUILTIN (OMP_DIM_SIZE, CODE_FOR_oacc_dim_size, + "dim_size", B_INSN, + _A2 (GCN_BTI_INT, GCN_BTI_INT), + gcn_expand_builtin_1) +DEF_BUILTIN (OMP_DIM_POS, CODE_FOR_oacc_dim_pos, + "dim_pos", B_INSN, + _A2 (GCN_BTI_INT, GCN_BTI_INT), + gcn_expand_builtin_1) + +/* OpenACC. */ + +DEF_BUILTIN (ACC_SINGLE_START, -1, "single_start", B_INSN, _A1 (GCN_BTI_BOOL), + gcn_expand_builtin_1) + +DEF_BUILTIN (ACC_SINGLE_COPY_START, -1, "single_copy_start", B_INSN, + _A1 (GCN_BTI_LDS_VOIDPTR), gcn_expand_builtin_1) + +DEF_BUILTIN (ACC_SINGLE_COPY_END, -1, "single_copy_end", B_INSN, + _A2 (GCN_BTI_VOID, GCN_BTI_LDS_VOIDPTR), gcn_expand_builtin_1) + +DEF_BUILTIN (ACC_BARRIER, -1, "acc_barrier", B_INSN, _A1 (GCN_BTI_VOID), + gcn_expand_builtin_1) + + +#undef _A1 +#undef _A2 +#undef _A3 +#undef _A4 +#undef _A5 diff --git a/gcc/config/gcn/gcn-hsa.h b/gcc/config/gcn/gcn-hsa.h new file mode 100644 index 00000000000..2e3746cdab8 --- /dev/null +++ b/gcc/config/gcn/gcn-hsa.h @@ -0,0 +1,115 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +#ifndef OBJECT_FORMAT_ELF + #error elf.h included before elfos.h +#endif + +#define TEXT_SECTION_ASM_OP "\t.section\t.text" +#define BSS_SECTION_ASM_OP "\t.section\t.bss" +#define GLOBAL_ASM_OP "\t.globl\t" +#define DATA_SECTION_ASM_OP "\t.data\t" +#define SET_ASM_OP "\t.set\t" +#define LOCAL_LABEL_PREFIX "." +#define USER_LABEL_PREFIX "" +#define ASM_COMMENT_START ";" +#define TARGET_ASM_NAMED_SECTION default_elf_asm_named_section + +#define ASM_OUTPUT_ALIGNED_BSS(FILE, DECL, NAME, SIZE, ALIGN) \ + asm_output_aligned_bss (FILE, DECL, NAME, SIZE, ALIGN) + +#undef ASM_DECLARE_FUNCTION_NAME +#define ASM_DECLARE_FUNCTION_NAME(FILE, NAME, DECL) \ + gcn_hsa_declare_function_name ((FILE), (NAME), (DECL)) + +/* Unlike GNU as, the LLVM assembler uses log2 alignments. */ +#undef ASM_OUTPUT_ALIGNED_COMMON +#define ASM_OUTPUT_ALIGNED_COMMON(FILE, NAME, SIZE, ALIGNMENT) \ + (fprintf ((FILE), "%s", COMMON_ASM_OP), \ + assemble_name ((FILE), (NAME)), \ + fprintf ((FILE), "," HOST_WIDE_INT_PRINT_UNSIGNED ",%u\n", \ + (SIZE) > 0 ? (SIZE) : 1, exact_log2 ((ALIGNMENT) / BITS_PER_UNIT))) + +#define ASM_OUTPUT_LABEL(FILE,NAME) \ + do { assemble_name (FILE, NAME); fputs (":\n", FILE); } while (0) + +#define ASM_OUTPUT_LABELREF(FILE, NAME) \ + asm_fprintf (FILE, "%U%s", default_strip_name_encoding (NAME)) + +extern unsigned int gcn_local_sym_hash (const char *name); + +#define ASM_OUTPUT_SYMBOL_REF(FILE, X) gcn_asm_output_symbol_ref (FILE, X) + +#define ASM_OUTPUT_ADDR_DIFF_ELT(FILE, BODY, VALUE, REL) \ + fprintf (FILE, "\t.word .L%d-.L%d\n", VALUE, REL) + +#define ASM_OUTPUT_ADDR_VEC_ELT(FILE, VALUE) \ + fprintf (FILE, "\t.word .L%d\n", VALUE) + +#define ASM_OUTPUT_ALIGN(FILE,LOG) \ + do { if (LOG!=0) fprintf (FILE, "\t.align\t%d\n", 1<<(LOG)); } while (0) +#define ASM_OUTPUT_ALIGN_WITH_NOP(FILE,LOG) \ + do { \ + if (LOG!=0) \ + fprintf (FILE, "\t.p2alignl\t%d, 0xBF800000" \ + " ; Fill value is 's_nop 0'\n", (LOG)); \ + } while (0) + +#define ASM_APP_ON "" +#define ASM_APP_OFF "" + +/* Avoid the default in ../../gcc.c, which adds "-pthread", which is not + supported for gcn. */ +#define GOMP_SELF_SPECS "" + +/* Use LLVM assembler and linker options. */ +#define ASM_SPEC "-triple=amdgcn--amdhsa " \ + "%:last_arg(%{march=*:-mcpu=%*}) " \ + "-filetype=obj" +/* Add -mlocal-symbol-id= unless the user (or mkoffload) + passes the option explicitly on the command line. The option also causes + several dump-matching tests to fail in the testsuite, so the option is not + added when or tree dump/compare-debug options used in the testsuite are + present. + This has the potential for surprise, but a user can still use an explicit + -mlocal-symbol-id= option manually together with -fdump-tree or + -fcompare-debug options. */ +#define CC1_SPEC "%{!mlocal-symbol-id=*:%{!fdump-tree-*:" \ + "%{!fdump-ipa-*:%{!fcompare-debug*:-mlocal-symbol-id=%b}}}}" +#define LINK_SPEC "--pie" +#define LIB_SPEC "-lc" + +/* Provides a _start symbol to keep the linker happy. */ +#define STARTFILE_SPEC "crt0.o%s" +#define ENDFILE_SPEC "" +#define STANDARD_STARTFILE_PREFIX_2 "" + +/* The LLVM assembler rejects multiple -mcpu options, so we must drop + all but the last. */ +extern const char *last_arg_spec_function (int argc, const char **argv); +#define EXTRA_SPEC_FUNCTIONS \ + { "last_arg", last_arg_spec_function }, + +#undef LOCAL_INCLUDE_DIR + +/* FIXME: Review debug info settings. + * In particular, EH_FRAME_THROUGH_COLLECT2 is probably the wrong + * thing but stuff fails to build without it. + * (Debug info is not a big deal until we get a debugger.) */ +#define PREFERRED_DEBUGGING_TYPE DWARF2_DEBUG +#define DWARF2_DEBUGGING_INFO 1 +#define DWARF2_ASM_LINE_DEBUG_INFO 1 +#define EH_FRAME_THROUGH_COLLECT2 1 diff --git a/gcc/config/gcn/gcn-modes.def b/gcc/config/gcn/gcn-modes.def new file mode 100644 index 00000000000..6d73100c4f4 --- /dev/null +++ b/gcc/config/gcn/gcn-modes.def @@ -0,0 +1,41 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +/* Half-precision floating point */ +FLOAT_MODE (HF, 2, 0); +/* FIXME: No idea what format it is. */ +ADJUST_FLOAT_FORMAT (HF, &ieee_half_format); + +/* Native vector modes. */ +VECTOR_MODE (INT, QI, 64); /* V64QI */ +VECTOR_MODE (INT, HI, 64); /* V64HI */ +VECTOR_MODE (INT, SI, 64); /* V64SI */ +VECTOR_MODE (INT, DI, 64); /* V64DI */ +VECTOR_MODE (INT, TI, 64); /* V64TI */ +VECTOR_MODE (FLOAT, HF, 64); /* V64HF */ +VECTOR_MODE (FLOAT, SF, 64); /* V64SF */ +VECTOR_MODE (FLOAT, DF, 64); /* V64DF */ + +/* Vector units handle reads independently and thus no large alignment + needed. */ +ADJUST_ALIGNMENT (V64QI, 1); +ADJUST_ALIGNMENT (V64HI, 2); +ADJUST_ALIGNMENT (V64SI, 4); +ADJUST_ALIGNMENT (V64DI, 8); +ADJUST_ALIGNMENT (V64TI, 16); +ADJUST_ALIGNMENT (V64HF, 2); +ADJUST_ALIGNMENT (V64SF, 4); +ADJUST_ALIGNMENT (V64DF, 8); diff --git a/gcc/config/gcn/gcn-opts.h b/gcc/config/gcn/gcn-opts.h new file mode 100644 index 00000000000..e2b32b6c461 --- /dev/null +++ b/gcc/config/gcn/gcn-opts.h @@ -0,0 +1,36 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +#ifndef GCN_OPTS_H +#define GCN_OPTS_H + +/* Which processor to generate code or schedule for. */ +enum processor_type +{ + PROCESSOR_CARRIZO, + PROCESSOR_FIJI, + PROCESSOR_VEGA +}; + +/* Set in gcn_option_override. */ +extern int gcn_isa; + +#define TARGET_GCN3 (gcn_isa == 3) +#define TARGET_GCN3_PLUS (gcn_isa >= 3) +#define TARGET_GCN5 (gcn_isa == 5) +#define TARGET_GCN5_PLUS (gcn_isa >= 5) + +#endif diff --git a/gcc/config/gcn/gcn-passes.def b/gcc/config/gcn/gcn-passes.def new file mode 100644 index 00000000000..50f583fbbbe --- /dev/null +++ b/gcc/config/gcn/gcn-passes.def @@ -0,0 +1,19 @@ +/* Copyright (C) 2017-2019 Free Software Foundation, Inc. + + This file is part of GCC. + + GCC is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3, or (at your option) any later + version. + + GCC is distributed in the hope that it will be useful, but WITHOUT ANY + WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +INSERT_PASS_AFTER (pass_omp_target_link, 1, pass_omp_gcn); diff --git a/gcc/config/gcn/gcn-protos.h b/gcc/config/gcn/gcn-protos.h new file mode 100644 index 00000000000..da7faf29c70 --- /dev/null +++ b/gcc/config/gcn/gcn-protos.h @@ -0,0 +1,146 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +#ifndef _GCN_PROTOS_ +#define _GCN_PROTOS_ + +extern void gcn_asm_output_symbol_ref (FILE *file, rtx x); +extern tree gcn_builtin_decl (unsigned code, bool initialize_p); +extern bool gcn_can_split_p (machine_mode, rtx); +extern bool gcn_constant64_p (rtx); +extern bool gcn_constant_p (rtx); +extern rtx gcn_convert_mask_mode (rtx reg); +extern char * gcn_expand_dpp_shr_insn (machine_mode, const char *, int, int); +extern void gcn_expand_epilogue (); +extern rtx gcn_expand_scaled_offsets (addr_space_t as, rtx base, rtx offsets, + rtx scale, bool unsigned_p, rtx exec); +extern void gcn_expand_prologue (); +extern rtx gcn_expand_reduc_scalar (machine_mode, rtx, int); +extern rtx gcn_expand_scalar_to_vector_address (machine_mode, rtx, rtx, rtx); +extern void gcn_expand_vector_init (rtx, rtx); +extern bool gcn_flat_address_p (rtx, machine_mode); +extern bool gcn_fp_constant_p (rtx, bool); +extern rtx gcn_full_exec (); +extern rtx gcn_full_exec_reg (); +extern rtx gcn_gen_undef (machine_mode); +extern bool gcn_global_address_p (rtx); +extern tree gcn_goacc_adjust_propagation_record (tree record_type, bool sender, + const char *name); +extern void gcn_goacc_adjust_gangprivate_decl (tree var); +extern void gcn_goacc_reduction (gcall *call); +extern bool gcn_hard_regno_rename_ok (unsigned int from_reg, + unsigned int to_reg); +extern machine_mode gcn_hard_regno_caller_save_mode (unsigned int regno, + unsigned int nregs, + machine_mode regmode); +extern bool gcn_hard_regno_mode_ok (int regno, machine_mode mode); +extern int gcn_hard_regno_nregs (int regno, machine_mode mode); +extern void gcn_hsa_declare_function_name (FILE *file, const char *name, + tree decl); +extern HOST_WIDE_INT gcn_initial_elimination_offset (int, int); +extern bool gcn_inline_constant64_p (rtx); +extern bool gcn_inline_constant_p (rtx); +extern int gcn_inline_fp_constant_p (rtx, bool); +extern reg_class gcn_mode_code_base_reg_class (machine_mode, addr_space_t, + int, int); +extern rtx gcn_oacc_dim_pos (int dim); +extern rtx gcn_oacc_dim_size (int dim); +extern rtx gcn_operand_doublepart (machine_mode, rtx, int); +extern rtx gcn_operand_part (machine_mode, rtx, int); +extern bool gcn_regno_mode_code_ok_for_base_p (int, machine_mode, + addr_space_t, int, int); +extern reg_class gcn_regno_reg_class (int regno); +extern rtx gcn_scalar_exec (); +extern rtx gcn_scalar_exec_reg (); +extern bool gcn_scalar_flat_address_p (rtx); +extern bool gcn_scalar_flat_mem_p (rtx); +extern bool gcn_sgpr_move_p (rtx, rtx); +extern bool gcn_valid_move_p (machine_mode, rtx, rtx); +extern rtx gcn_vec_constant (machine_mode, int); +extern rtx gcn_vec_constant (machine_mode, rtx); +extern bool gcn_vgpr_move_p (rtx, rtx); +extern void print_operand_address (FILE *file, register rtx addr); +extern void print_operand (FILE *file, rtx x, int code); +extern bool regno_ok_for_index_p (int); + +enum gcn_cvt_t +{ + fix_trunc_cvt, + fixuns_trunc_cvt, + float_cvt, + floatuns_cvt, + extend_cvt, + trunc_cvt +}; + +extern bool gcn_valid_cvt_p (machine_mode from, machine_mode to, + enum gcn_cvt_t op); + +#ifdef TREE_CODE +extern void gcn_init_cumulative_args (CUMULATIVE_ARGS *, tree, rtx, tree, + int); +class gimple_opt_pass; +extern gimple_opt_pass *make_pass_omp_gcn (gcc::context *ctxt); +#endif + +/* Return true if MODE is valid for 1 VGPR register. */ + +inline bool +vgpr_1reg_mode_p (machine_mode mode) +{ + return (mode == SImode || mode == SFmode || mode == HImode || mode == QImode + || mode == V64QImode || mode == V64HImode || mode == V64SImode + || mode == V64HFmode || mode == V64SFmode || mode == BImode); +} + +/* Return true if MODE is valid for 1 SGPR register. */ + +inline bool +sgpr_1reg_mode_p (machine_mode mode) +{ + return (mode == SImode || mode == SFmode || mode == HImode + || mode == QImode || mode == BImode); +} + +/* Return true if MODE is valid for pair of VGPR registers. */ + +inline bool +vgpr_2reg_mode_p (machine_mode mode) +{ + return (mode == DImode || mode == DFmode + || mode == V64DImode || mode == V64DFmode); +} + +/* Return true if MODE can be handled directly by VGPR operations. */ + +inline bool +vgpr_vector_mode_p (machine_mode mode) +{ + return (mode == V64QImode || mode == V64HImode + || mode == V64SImode || mode == V64DImode + || mode == V64HFmode || mode == V64SFmode || mode == V64DFmode); +} + + +/* Return true if MODE is valid for pair of SGPR registers. */ + +inline bool +sgpr_2reg_mode_p (machine_mode mode) +{ + return mode == DImode || mode == DFmode; +} + +#endif diff --git a/gcc/config/gcn/gcn-run.c b/gcc/config/gcn/gcn-run.c new file mode 100644 index 00000000000..58089843ef8 --- /dev/null +++ b/gcc/config/gcn/gcn-run.c @@ -0,0 +1,850 @@ +/* Run a stand-alone AMD GCN kernel. + + Copyright 2017 Mentor Graphics Corporation + Copyright 2018-2019 Free Software Foundation, Inc. + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . */ + +/* This program will run a compiled stand-alone GCN kernel on a GPU. + + The kernel entry point's signature must use a standard main signature: + + int main(int argc, char **argv) +*/ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* These probably won't be in elf.h for a while. */ +#ifndef R_AMDGPU_NONE +#define R_AMDGPU_NONE 0 +#define R_AMDGPU_ABS32_LO 1 /* (S + A) & 0xFFFFFFFF */ +#define R_AMDGPU_ABS32_HI 2 /* (S + A) >> 32 */ +#define R_AMDGPU_ABS64 3 /* S + A */ +#define R_AMDGPU_REL32 4 /* S + A - P */ +#define R_AMDGPU_REL64 5 /* S + A - P */ +#define R_AMDGPU_ABS32 6 /* S + A */ +#define R_AMDGPU_GOTPCREL 7 /* G + GOT + A - P */ +#define R_AMDGPU_GOTPCREL32_LO 8 /* (G + GOT + A - P) & 0xFFFFFFFF */ +#define R_AMDGPU_GOTPCREL32_HI 9 /* (G + GOT + A - P) >> 32 */ +#define R_AMDGPU_REL32_LO 10 /* (S + A - P) & 0xFFFFFFFF */ +#define R_AMDGPU_REL32_HI 11 /* (S + A - P) >> 32 */ +#define reserved 12 +#define R_AMDGPU_RELATIVE64 13 /* B + A */ +#endif + +#include "hsa.h" + +#ifndef HSA_RUNTIME_LIB +#define HSA_RUNTIME_LIB "libhsa-runtime64.so" +#endif + +#ifndef VERSION_STRING +#define VERSION_STRING "(version unknown)" +#endif + +bool debug = false; + +hsa_agent_t device = { 0 }; +hsa_queue_t *queue = NULL; +uint64_t kernel = 0; +hsa_executable_t executable = { 0 }; + +hsa_region_t kernargs_region = { 0 }; +uint32_t kernarg_segment_size = 0; +uint32_t group_segment_size = 0; +uint32_t private_segment_size = 0; + +static void +usage (const char *progname) +{ + printf ("Usage: %s [options] kernel [kernel-args]\n\n" + "Options:\n" + " --help\n" + " --version\n" + " --debug\n", progname); +} + +static void +version (const char *progname) +{ + printf ("%s " VERSION_STRING "\n", progname); +} + +/* As an HSA runtime is dlopened, following structure defines the necessary + function pointers. + Code adapted from libgomp. */ + +struct hsa_runtime_fn_info +{ + /* HSA runtime. */ + hsa_status_t (*hsa_status_string_fn) (hsa_status_t status, + const char **status_string); + hsa_status_t (*hsa_agent_get_info_fn) (hsa_agent_t agent, + hsa_agent_info_t attribute, + void *value); + hsa_status_t (*hsa_init_fn) (void); + hsa_status_t (*hsa_iterate_agents_fn) + (hsa_status_t (*callback) (hsa_agent_t agent, void *data), void *data); + hsa_status_t (*hsa_region_get_info_fn) (hsa_region_t region, + hsa_region_info_t attribute, + void *value); + hsa_status_t (*hsa_queue_create_fn) + (hsa_agent_t agent, uint32_t size, hsa_queue_type_t type, + void (*callback) (hsa_status_t status, hsa_queue_t *source, void *data), + void *data, uint32_t private_segment_size, + uint32_t group_segment_size, hsa_queue_t **queue); + hsa_status_t (*hsa_agent_iterate_regions_fn) + (hsa_agent_t agent, + hsa_status_t (*callback) (hsa_region_t region, void *data), void *data); + hsa_status_t (*hsa_executable_destroy_fn) (hsa_executable_t executable); + hsa_status_t (*hsa_executable_create_fn) + (hsa_profile_t profile, hsa_executable_state_t executable_state, + const char *options, hsa_executable_t *executable); + hsa_status_t (*hsa_executable_global_variable_define_fn) + (hsa_executable_t executable, const char *variable_name, void *address); + hsa_status_t (*hsa_executable_load_code_object_fn) + (hsa_executable_t executable, hsa_agent_t agent, + hsa_code_object_t code_object, const char *options); + hsa_status_t (*hsa_executable_freeze_fn) (hsa_executable_t executable, + const char *options); + hsa_status_t (*hsa_signal_create_fn) (hsa_signal_value_t initial_value, + uint32_t num_consumers, + const hsa_agent_t *consumers, + hsa_signal_t *signal); + hsa_status_t (*hsa_memory_allocate_fn) (hsa_region_t region, size_t size, + void **ptr); + hsa_status_t (*hsa_memory_copy_fn) (void *dst, const void *src, + size_t size); + hsa_status_t (*hsa_memory_free_fn) (void *ptr); + hsa_status_t (*hsa_signal_destroy_fn) (hsa_signal_t signal); + hsa_status_t (*hsa_executable_get_symbol_fn) + (hsa_executable_t executable, const char *module_name, + const char *symbol_name, hsa_agent_t agent, int32_t call_convention, + hsa_executable_symbol_t *symbol); + hsa_status_t (*hsa_executable_symbol_get_info_fn) + (hsa_executable_symbol_t executable_symbol, + hsa_executable_symbol_info_t attribute, void *value); + void (*hsa_signal_store_relaxed_fn) (hsa_signal_t signal, + hsa_signal_value_t value); + hsa_signal_value_t (*hsa_signal_wait_acquire_fn) + (hsa_signal_t signal, hsa_signal_condition_t condition, + hsa_signal_value_t compare_value, uint64_t timeout_hint, + hsa_wait_state_t wait_state_hint); + hsa_signal_value_t (*hsa_signal_wait_relaxed_fn) + (hsa_signal_t signal, hsa_signal_condition_t condition, + hsa_signal_value_t compare_value, uint64_t timeout_hint, + hsa_wait_state_t wait_state_hint); + hsa_status_t (*hsa_queue_destroy_fn) (hsa_queue_t *queue); + hsa_status_t (*hsa_code_object_deserialize_fn) + (void *serialized_code_object, size_t serialized_code_object_size, + const char *options, hsa_code_object_t *code_object); + uint64_t (*hsa_queue_load_write_index_relaxed_fn) + (const hsa_queue_t *queue); + void (*hsa_queue_store_write_index_relaxed_fn) + (const hsa_queue_t *queue, uint64_t value); + hsa_status_t (*hsa_shut_down_fn) (); +}; + +/* HSA runtime functions that are initialized in init_hsa_context. + Code adapted from libgomp. */ + +static struct hsa_runtime_fn_info hsa_fns; + +#define DLSYM_FN(function) \ + *(void**)(&hsa_fns.function##_fn) = dlsym (handle, #function); \ + if (hsa_fns.function##_fn == NULL) \ + goto fail; + +static void +init_hsa_runtime_functions (void) +{ + void *handle = dlopen (HSA_RUNTIME_LIB, RTLD_LAZY); + if (handle == NULL) + { + fprintf (stderr, + "The HSA runtime is required to run GCN kernels on hardware.\n" + "%s: File not found or could not be opened\n", + HSA_RUNTIME_LIB); + exit (1); + } + + DLSYM_FN (hsa_status_string) + DLSYM_FN (hsa_agent_get_info) + DLSYM_FN (hsa_init) + DLSYM_FN (hsa_iterate_agents) + DLSYM_FN (hsa_region_get_info) + DLSYM_FN (hsa_queue_create) + DLSYM_FN (hsa_agent_iterate_regions) + DLSYM_FN (hsa_executable_destroy) + DLSYM_FN (hsa_executable_create) + DLSYM_FN (hsa_executable_global_variable_define) + DLSYM_FN (hsa_executable_load_code_object) + DLSYM_FN (hsa_executable_freeze) + DLSYM_FN (hsa_signal_create) + DLSYM_FN (hsa_memory_allocate) + DLSYM_FN (hsa_memory_copy) + DLSYM_FN (hsa_memory_free) + DLSYM_FN (hsa_signal_destroy) + DLSYM_FN (hsa_executable_get_symbol) + DLSYM_FN (hsa_executable_symbol_get_info) + DLSYM_FN (hsa_signal_wait_acquire) + DLSYM_FN (hsa_signal_wait_relaxed) + DLSYM_FN (hsa_signal_store_relaxed) + DLSYM_FN (hsa_queue_destroy) + DLSYM_FN (hsa_code_object_deserialize) + DLSYM_FN (hsa_queue_load_write_index_relaxed) + DLSYM_FN (hsa_queue_store_write_index_relaxed) + DLSYM_FN (hsa_shut_down) + + return; + +fail: + fprintf (stderr, "Failed to find HSA functions in " HSA_RUNTIME_LIB "\n"); + exit (1); +} + +#undef DLSYM_FN + +/* Report a fatal error STR together with the HSA error corresponding to + STATUS and terminate execution of the current process. */ + +static void +hsa_fatal (const char *str, hsa_status_t status) +{ + const char *hsa_error_msg; + hsa_fns.hsa_status_string_fn (status, &hsa_error_msg); + fprintf (stderr, "%s: FAILED\nHSA Runtime message: %s\n", str, + hsa_error_msg); + exit (1); +} + +/* Helper macros to ensure we check the return values from the HSA Runtime. + These just keep the rest of the code a bit cleaner. */ + +#define XHSA_CMP(FN, CMP, MSG) \ + do { \ + hsa_status_t status = (FN); \ + if (!(CMP)) \ + hsa_fatal ((MSG), status); \ + else if (debug) \ + fprintf (stderr, "%s: OK\n", (MSG)); \ + } while (0) +#define XHSA(FN, MSG) XHSA_CMP(FN, status == HSA_STATUS_SUCCESS, MSG) + +/* Callback of hsa_iterate_agents. + Called once for each available device, and returns "break" when a + suitable one has been found. */ + +static hsa_status_t +get_gpu_agent (hsa_agent_t agent, void *data __attribute__ ((unused))) +{ + hsa_device_type_t device_type; + XHSA (hsa_fns.hsa_agent_get_info_fn (agent, HSA_AGENT_INFO_DEVICE, + &device_type), + "Get agent type"); + + /* Select only GPU devices. */ + /* TODO: support selecting from multiple GPUs. */ + if (HSA_DEVICE_TYPE_GPU == device_type) + { + device = agent; + return HSA_STATUS_INFO_BREAK; + } + + /* The device was not suitable. */ + return HSA_STATUS_SUCCESS; +} + +/* Callback of hsa_iterate_regions. + Called once for each available memory region, and returns "break" when a + suitable one has been found. */ + +static hsa_status_t +get_kernarg_region (hsa_region_t region, void *data __attribute__ ((unused))) +{ + /* Reject non-global regions. */ + hsa_region_segment_t segment; + hsa_fns.hsa_region_get_info_fn (region, HSA_REGION_INFO_SEGMENT, &segment); + if (HSA_REGION_SEGMENT_GLOBAL != segment) + return HSA_STATUS_SUCCESS; + + /* Find a region with the KERNARG flag set. */ + hsa_region_global_flag_t flags; + hsa_fns.hsa_region_get_info_fn (region, HSA_REGION_INFO_GLOBAL_FLAGS, + &flags); + if (flags & HSA_REGION_GLOBAL_FLAG_KERNARG) + { + kernargs_region = region; + return HSA_STATUS_INFO_BREAK; + } + + /* The region was not suitable. */ + return HSA_STATUS_SUCCESS; +} + +/* Initialize the HSA Runtime library and GPU device. */ + +static void +init_device () +{ + /* Load the shared library and find the API functions. */ + init_hsa_runtime_functions (); + + /* Initialize the HSA Runtime. */ + XHSA (hsa_fns.hsa_init_fn (), + "Initialize run-time"); + + /* Select a suitable device. + The call-back function, get_gpu_agent, does the selection. */ + XHSA_CMP (hsa_fns.hsa_iterate_agents_fn (get_gpu_agent, NULL), + status == HSA_STATUS_SUCCESS || status == HSA_STATUS_INFO_BREAK, + "Find a device"); + + /* Initialize the queue used for launching kernels. */ + uint32_t queue_size = 0; + XHSA (hsa_fns.hsa_agent_get_info_fn (device, HSA_AGENT_INFO_QUEUE_MAX_SIZE, + &queue_size), + "Find max queue size"); + XHSA (hsa_fns.hsa_queue_create_fn (device, queue_size, + HSA_QUEUE_TYPE_SINGLE, NULL, + NULL, UINT32_MAX, UINT32_MAX, &queue), + "Set up a device queue"); + + /* Select a memory region for the kernel arguments. + The call-back function, get_kernarg_region, does the selection. */ + XHSA_CMP (hsa_fns.hsa_agent_iterate_regions_fn (device, get_kernarg_region, + NULL), + status == HSA_STATUS_SUCCESS || status == HSA_STATUS_INFO_BREAK, + "Locate kernargs memory"); +} + + +/* Read a whole input file. + Code copied from mkoffload. */ + +static char * +read_file (const char *filename, size_t *plen) +{ + size_t alloc = 16384; + size_t base = 0; + char *buffer; + + FILE *stream = fopen (filename, "rb"); + if (!stream) + { + perror (filename); + exit (1); + } + + if (!fseek (stream, 0, SEEK_END)) + { + /* Get the file size. */ + long s = ftell (stream); + if (s >= 0) + alloc = s + 100; + fseek (stream, 0, SEEK_SET); + } + buffer = malloc (alloc); + + for (;;) + { + size_t n = fread (buffer + base, 1, alloc - base - 1, stream); + + if (!n) + break; + base += n; + if (base + 1 == alloc) + { + alloc *= 2; + buffer = realloc (buffer, alloc); + } + } + buffer[base] = 0; + *plen = base; + + fclose (stream); + + return buffer; +} + +/* Read a HSA Code Object (HSACO) from file, and load it into the device. */ + +static void +load_image (const char *filename) +{ + size_t image_size; + Elf64_Ehdr *image = (void *) read_file (filename, &image_size); + + /* An "executable" consists of one or more code objects. */ + XHSA (hsa_fns.hsa_executable_create_fn (HSA_PROFILE_FULL, + HSA_EXECUTABLE_STATE_UNFROZEN, "", + &executable), + "Initialize GCN executable"); + + /* Hide relocations from the HSA runtime loader. + Keep a copy of the unmodified section headers to use later. */ + Elf64_Shdr *image_sections = + (Elf64_Shdr *) ((char *) image + image->e_shoff); + Elf64_Shdr *sections = malloc (sizeof (Elf64_Shdr) * image->e_shnum); + memcpy (sections, image_sections, sizeof (Elf64_Shdr) * image->e_shnum); + for (int i = image->e_shnum - 1; i >= 0; i--) + { + if (image_sections[i].sh_type == SHT_RELA + || image_sections[i].sh_type == SHT_REL) + /* Change section type to something harmless. */ + image_sections[i].sh_type = SHT_NOTE; + } + + /* Add the HSACO to the executable. */ + hsa_code_object_t co = { 0 }; + XHSA (hsa_fns.hsa_code_object_deserialize_fn (image, image_size, NULL, &co), + "Deserialize GCN code object"); + XHSA (hsa_fns.hsa_executable_load_code_object_fn (executable, device, co, + ""), + "Load GCN code object"); + + /* We're done modifying he executable. */ + XHSA (hsa_fns.hsa_executable_freeze_fn (executable, ""), + "Freeze GCN executable"); + + /* Locate the "main" function, and read the kernel's properties. */ + hsa_executable_symbol_t symbol; + XHSA (hsa_fns.hsa_executable_get_symbol_fn (executable, NULL, "main", + device, 0, &symbol), + "Find 'main' function"); + XHSA (hsa_fns.hsa_executable_symbol_get_info_fn + (symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT, &kernel), + "Extract kernel object"); + XHSA (hsa_fns.hsa_executable_symbol_get_info_fn + (symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_KERNARG_SEGMENT_SIZE, + &kernarg_segment_size), + "Extract kernarg segment size"); + XHSA (hsa_fns.hsa_executable_symbol_get_info_fn + (symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_GROUP_SEGMENT_SIZE, + &group_segment_size), + "Extract group segment size"); + XHSA (hsa_fns.hsa_executable_symbol_get_info_fn + (symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_PRIVATE_SEGMENT_SIZE, + &private_segment_size), + "Extract private segment size"); + + /* Find main function in ELF, and calculate actual load offset. */ + Elf64_Addr load_offset; + XHSA (hsa_fns.hsa_executable_symbol_get_info_fn + (symbol, HSA_EXECUTABLE_SYMBOL_INFO_VARIABLE_ADDRESS, + &load_offset), + "Extract 'main' symbol address"); + for (int i = 0; i < image->e_shnum; i++) + if (sections[i].sh_type == SHT_SYMTAB) + { + Elf64_Shdr *strtab = §ions[sections[i].sh_link]; + char *strings = (char *) image + strtab->sh_offset; + + for (size_t offset = 0; + offset < sections[i].sh_size; + offset += sections[i].sh_entsize) + { + Elf64_Sym *sym = (Elf64_Sym *) ((char *) image + + sections[i].sh_offset + offset); + if (strcmp ("main", strings + sym->st_name) == 0) + { + load_offset -= sym->st_value; + goto found_main; + } + } + } + /* We only get here when main was not found. + This should never happen. */ + fprintf (stderr, "Error: main function not found.\n"); + abort (); +found_main:; + + /* Find dynamic symbol table. */ + Elf64_Shdr *dynsym = NULL; + for (int i = 0; i < image->e_shnum; i++) + if (sections[i].sh_type == SHT_DYNSYM) + { + dynsym = §ions[i]; + break; + } + + /* Fix up relocations. */ + for (int i = 0; i < image->e_shnum; i++) + { + if (sections[i].sh_type == SHT_RELA) + for (size_t offset = 0; + offset < sections[i].sh_size; + offset += sections[i].sh_entsize) + { + Elf64_Rela *reloc = (Elf64_Rela *) ((char *) image + + sections[i].sh_offset + + offset); + Elf64_Sym *sym = + (dynsym + ? (Elf64_Sym *) ((char *) image + + dynsym->sh_offset + + (dynsym->sh_entsize + * ELF64_R_SYM (reloc->r_info))) : NULL); + + int64_t S = (sym ? sym->st_value : 0); + int64_t P = reloc->r_offset + load_offset; + int64_t A = reloc->r_addend; + int64_t B = load_offset; + int64_t V, size; + switch (ELF64_R_TYPE (reloc->r_info)) + { + case R_AMDGPU_ABS32_LO: + V = (S + A) & 0xFFFFFFFF; + size = 4; + break; + case R_AMDGPU_ABS32_HI: + V = (S + A) >> 32; + size = 4; + break; + case R_AMDGPU_ABS64: + V = S + A; + size = 8; + break; + case R_AMDGPU_REL32: + V = S + A - P; + size = 4; + break; + case R_AMDGPU_REL64: + /* FIXME + LLD seems to emit REL64 where the the assembler has ABS64. + This is clearly wrong because it's not what the compiler + is expecting. Let's assume, for now, that it's a bug. + In any case, GCN kernels are always self contained and + therefore relative relocations will have been resolved + already, so this should be a safe workaround. */ + V = S + A /* - P */ ; + size = 8; + break; + case R_AMDGPU_ABS32: + V = S + A; + size = 4; + break; + /* TODO R_AMDGPU_GOTPCREL */ + /* TODO R_AMDGPU_GOTPCREL32_LO */ + /* TODO R_AMDGPU_GOTPCREL32_HI */ + case R_AMDGPU_REL32_LO: + V = (S + A - P) & 0xFFFFFFFF; + size = 4; + break; + case R_AMDGPU_REL32_HI: + V = (S + A - P) >> 32; + size = 4; + break; + case R_AMDGPU_RELATIVE64: + V = B + A; + size = 8; + break; + default: + fprintf (stderr, "Error: unsupported relocation type.\n"); + exit (1); + } + XHSA (hsa_fns.hsa_memory_copy_fn ((void *) P, &V, size), + "Fix up relocation"); + } + } +} + +/* Allocate some device memory from the kernargs region. + The returned address will be 32-bit (with excess zeroed on 64-bit host), + and accessible via the same address on both host and target (via + __flat_scalar GCN address space). */ + +static void * +device_malloc (size_t size) +{ + void *result; + XHSA (hsa_fns.hsa_memory_allocate_fn (kernargs_region, size, &result), + "Allocate device memory"); + return result; +} + +/* These are the device pointers that will be transferred to the target. + The HSA Runtime points the kernargs register here. + They correspond to function signature: + int main (int argc, char *argv[], int *return_value) + The compiler expects this, for kernel functions, and will + automatically assign the exit value to *return_value. */ +struct kernargs +{ + /* Kernargs. */ + int32_t argc; + int64_t argv; + int64_t out_ptr; + int64_t heap_ptr; + + /* Output data. */ + struct output + { + int return_value; + int next_output; + struct printf_data + { + int written; + char msg[128]; + int type; + union + { + int64_t ivalue; + double dvalue; + char text[128]; + }; + } queue[1000]; + } output_data; + + struct heap + { + int64_t size; + char data[0]; + } heap; +}; + +/* Print any console output from the kernel. + We print all entries from print_index to the next entry without a "written" + flag. Subsequent calls should use the returned print_index value to resume + from the same point. */ +void +gomp_print_output (struct kernargs *kernargs, int *print_index) +{ + int limit = (sizeof (kernargs->output_data.queue) + / sizeof (kernargs->output_data.queue[0])); + + int i; + for (i = *print_index; i < limit; i++) + { + struct printf_data *data = &kernargs->output_data.queue[i]; + + if (!data->written) + break; + + switch (data->type) + { + case 0: + printf ("%.128s%ld\n", data->msg, data->ivalue); + break; + case 1: + printf ("%.128s%f\n", data->msg, data->dvalue); + break; + case 2: + printf ("%.128s%.128s\n", data->msg, data->text); + break; + case 3: + printf ("%.128s%.128s", data->msg, data->text); + break; + } + + data->written = 0; + } + + if (*print_index < limit && i == limit + && kernargs->output_data.next_output > limit) + printf ("WARNING: GCN print buffer exhausted.\n"); + + *print_index = i; +} + +/* Execute an already-loaded kernel on the device. */ + +static void +run (void *kernargs) +{ + /* A "signal" is used to launch and monitor the kernel. */ + hsa_signal_t signal; + XHSA (hsa_fns.hsa_signal_create_fn (1, 0, NULL, &signal), + "Create signal"); + + /* Configure for a single-worker kernel. */ + uint64_t index = hsa_fns.hsa_queue_load_write_index_relaxed_fn (queue); + const uint32_t queueMask = queue->size - 1; + hsa_kernel_dispatch_packet_t *dispatch_packet = + &(((hsa_kernel_dispatch_packet_t *) (queue->base_address))[index & + queueMask]); + dispatch_packet->setup |= 3 << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS; + dispatch_packet->workgroup_size_x = (uint16_t) 1; + dispatch_packet->workgroup_size_y = (uint16_t) 64; + dispatch_packet->workgroup_size_z = (uint16_t) 1; + dispatch_packet->grid_size_x = 1; + dispatch_packet->grid_size_y = 64; + dispatch_packet->grid_size_z = 1; + dispatch_packet->completion_signal = signal; + dispatch_packet->kernel_object = kernel; + dispatch_packet->kernarg_address = (void *) kernargs; + dispatch_packet->private_segment_size = private_segment_size; + dispatch_packet->group_segment_size = group_segment_size; + + uint16_t header = 0; + header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE; + header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE; + header |= HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE; + + __atomic_store_n ((uint32_t *) dispatch_packet, + header | (dispatch_packet->setup << 16), + __ATOMIC_RELEASE); + + if (debug) + fprintf (stderr, "Launch kernel\n"); + + hsa_fns.hsa_queue_store_write_index_relaxed_fn (queue, index + 1); + hsa_fns.hsa_signal_store_relaxed_fn (queue->doorbell_signal, index); + /* Kernel running ...... */ + int print_index = 0; + while (hsa_fns.hsa_signal_wait_relaxed_fn (signal, HSA_SIGNAL_CONDITION_LT, + 1, 1000000, + HSA_WAIT_STATE_ACTIVE) != 0) + { + usleep (10000); + gomp_print_output (kernargs, &print_index); + } + + gomp_print_output (kernargs, &print_index); + + if (debug) + fprintf (stderr, "Kernel exited\n"); + + XHSA (hsa_fns.hsa_signal_destroy_fn (signal), + "Clean up signal"); +} + +int +main (int argc, char *argv[]) +{ + int kernel_arg = 0; + for (int i = 1; i < argc; i++) + { + if (!strcmp (argv[i], "--help")) + { + usage (argv[0]); + return 0; + } + else if (!strcmp (argv[i], "--version")) + { + version (argv[0]); + return 0; + } + else if (!strcmp (argv[i], "--debug")) + debug = true; + else if (argv[i][0] == '-') + { + usage (argv[0]); + return 1; + } + else + { + kernel_arg = i; + break; + } + } + + if (!kernel_arg) + { + /* No kernel arguments were found. */ + usage (argv[0]); + return 1; + } + + /* The remaining arguments are for the GCN kernel. */ + int kernel_argc = argc - kernel_arg; + char **kernel_argv = &argv[kernel_arg]; + + init_device (); + load_image (kernel_argv[0]); + + /* Calculate size of function parameters + argv data. */ + size_t args_size = 0; + for (int i = 0; i < kernel_argc; i++) + args_size += strlen (kernel_argv[i]) + 1; + + /* Allocate device memory for both function parameters and the argv + data. */ + size_t heap_size = 10 * 1024 * 1024; /* 10MB. */ + struct kernargs *kernargs = device_malloc (sizeof (*kernargs) + heap_size); + struct argdata + { + int64_t argv_data[kernel_argc]; + char strings[args_size]; + } *args = device_malloc (sizeof (struct argdata)); + + /* Write the data to the target. */ + kernargs->argc = kernel_argc; + kernargs->argv = (int64_t) args->argv_data; + kernargs->out_ptr = (int64_t) &kernargs->output_data; + kernargs->output_data.return_value = 0xcafe0000; /* Default return value. */ + kernargs->output_data.next_output = 0; + for (unsigned i = 0; i < (sizeof (kernargs->output_data.queue) + / sizeof (kernargs->output_data.queue[0])); i++) + kernargs->output_data.queue[i].written = 0; + int offset = 0; + for (int i = 0; i < kernel_argc; i++) + { + size_t arg_len = strlen (kernel_argv[i]) + 1; + args->argv_data[i] = (int64_t) &args->strings[offset]; + memcpy (&args->strings[offset], kernel_argv[i], arg_len + 1); + offset += arg_len; + } + kernargs->heap_ptr = (int64_t) &kernargs->heap; + kernargs->heap.size = heap_size; + + /* Run the kernel on the GPU. */ + run (kernargs); + unsigned int return_value = + (unsigned int) kernargs->output_data.return_value; + + unsigned int upper = (return_value & ~0xffff) >> 16; + if (upper == 0xcafe) + printf ("Kernel exit value was never set\n"); + else if (upper == 0xffff) + ; /* Set by exit. */ + else if (upper == 0) + ; /* Set by return from main. */ + else + printf ("Possible kernel exit value corruption, 2 most significant bytes " + "aren't 0xffff, 0xcafe, or 0: 0x%x\n", return_value); + + if (upper == 0xffff) + { + unsigned int signal = (return_value >> 8) & 0xff; + if (signal == SIGABRT) + printf ("Kernel aborted\n"); + else if (signal != 0) + printf ("Kernel received unkown signal\n"); + } + + if (debug) + printf ("Kernel exit value: %d\n", return_value & 0xff); + + /* Clean shut down. */ + XHSA (hsa_fns.hsa_memory_free_fn (kernargs), + "Clean up device memory"); + XHSA (hsa_fns.hsa_executable_destroy_fn (executable), + "Clean up GCN executable"); + XHSA (hsa_fns.hsa_queue_destroy_fn (queue), + "Clean up device queue"); + XHSA (hsa_fns.hsa_shut_down_fn (), + "Shut down run-time"); + + return return_value & 0xff; +} diff --git a/gcc/config/gcn/gcn-tree.c b/gcc/config/gcn/gcn-tree.c new file mode 100644 index 00000000000..c6b6302e9ed --- /dev/null +++ b/gcc/config/gcn/gcn-tree.c @@ -0,0 +1,721 @@ +/* Copyright (C) 2017-2019 Free Software Foundation, Inc. + + This file is part of GCC. + + GCC is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3, or (at your option) any later + version. + + GCC is distributed in the hope that it will be useful, but WITHOUT ANY + WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +/* {{{ Includes. */ + +#include "config.h" +#include "system.h" +#include "coretypes.h" +#include "backend.h" +#include "target.h" +#include "tree.h" +#include "gimple.h" +#include "tree-pass.h" +#include "gimple-iterator.h" +#include "cfghooks.h" +#include "cfgloop.h" +#include "tm_p.h" +#include "stringpool.h" +#include "fold-const.h" +#include "varasm.h" +#include "omp-low.h" +#include "omp-general.h" +#include "internal-fn.h" +#include "tree-vrp.h" +#include "tree-ssanames.h" +#include "tree-ssa-operands.h" +#include "gimplify.h" +#include "tree-phinodes.h" +#include "cgraph.h" +#include "targhooks.h" +#include "langhooks-def.h" + +/* }}} */ +/* {{{ OMP GCN pass. + + This pass is intended to make any GCN-specfic transformations to OpenMP + target regions. + + At present, its only purpose is to convert some "omp" built-in functions + to use closer-to-the-metal "gcn" built-in functions. */ + +unsigned int +execute_omp_gcn (void) +{ + tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM); + tree thr_num_id = DECL_NAME (thr_num_tree); + tree team_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM); + tree team_num_id = DECL_NAME (team_num_tree); + basic_block bb; + gimple_stmt_iterator gsi; + unsigned int todo = 0; + + FOR_EACH_BB_FN (bb, cfun) + for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) + { + gimple *call = gsi_stmt (gsi); + tree decl; + + if (is_gimple_call (call) && (decl = gimple_call_fndecl (call))) + { + tree decl_id = DECL_NAME (decl); + tree lhs = gimple_get_lhs (call); + + if (decl_id == thr_num_id) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Replace '%s' with __builtin_gcn_dim_pos.\n", + IDENTIFIER_POINTER (decl_id)); + + /* Transform this: + lhs = __builtin_omp_get_thread_num () + to this: + lhs = __builtin_gcn_dim_pos (1) */ + tree fn = targetm.builtin_decl (GCN_BUILTIN_OMP_DIM_POS, 0); + tree fnarg = build_int_cst (unsigned_type_node, 1); + gimple *stmt = gimple_build_call (fn, 1, fnarg); + gimple_call_set_lhs (stmt, lhs); + gsi_replace (&gsi, stmt, true); + + todo |= TODO_update_ssa; + } + else if (decl_id == team_num_id) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Replace '%s' with __builtin_gcn_dim_pos.\n", + IDENTIFIER_POINTER (decl_id)); + + /* Transform this: + lhs = __builtin_omp_get_team_num () + to this: + lhs = __builtin_gcn_dim_pos (0) */ + tree fn = targetm.builtin_decl (GCN_BUILTIN_OMP_DIM_POS, 0); + tree fnarg = build_zero_cst (unsigned_type_node); + gimple *stmt = gimple_build_call (fn, 1, fnarg); + gimple_call_set_lhs (stmt, lhs); + gsi_replace (&gsi, stmt, true); + + todo |= TODO_update_ssa; + } + } + } + + return todo; +} + +namespace +{ + + const pass_data pass_data_omp_gcn = { + GIMPLE_PASS, + "omp_gcn", /* name */ + OPTGROUP_NONE, /* optinfo_flags */ + TV_NONE, /* tv_id */ + 0, /* properties_required */ + 0, /* properties_provided */ + 0, /* properties_destroyed */ + 0, /* todo_flags_start */ + TODO_df_finish, /* todo_flags_finish */ + }; + + class pass_omp_gcn : public gimple_opt_pass + { + public: + pass_omp_gcn (gcc::context *ctxt) + : gimple_opt_pass (pass_data_omp_gcn, ctxt) + { + } + + /* opt_pass methods: */ + virtual bool gate (function *) + { + return flag_openmp; + } + + virtual unsigned int execute (function *) + { + return execute_omp_gcn (); + } + + }; /* class pass_omp_gcn. */ + +} /* anon namespace. */ + +gimple_opt_pass * +make_pass_omp_gcn (gcc::context *ctxt) +{ + return new pass_omp_gcn (ctxt); +} + +/* }}} */ +/* {{{ OpenACC reductions. */ + +/* Global lock variable, needed for 128bit worker & gang reductions. */ + +static GTY(()) tree global_lock_var; + +/* Lazily generate the global_lock_var decl and return its address. */ + +static tree +gcn_global_lock_addr () +{ + tree v = global_lock_var; + + if (!v) + { + tree name = get_identifier ("__reduction_lock"); + tree type = build_qualified_type (unsigned_type_node, + TYPE_QUAL_VOLATILE); + v = build_decl (BUILTINS_LOCATION, VAR_DECL, name, type); + global_lock_var = v; + DECL_ARTIFICIAL (v) = 1; + DECL_EXTERNAL (v) = 1; + TREE_STATIC (v) = 1; + TREE_PUBLIC (v) = 1; + TREE_USED (v) = 1; + mark_addressable (v); + mark_decl_referenced (v); + } + + return build_fold_addr_expr (v); +} + +/* Helper function for gcn_reduction_update. + + Insert code to locklessly update *PTR with *PTR OP VAR just before + GSI. We use a lockless scheme for nearly all case, which looks + like: + actual = initval (OP); + do { + guess = actual; + write = guess OP myval; + actual = cmp&swap (ptr, guess, write) + } while (actual bit-different-to guess); + return write; + + This relies on a cmp&swap instruction, which is available for 32- and + 64-bit types. Larger types must use a locking scheme. */ + +static tree +gcn_lockless_update (location_t loc, gimple_stmt_iterator *gsi, + tree ptr, tree var, tree_code op) +{ + unsigned fn = GCN_BUILTIN_CMP_SWAP; + tree_code code = NOP_EXPR; + tree arg_type = unsigned_type_node; + tree var_type = TREE_TYPE (var); + + if (TREE_CODE (var_type) == COMPLEX_TYPE + || TREE_CODE (var_type) == REAL_TYPE) + code = VIEW_CONVERT_EXPR; + + if (TYPE_SIZE (var_type) == TYPE_SIZE (long_long_unsigned_type_node)) + { + arg_type = long_long_unsigned_type_node; + fn = GCN_BUILTIN_CMP_SWAPLL; + } + + tree swap_fn = gcn_builtin_decl (fn, true); + + gimple_seq init_seq = NULL; + tree init_var = make_ssa_name (arg_type); + tree init_expr = omp_reduction_init_op (loc, op, var_type); + init_expr = fold_build1 (code, arg_type, init_expr); + gimplify_assign (init_var, init_expr, &init_seq); + gimple *init_end = gimple_seq_last (init_seq); + + gsi_insert_seq_before (gsi, init_seq, GSI_SAME_STMT); + + /* Split the block just after the init stmts. */ + basic_block pre_bb = gsi_bb (*gsi); + edge pre_edge = split_block (pre_bb, init_end); + basic_block loop_bb = pre_edge->dest; + pre_bb = pre_edge->src; + /* Reset the iterator. */ + *gsi = gsi_for_stmt (gsi_stmt (*gsi)); + + tree expect_var = make_ssa_name (arg_type); + tree actual_var = make_ssa_name (arg_type); + tree write_var = make_ssa_name (arg_type); + + /* Build and insert the reduction calculation. */ + gimple_seq red_seq = NULL; + tree write_expr = fold_build1 (code, var_type, expect_var); + write_expr = fold_build2 (op, var_type, write_expr, var); + write_expr = fold_build1 (code, arg_type, write_expr); + gimplify_assign (write_var, write_expr, &red_seq); + + gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); + + /* Build & insert the cmp&swap sequence. */ + gimple_seq latch_seq = NULL; + tree swap_expr = build_call_expr_loc (loc, swap_fn, 3, + ptr, expect_var, write_var); + gimplify_assign (actual_var, swap_expr, &latch_seq); + + gcond *cond = gimple_build_cond (EQ_EXPR, actual_var, expect_var, + NULL_TREE, NULL_TREE); + gimple_seq_add_stmt (&latch_seq, cond); + + gimple *latch_end = gimple_seq_last (latch_seq); + gsi_insert_seq_before (gsi, latch_seq, GSI_SAME_STMT); + + /* Split the block just after the latch stmts. */ + edge post_edge = split_block (loop_bb, latch_end); + basic_block post_bb = post_edge->dest; + loop_bb = post_edge->src; + *gsi = gsi_for_stmt (gsi_stmt (*gsi)); + + post_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; + /* post_edge->probability = profile_probability::even (); */ + edge loop_edge = make_edge (loop_bb, loop_bb, EDGE_FALSE_VALUE); + /* loop_edge->probability = profile_probability::even (); */ + set_immediate_dominator (CDI_DOMINATORS, loop_bb, pre_bb); + set_immediate_dominator (CDI_DOMINATORS, post_bb, loop_bb); + + gphi *phi = create_phi_node (expect_var, loop_bb); + add_phi_arg (phi, init_var, pre_edge, loc); + add_phi_arg (phi, actual_var, loop_edge, loc); + + loop *loop = alloc_loop (); + loop->header = loop_bb; + loop->latch = loop_bb; + add_loop (loop, loop_bb->loop_father); + + return fold_build1 (code, var_type, write_var); +} + +/* Helper function for gcn_reduction_update. + + Insert code to lockfully update *PTR with *PTR OP VAR just before + GSI. This is necessary for types larger than 64 bits, where there + is no cmp&swap instruction to implement a lockless scheme. We use + a lock variable in global memory. + + while (cmp&swap (&lock_var, 0, 1)) + continue; + T accum = *ptr; + accum = accum OP var; + *ptr = accum; + cmp&swap (&lock_var, 1, 0); + return accum; + + A lock in global memory is necessary to force execution engine + descheduling and avoid resource starvation that can occur if the + lock is in shared memory. */ + +static tree +gcn_lockfull_update (location_t loc, gimple_stmt_iterator *gsi, + tree ptr, tree var, tree_code op) +{ + tree var_type = TREE_TYPE (var); + tree swap_fn = gcn_builtin_decl (GCN_BUILTIN_CMP_SWAP, true); + tree uns_unlocked = build_int_cst (unsigned_type_node, 0); + tree uns_locked = build_int_cst (unsigned_type_node, 1); + + /* Split the block just before the gsi. Insert a gimple nop to make + this easier. */ + gimple *nop = gimple_build_nop (); + gsi_insert_before (gsi, nop, GSI_SAME_STMT); + basic_block entry_bb = gsi_bb (*gsi); + edge entry_edge = split_block (entry_bb, nop); + basic_block lock_bb = entry_edge->dest; + /* Reset the iterator. */ + *gsi = gsi_for_stmt (gsi_stmt (*gsi)); + + /* Build and insert the locking sequence. */ + gimple_seq lock_seq = NULL; + tree lock_var = make_ssa_name (unsigned_type_node); + tree lock_expr = gcn_global_lock_addr (); + lock_expr = build_call_expr_loc (loc, swap_fn, 3, lock_expr, + uns_unlocked, uns_locked); + gimplify_assign (lock_var, lock_expr, &lock_seq); + gcond *cond = gimple_build_cond (EQ_EXPR, lock_var, uns_unlocked, + NULL_TREE, NULL_TREE); + gimple_seq_add_stmt (&lock_seq, cond); + gimple *lock_end = gimple_seq_last (lock_seq); + gsi_insert_seq_before (gsi, lock_seq, GSI_SAME_STMT); + + /* Split the block just after the lock sequence. */ + edge locked_edge = split_block (lock_bb, lock_end); + basic_block update_bb = locked_edge->dest; + lock_bb = locked_edge->src; + *gsi = gsi_for_stmt (gsi_stmt (*gsi)); + + /* Create the lock loop. */ + locked_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; + locked_edge->probability = profile_probability::even (); + edge loop_edge = make_edge (lock_bb, lock_bb, EDGE_FALSE_VALUE); + loop_edge->probability = profile_probability::even (); + set_immediate_dominator (CDI_DOMINATORS, lock_bb, entry_bb); + set_immediate_dominator (CDI_DOMINATORS, update_bb, lock_bb); + + /* Create the loop structure. */ + loop *lock_loop = alloc_loop (); + lock_loop->header = lock_bb; + lock_loop->latch = lock_bb; + lock_loop->nb_iterations_estimate = 1; + lock_loop->any_estimate = true; + add_loop (lock_loop, entry_bb->loop_father); + + /* Build and insert the reduction calculation. */ + gimple_seq red_seq = NULL; + tree acc_in = make_ssa_name (var_type); + tree ref_in = build_simple_mem_ref (ptr); + TREE_THIS_VOLATILE (ref_in) = 1; + gimplify_assign (acc_in, ref_in, &red_seq); + + tree acc_out = make_ssa_name (var_type); + tree update_expr = fold_build2 (op, var_type, ref_in, var); + gimplify_assign (acc_out, update_expr, &red_seq); + + tree ref_out = build_simple_mem_ref (ptr); + TREE_THIS_VOLATILE (ref_out) = 1; + gimplify_assign (ref_out, acc_out, &red_seq); + + gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); + + /* Build & insert the unlock sequence. */ + gimple_seq unlock_seq = NULL; + tree unlock_expr = gcn_global_lock_addr (); + unlock_expr = build_call_expr_loc (loc, swap_fn, 3, unlock_expr, + uns_locked, uns_unlocked); + gimplify_and_add (unlock_expr, &unlock_seq); + gsi_insert_seq_before (gsi, unlock_seq, GSI_SAME_STMT); + + return acc_out; +} + +/* Emit a sequence to update a reduction accumulator at *PTR with the + value held in VAR using operator OP. Return the updated value. + + TODO: optimize for atomic ops and independent complex ops. */ + +static tree +gcn_reduction_update (location_t loc, gimple_stmt_iterator *gsi, + tree ptr, tree var, tree_code op) +{ + tree type = TREE_TYPE (var); + tree size = TYPE_SIZE (type); + + if (size == TYPE_SIZE (unsigned_type_node) + || size == TYPE_SIZE (long_long_unsigned_type_node)) + return gcn_lockless_update (loc, gsi, ptr, var, op); + else + return gcn_lockfull_update (loc, gsi, ptr, var, op); +} + +/* Return a temporary variable decl to use for an OpenACC worker reduction. */ + +static tree +gcn_goacc_get_worker_red_decl (tree type, unsigned offset) +{ + machine_function *machfun = cfun->machine; + tree existing_decl; + + if (TREE_CODE (type) == REFERENCE_TYPE) + type = TREE_TYPE (type); + + tree var_type + = build_qualified_type (type, + (TYPE_QUALS (type) + | ENCODE_QUAL_ADDR_SPACE (ADDR_SPACE_LDS))); + + if (machfun->reduc_decls + && offset < machfun->reduc_decls->length () + && (existing_decl = (*machfun->reduc_decls)[offset])) + { + gcc_assert (TREE_TYPE (existing_decl) == var_type); + return existing_decl; + } + else + { + char name[50]; + sprintf (name, ".oacc_reduction_%u", offset); + tree decl = create_tmp_var_raw (var_type, name); + + DECL_CONTEXT (decl) = NULL_TREE; + TREE_STATIC (decl) = 1; + + varpool_node::finalize_decl (decl); + + vec_safe_grow_cleared (machfun->reduc_decls, offset + 1); + (*machfun->reduc_decls)[offset] = decl; + + return decl; + } + + return NULL_TREE; +} + +/* Expand IFN_GOACC_REDUCTION_SETUP. */ + +static void +gcn_goacc_reduction_setup (gcall *call) +{ + gimple_stmt_iterator gsi = gsi_for_stmt (call); + tree lhs = gimple_call_lhs (call); + tree var = gimple_call_arg (call, 2); + int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); + gimple_seq seq = NULL; + + push_gimplify_context (true); + + if (level != GOMP_DIM_GANG) + { + /* Copy the receiver object. */ + tree ref_to_res = gimple_call_arg (call, 1); + + if (!integer_zerop (ref_to_res)) + var = build_simple_mem_ref (ref_to_res); + } + + if (level == GOMP_DIM_WORKER) + { + tree var_type = TREE_TYPE (var); + /* Store incoming value to worker reduction buffer. */ + tree offset = gimple_call_arg (call, 5); + tree decl + = gcn_goacc_get_worker_red_decl (var_type, TREE_INT_CST_LOW (offset)); + + gimplify_assign (decl, var, &seq); + } + + if (lhs) + gimplify_assign (lhs, var, &seq); + + pop_gimplify_context (NULL); + gsi_replace_with_seq (&gsi, seq, true); +} + +/* Expand IFN_GOACC_REDUCTION_INIT. */ + +static void +gcn_goacc_reduction_init (gcall *call) +{ + gimple_stmt_iterator gsi = gsi_for_stmt (call); + tree lhs = gimple_call_lhs (call); + tree var = gimple_call_arg (call, 2); + int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); + enum tree_code rcode + = (enum tree_code) TREE_INT_CST_LOW (gimple_call_arg (call, 4)); + tree init = omp_reduction_init_op (gimple_location (call), rcode, + TREE_TYPE (var)); + gimple_seq seq = NULL; + + push_gimplify_context (true); + + if (level == GOMP_DIM_GANG) + { + /* If there's no receiver object, propagate the incoming VAR. */ + tree ref_to_res = gimple_call_arg (call, 1); + if (integer_zerop (ref_to_res)) + init = var; + } + + if (lhs) + gimplify_assign (lhs, init, &seq); + + pop_gimplify_context (NULL); + gsi_replace_with_seq (&gsi, seq, true); +} + +/* Expand IFN_GOACC_REDUCTION_FINI. */ + +static void +gcn_goacc_reduction_fini (gcall *call) +{ + gimple_stmt_iterator gsi = gsi_for_stmt (call); + tree lhs = gimple_call_lhs (call); + tree ref_to_res = gimple_call_arg (call, 1); + tree var = gimple_call_arg (call, 2); + int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); + enum tree_code op + = (enum tree_code) TREE_INT_CST_LOW (gimple_call_arg (call, 4)); + gimple_seq seq = NULL; + tree r = NULL_TREE;; + + push_gimplify_context (true); + + tree accum = NULL_TREE; + + if (level == GOMP_DIM_WORKER) + { + tree var_type = TREE_TYPE (var); + tree offset = gimple_call_arg (call, 5); + tree decl + = gcn_goacc_get_worker_red_decl (var_type, TREE_INT_CST_LOW (offset)); + + accum = build_fold_addr_expr (decl); + } + else if (integer_zerop (ref_to_res)) + r = var; + else + accum = ref_to_res; + + if (accum) + { + /* UPDATE the accumulator. */ + gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); + seq = NULL; + r = gcn_reduction_update (gimple_location (call), &gsi, accum, var, op); + } + + if (lhs) + gimplify_assign (lhs, r, &seq); + pop_gimplify_context (NULL); + + gsi_replace_with_seq (&gsi, seq, true); +} + +/* Expand IFN_GOACC_REDUCTION_TEARDOWN. */ + +static void +gcn_goacc_reduction_teardown (gcall *call) +{ + gimple_stmt_iterator gsi = gsi_for_stmt (call); + tree lhs = gimple_call_lhs (call); + tree var = gimple_call_arg (call, 2); + int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); + gimple_seq seq = NULL; + + push_gimplify_context (true); + + if (level == GOMP_DIM_WORKER) + { + tree var_type = TREE_TYPE (var); + + /* Read the worker reduction buffer. */ + tree offset = gimple_call_arg (call, 5); + tree decl + = gcn_goacc_get_worker_red_decl (var_type, TREE_INT_CST_LOW (offset)); + var = decl; + } + + if (level != GOMP_DIM_GANG) + { + /* Write to the receiver object. */ + tree ref_to_res = gimple_call_arg (call, 1); + + if (!integer_zerop (ref_to_res)) + gimplify_assign (build_simple_mem_ref (ref_to_res), var, &seq); + } + + if (lhs) + gimplify_assign (lhs, var, &seq); + + pop_gimplify_context (NULL); + + gsi_replace_with_seq (&gsi, seq, true); +} + +/* Implement TARGET_GOACC_REDUCTION. + + Expand calls to the GOACC REDUCTION internal function, into a sequence of + gimple instructions. */ + +void +gcn_goacc_reduction (gcall *call) +{ + int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); + + if (level == GOMP_DIM_VECTOR) + { + default_goacc_reduction (call); + return; + } + + unsigned code = (unsigned) TREE_INT_CST_LOW (gimple_call_arg (call, 0)); + + switch (code) + { + case IFN_GOACC_REDUCTION_SETUP: + gcn_goacc_reduction_setup (call); + break; + + case IFN_GOACC_REDUCTION_INIT: + gcn_goacc_reduction_init (call); + break; + + case IFN_GOACC_REDUCTION_FINI: + gcn_goacc_reduction_fini (call); + break; + + case IFN_GOACC_REDUCTION_TEARDOWN: + gcn_goacc_reduction_teardown (call); + break; + + default: + gcc_unreachable (); + } +} + +/* Implement TARGET_GOACC_ADJUST_PROPAGATION_RECORD. + + Tweak (worker) propagation record, e.g. to put it in shared memory. */ + +tree +gcn_goacc_adjust_propagation_record (tree record_type, bool sender, + const char *name) +{ + tree type = record_type; + + TYPE_ADDR_SPACE (type) = ADDR_SPACE_LDS; + + if (!sender) + type = build_pointer_type (type); + + tree decl = create_tmp_var_raw (type, name); + + if (sender) + { + DECL_CONTEXT (decl) = NULL_TREE; + TREE_STATIC (decl) = 1; + } + + if (sender) + varpool_node::finalize_decl (decl); + + return decl; +} + +void +gcn_goacc_adjust_gangprivate_decl (tree var) +{ + tree type = TREE_TYPE (var); + tree lds_type = build_qualified_type (type, + TYPE_QUALS_NO_ADDR_SPACE (type) + | ENCODE_QUAL_ADDR_SPACE (ADDR_SPACE_LDS)); + machine_function *machfun = cfun->machine; + + TREE_TYPE (var) = lds_type; + TREE_STATIC (var) = 1; + + /* We're making VAR static. We have to mangle the name to avoid collisions + between different local variables that share the same names. */ + lhd_set_decl_assembler_name (var); + + varpool_node::finalize_decl (var); + + if (machfun) + machfun->use_flat_addressing = true; +} + +/* }}} */ diff --git a/gcc/config/gcn/gcn.c b/gcc/config/gcn/gcn.c new file mode 100644 index 00000000000..bd8ea55ec03 --- /dev/null +++ b/gcc/config/gcn/gcn.c @@ -0,0 +1,6123 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +/* {{{ Includes. */ + +/* We want GET_MODE_SIZE et al to return integers, please. */ +#define IN_TARGET_CODE 1 + +#include "config.h" +#include "system.h" +#include "coretypes.h" +#include "backend.h" +#include "target.h" +#include "memmodel.h" +#include "rtl.h" +#include "tree.h" +#include "df.h" +#include "tm_p.h" +#include "stringpool.h" +#include "optabs.h" +#include "regs.h" +#include "emit-rtl.h" +#include "recog.h" +#include "diagnostic-core.h" +#include "insn-attr.h" +#include "fold-const.h" +#include "calls.h" +#include "explow.h" +#include "expr.h" +#include "output.h" +#include "cfgrtl.h" +#include "langhooks.h" +#include "builtins.h" +#include "omp-general.h" +#include "print-rtl.h" +#include "attribs.h" +#include "varasm.h" +#include "intl.h" +#include "rtl-iter.h" + +/* This file should be included last. */ +#include "target-def.h" + +/* }}} */ +/* {{{ Global variables. */ + +/* Constants used by FP instructions. */ + +static REAL_VALUE_TYPE dconst4, dconst1over2pi; +static bool ext_gcn_constants_init = 0; + +/* Holds the ISA variant, derived from the command line parameters. */ + +int gcn_isa = 3; /* Default to GCN3. */ + +/* Reserve this much space for LDS (for propagating variables from + worker-single mode to worker-partitioned mode), per workgroup. Global + analysis could calculate an exact bound, but we don't do that yet. + + We reserve the whole LDS, which also prevents any other workgroup + sharing the Compute Unit. */ + +#define LDS_SIZE 65536 + +/* }}} */ +/* {{{ Initialization and options. */ + +/* Initialize machine_function. */ + +static struct machine_function * +gcn_init_machine_status (void) +{ + struct machine_function *f; + + f = ggc_cleared_alloc (); + + /* Set up LDS allocation for broadcasting for this function. */ + f->lds_allocated = 32; + f->lds_allocs = hash_map::create_ggc (64); + + /* And LDS temporary decls for worker reductions. */ + vec_alloc (f->reduc_decls, 0); + + if (TARGET_GCN3) + f->use_flat_addressing = true; + + return f; +} + +/* Implement TARGET_OPTION_OVERRIDE. + + Override option settings where defaults are variable, or we have specific + needs to consider. */ + +static void +gcn_option_override (void) +{ + init_machine_status = gcn_init_machine_status; + + /* The HSA runtime does not respect ELF load addresses, so force PIE. */ + if (!flag_pie) + flag_pie = 2; + if (!flag_pic) + flag_pic = flag_pie; + + gcn_isa = gcn_arch == PROCESSOR_VEGA ? 5 : 3; + + /* The default stack size needs to be small for offload kernels because + there may be many, many threads. Also, a smaller stack gives a + measureable performance boost. But, a small stack is insufficient + for running the testsuite, so we use a larger default for the stand + alone case. */ + if (stack_size_opt == -1) + { + if (flag_openacc || flag_openmp) + /* 512 bytes per work item = 32kB total. */ + stack_size_opt = 512 * 64; + else + /* 1MB total. */ + stack_size_opt = 1048576; + } +} + +/* }}} */ +/* {{{ Attributes. */ + +/* This table defines the arguments that are permitted in + __attribute__ ((amdgpu_hsa_kernel (...))). + + The names and values correspond to the HSA metadata that is encoded + into the assembler file and binary. */ + +static const struct gcn_kernel_arg_type +{ + const char *name; + const char *header_pseudo; + machine_mode mode; + + /* This should be set to -1 or -2 for a dynamically allocated register + number. Use -1 if this argument contributes to the user_sgpr_count, + -2 otherwise. */ + int fixed_regno; +} gcn_kernel_arg_types[] = { + {"exec", NULL, DImode, EXEC_REG}, +#define PRIVATE_SEGMENT_BUFFER_ARG 1 + {"private_segment_buffer", + "enable_sgpr_private_segment_buffer", TImode, -1}, +#define DISPATCH_PTR_ARG 2 + {"dispatch_ptr", "enable_sgpr_dispatch_ptr", DImode, -1}, +#define QUEUE_PTR_ARG 3 + {"queue_ptr", "enable_sgpr_queue_ptr", DImode, -1}, +#define KERNARG_SEGMENT_PTR_ARG 4 + {"kernarg_segment_ptr", "enable_sgpr_kernarg_segment_ptr", DImode, -1}, + {"dispatch_id", "enable_sgpr_dispatch_id", DImode, -1}, +#define FLAT_SCRATCH_INIT_ARG 6 + {"flat_scratch_init", "enable_sgpr_flat_scratch_init", DImode, -1}, +#define FLAT_SCRATCH_SEGMENT_SIZE_ARG 7 + {"private_segment_size", "enable_sgpr_private_segment_size", SImode, -1}, + {"grid_workgroup_count_X", + "enable_sgpr_grid_workgroup_count_x", SImode, -1}, + {"grid_workgroup_count_Y", + "enable_sgpr_grid_workgroup_count_y", SImode, -1}, + {"grid_workgroup_count_Z", + "enable_sgpr_grid_workgroup_count_z", SImode, -1}, +#define WORKGROUP_ID_X_ARG 11 + {"workgroup_id_X", "enable_sgpr_workgroup_id_x", SImode, -2}, + {"workgroup_id_Y", "enable_sgpr_workgroup_id_y", SImode, -2}, + {"workgroup_id_Z", "enable_sgpr_workgroup_id_z", SImode, -2}, + {"workgroup_info", "enable_sgpr_workgroup_info", SImode, -1}, +#define PRIVATE_SEGMENT_WAVE_OFFSET_ARG 15 + {"private_segment_wave_offset", + "enable_sgpr_private_segment_wave_byte_offset", SImode, -2}, +#define WORK_ITEM_ID_X_ARG 16 + {"work_item_id_X", NULL, V64SImode, FIRST_VGPR_REG}, +#define WORK_ITEM_ID_Y_ARG 17 + {"work_item_id_Y", NULL, V64SImode, FIRST_VGPR_REG + 1}, +#define WORK_ITEM_ID_Z_ARG 18 + {"work_item_id_Z", NULL, V64SImode, FIRST_VGPR_REG + 2} +}; + +/* Extract parameter settings from __attribute__((amdgpu_hsa_kernel ())). + This function also sets the default values for some arguments. + + Return true on success, with ARGS populated. */ + +static bool +gcn_parse_amdgpu_hsa_kernel_attribute (struct gcn_kernel_args *args, + tree list) +{ + bool err = false; + args->requested = ((1 << PRIVATE_SEGMENT_BUFFER_ARG) + | (1 << QUEUE_PTR_ARG) + | (1 << KERNARG_SEGMENT_PTR_ARG) + | (1 << PRIVATE_SEGMENT_WAVE_OFFSET_ARG)); + args->nargs = 0; + + for (int a = 0; a < GCN_KERNEL_ARG_TYPES; a++) + args->reg[a] = -1; + + for (; list; list = TREE_CHAIN (list)) + { + const char *str; + if (TREE_CODE (TREE_VALUE (list)) != STRING_CST) + { + error ("amdgpu_hsa_kernel attribute requires string constant " + "arguments"); + break; + } + str = TREE_STRING_POINTER (TREE_VALUE (list)); + int a; + for (a = 0; a < GCN_KERNEL_ARG_TYPES; a++) + { + if (!strcmp (str, gcn_kernel_arg_types[a].name)) + break; + } + if (a == GCN_KERNEL_ARG_TYPES) + { + error ("unknown specifier %s in amdgpu_hsa_kernel attribute", str); + err = true; + break; + } + if (args->requested & (1 << a)) + { + error ("duplicated parameter specifier %s in amdgpu_hsa_kernel " + "attribute", str); + err = true; + break; + } + args->requested |= (1 << a); + args->order[args->nargs++] = a; + } + args->requested |= (1 << WORKGROUP_ID_X_ARG); + args->requested |= (1 << WORK_ITEM_ID_Z_ARG); + + /* Requesting WORK_ITEM_ID_Z_ARG implies requesting WORK_ITEM_ID_X_ARG and + WORK_ITEM_ID_Y_ARG. Similarly, requesting WORK_ITEM_ID_Y_ARG implies + requesting WORK_ITEM_ID_X_ARG. */ + if (args->requested & (1 << WORK_ITEM_ID_Z_ARG)) + args->requested |= (1 << WORK_ITEM_ID_Y_ARG); + if (args->requested & (1 << WORK_ITEM_ID_Y_ARG)) + args->requested |= (1 << WORK_ITEM_ID_X_ARG); + + /* Always enable this so that kernargs is in a predictable place for + gomp_print, etc. */ + args->requested |= (1 << DISPATCH_PTR_ARG); + + int sgpr_regno = FIRST_SGPR_REG; + args->nsgprs = 0; + for (int a = 0; a < GCN_KERNEL_ARG_TYPES; a++) + { + if (!(args->requested & (1 << a))) + continue; + + if (gcn_kernel_arg_types[a].fixed_regno >= 0) + args->reg[a] = gcn_kernel_arg_types[a].fixed_regno; + else + { + int reg_count; + + switch (gcn_kernel_arg_types[a].mode) + { + case E_SImode: + reg_count = 1; + break; + case E_DImode: + reg_count = 2; + break; + case E_TImode: + reg_count = 4; + break; + default: + gcc_unreachable (); + } + args->reg[a] = sgpr_regno; + sgpr_regno += reg_count; + if (gcn_kernel_arg_types[a].fixed_regno == -1) + args->nsgprs += reg_count; + } + } + if (sgpr_regno > FIRST_SGPR_REG + 16) + { + error ("too many arguments passed in sgpr registers"); + } + return err; +} + +/* Referenced by TARGET_ATTRIBUTE_TABLE. + + Validates target specific attributes. */ + +static tree +gcn_handle_amdgpu_hsa_kernel_attribute (tree *node, tree name, + tree args, int, bool *no_add_attrs) +{ + if (FUNC_OR_METHOD_TYPE_P (*node) + && TREE_CODE (*node) != FIELD_DECL + && TREE_CODE (*node) != TYPE_DECL) + { + warning (OPT_Wattributes, "%qE attribute only applies to functions", + name); + *no_add_attrs = true; + return NULL_TREE; + } + + /* Can combine regparm with all attributes but fastcall, and thiscall. */ + if (is_attribute_p ("gcnhsa_kernel", name)) + { + struct gcn_kernel_args kernelarg; + + if (gcn_parse_amdgpu_hsa_kernel_attribute (&kernelarg, args)) + *no_add_attrs = true; + + return NULL_TREE; + } + + return NULL_TREE; +} + +/* Implement TARGET_ATTRIBUTE_TABLE. + + Create target-specific __attribute__ types. */ + +static const struct attribute_spec gcn_attribute_table[] = { + /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler, + affects_type_identity } */ + {"amdgpu_hsa_kernel", 0, GCN_KERNEL_ARG_TYPES, false, true, + true, true, gcn_handle_amdgpu_hsa_kernel_attribute, NULL}, + /* End element. */ + {NULL, 0, 0, false, false, false, false, NULL, NULL} +}; + +/* }}} */ +/* {{{ Registers and modes. */ + +/* Implement TARGET_CLASS_MAX_NREGS. + + Return the number of hard registers needed to hold a value of MODE in + a register of class RCLASS. */ + +static unsigned char +gcn_class_max_nregs (reg_class_t rclass, machine_mode mode) +{ + /* Scalar registers are 32bit, vector registers are in fact tuples of + 64 lanes. */ + if (rclass == VGPR_REGS) + { + if (vgpr_1reg_mode_p (mode)) + return 1; + if (vgpr_2reg_mode_p (mode)) + return 2; + /* TImode is used by DImode compare_and_swap. */ + if (mode == TImode) + return 4; + } + else if (rclass == VCC_CONDITIONAL_REG && mode == BImode) + return 2; + return CEIL (GET_MODE_SIZE (mode), 4); +} + +/* Implement TARGET_HARD_REGNO_NREGS. + + Return the number of hard registers needed to hold a value of MODE in + REGNO. */ + +unsigned int +gcn_hard_regno_nregs (unsigned int regno, machine_mode mode) +{ + return gcn_class_max_nregs (REGNO_REG_CLASS (regno), mode); +} + +/* Implement TARGET_HARD_REGNO_MODE_OK. + + Return true if REGNO can hold value in MODE. */ + +bool +gcn_hard_regno_mode_ok (unsigned int regno, machine_mode mode) +{ + /* Treat a complex mode as if it were a scalar mode of the same overall + size for the purposes of allocating hard registers. */ + if (COMPLEX_MODE_P (mode)) + switch (mode) + { + case E_CQImode: + case E_CHImode: + mode = SImode; + break; + case E_CSImode: + mode = DImode; + break; + case E_CDImode: + mode = TImode; + break; + case E_HCmode: + mode = SFmode; + break; + case E_SCmode: + mode = DFmode; + break; + default: + /* Not supported. */ + return false; + } + + switch (regno) + { + case FLAT_SCRATCH_LO_REG: + case XNACK_MASK_LO_REG: + case TBA_LO_REG: + case TMA_LO_REG: + return (mode == SImode || mode == DImode); + case VCC_LO_REG: + case EXEC_LO_REG: + return (mode == BImode || mode == SImode || mode == DImode); + case M0_REG: + case FLAT_SCRATCH_HI_REG: + case XNACK_MASK_HI_REG: + case TBA_HI_REG: + case TMA_HI_REG: + return mode == SImode; + case VCC_HI_REG: + return false; + case EXEC_HI_REG: + return mode == SImode /*|| mode == V32BImode */ ; + case SCC_REG: + case VCCZ_REG: + case EXECZ_REG: + return mode == BImode; + } + if (regno == ARG_POINTER_REGNUM || regno == FRAME_POINTER_REGNUM) + return true; + if (SGPR_REGNO_P (regno)) + /* We restrict double register values to aligned registers. */ + return (sgpr_1reg_mode_p (mode) + || (!((regno - FIRST_SGPR_REG) & 1) && sgpr_2reg_mode_p (mode)) + || (((regno - FIRST_SGPR_REG) & 3) == 0 && mode == TImode)); + if (VGPR_REGNO_P (regno)) + return (vgpr_1reg_mode_p (mode) || vgpr_2reg_mode_p (mode) + /* TImode is used by DImode compare_and_swap. */ + || mode == TImode); + return false; +} + +/* Implement REGNO_REG_CLASS via gcn.h. + + Return smallest class containing REGNO. */ + +enum reg_class +gcn_regno_reg_class (int regno) +{ + switch (regno) + { + case SCC_REG: + return SCC_CONDITIONAL_REG; + case VCCZ_REG: + return VCCZ_CONDITIONAL_REG; + case EXECZ_REG: + return EXECZ_CONDITIONAL_REG; + case EXEC_LO_REG: + case EXEC_HI_REG: + return EXEC_MASK_REG; + } + if (VGPR_REGNO_P (regno)) + return VGPR_REGS; + if (SGPR_REGNO_P (regno)) + return SGPR_REGS; + if (regno < FIRST_VGPR_REG) + return GENERAL_REGS; + if (regno == ARG_POINTER_REGNUM || regno == FRAME_POINTER_REGNUM) + return AFP_REGS; + return ALL_REGS; +} + +/* Implement TARGET_CAN_CHANGE_MODE_CLASS. + + GCC assumes that lowpart contains first part of value as stored in memory. + This is not the case for vector registers. */ + +bool +gcn_can_change_mode_class (machine_mode from, machine_mode to, + reg_class_t regclass) +{ + if (!vgpr_vector_mode_p (from) && !vgpr_vector_mode_p (to)) + return true; + return (gcn_class_max_nregs (regclass, from) + == gcn_class_max_nregs (regclass, to)); +} + +/* Implement TARGET_SMALL_REGISTER_CLASSES_FOR_MODE_P. + + When this hook returns true for MODE, the compiler allows + registers explicitly used in the rtl to be used as spill registers + but prevents the compiler from extending the lifetime of these + registers. */ + +bool +gcn_small_register_classes_for_mode_p (machine_mode mode) +{ + /* We allocate into exec and vcc regs. Those make small register class. */ + return mode == DImode || mode == SImode; +} + +/* Implement TARGET_CLASS_LIKELY_SPILLED_P. + + Returns true if pseudos that have been assigned to registers of class RCLASS + would likely be spilled because registers of RCLASS are needed for spill + registers. */ + +static bool +gcn_class_likely_spilled_p (reg_class_t rclass) +{ + return (rclass == EXEC_MASK_REG + || reg_classes_intersect_p (ALL_CONDITIONAL_REGS, rclass)); +} + +/* Implement TARGET_MODES_TIEABLE_P. + + Returns true if a value of MODE1 is accessible in MODE2 without + copying. */ + +bool +gcn_modes_tieable_p (machine_mode mode1, machine_mode mode2) +{ + return (GET_MODE_BITSIZE (mode1) <= MAX_FIXED_MODE_SIZE + && GET_MODE_BITSIZE (mode2) <= MAX_FIXED_MODE_SIZE); +} + +/* Implement TARGET_TRULY_NOOP_TRUNCATION. + + Returns true if it is safe to “convert” a value of INPREC bits to one of + OUTPREC bits (where OUTPREC is smaller than INPREC) by merely operating on + it as if it had only OUTPREC bits. */ + +bool +gcn_truly_noop_truncation (poly_uint64 outprec, poly_uint64 inprec) +{ + return ((inprec <= 32) && (outprec <= inprec)); +} + +/* Return N-th part of value occupying multiple registers. */ + +rtx +gcn_operand_part (machine_mode mode, rtx op, int n) +{ + if (GET_MODE_SIZE (mode) >= 256) + { + /*gcc_assert (GET_MODE_SIZE (mode) == 256 || n == 0); */ + + if (REG_P (op)) + { + gcc_assert (REGNO (op) + n < FIRST_PSEUDO_REGISTER); + return gen_rtx_REG (V64SImode, REGNO (op) + n); + } + if (GET_CODE (op) == CONST_VECTOR) + { + int units = GET_MODE_NUNITS (mode); + rtvec v = rtvec_alloc (units); + + for (int i = 0; i < units; ++i) + RTVEC_ELT (v, i) = gcn_operand_part (GET_MODE_INNER (mode), + CONST_VECTOR_ELT (op, i), n); + + return gen_rtx_CONST_VECTOR (V64SImode, v); + } + if (GET_CODE (op) == UNSPEC && XINT (op, 1) == UNSPEC_VECTOR) + return gcn_gen_undef (V64SImode); + gcc_unreachable (); + } + else if (GET_MODE_SIZE (mode) == 8 && REG_P (op)) + { + gcc_assert (REGNO (op) + n < FIRST_PSEUDO_REGISTER); + return gen_rtx_REG (SImode, REGNO (op) + n); + } + else + { + if (GET_CODE (op) == UNSPEC && XINT (op, 1) == UNSPEC_VECTOR) + return gcn_gen_undef (SImode); + + /* If it's a constant then let's assume it is of the largest mode + available, otherwise simplify_gen_subreg will fail. */ + if (mode == VOIDmode && CONST_INT_P (op)) + mode = DImode; + return simplify_gen_subreg (SImode, op, mode, n * 4); + } +} + +/* Return N-th part of value occupying multiple registers. */ + +rtx +gcn_operand_doublepart (machine_mode mode, rtx op, int n) +{ + return simplify_gen_subreg (DImode, op, mode, n * 8); +} + +/* Return true if OP can be split into subregs or high/low parts. + This is always true for scalars, but not normally true for vectors. + However, for vectors in hardregs we can use the low and high registers. */ + +bool +gcn_can_split_p (machine_mode, rtx op) +{ + if (vgpr_vector_mode_p (GET_MODE (op))) + { + if (GET_CODE (op) == SUBREG) + op = SUBREG_REG (op); + if (!REG_P (op)) + return true; + return REGNO (op) <= FIRST_PSEUDO_REGISTER; + } + return true; +} + +/* Implement TARGET_SPILL_CLASS. + + Return class of registers which could be used for pseudo of MODE + and of class RCLASS for spilling instead of memory. Return NO_REGS + if it is not possible or non-profitable. */ + +static reg_class_t +gcn_spill_class (reg_class_t c, machine_mode /*mode */ ) +{ + if (reg_classes_intersect_p (ALL_CONDITIONAL_REGS, c)) + return SGPR_REGS; + else + return NO_REGS; +} + +/* Implement TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS. + + Change allocno class for given pseudo from allocno and best class + calculated by IRA. */ + +static reg_class_t +gcn_ira_change_pseudo_allocno_class (int regno, reg_class_t cl, + reg_class_t best_cl) +{ + /* Avoid returning classes that contain both vgpr and sgpr registers. */ + if (cl != ALL_REGS && cl != SRCDST_REGS && cl != ALL_GPR_REGS) + return cl; + if (best_cl != ALL_REGS && best_cl != SRCDST_REGS + && best_cl != ALL_GPR_REGS) + return best_cl; + + machine_mode mode = PSEUDO_REGNO_MODE (regno); + if (vgpr_vector_mode_p (mode)) + return VGPR_REGS; + + return GENERAL_REGS; +} + +/* Create a new DImode pseudo reg and emit an instruction to initialize + it to VAL. */ + +static rtx +get_exec (int64_t val) +{ + rtx reg = gen_reg_rtx (DImode); + emit_insn (gen_rtx_SET (reg, gen_int_mode (val, DImode))); + return reg; +} + +/* Return value of scalar exec register. */ + +rtx +gcn_scalar_exec () +{ + return const1_rtx; +} + +/* Return pseudo holding scalar exec register. */ + +rtx +gcn_scalar_exec_reg () +{ + return get_exec (1); +} + +/* Return value of full exec register. */ + +rtx +gcn_full_exec () +{ + return constm1_rtx; +} + +/* Return pseudo holding full exec register. */ + +rtx +gcn_full_exec_reg () +{ + return get_exec (-1); +} + +/* }}} */ +/* {{{ Immediate constants. */ + +/* Initialize shared numeric constants. */ + +static void +init_ext_gcn_constants (void) +{ + real_from_integer (&dconst4, DFmode, 4, SIGNED); + + /* FIXME: this constant probably does not match what hardware really loads. + Reality check it eventually. */ + real_from_string (&dconst1over2pi, + "0.1591549430918953357663423455968866839"); + real_convert (&dconst1over2pi, SFmode, &dconst1over2pi); + + ext_gcn_constants_init = 1; +} + +/* Return non-zero if X is a constant that can appear as an inline operand. + This is 0, 0.5, -0.5, 1, -1, 2, -2, 4,-4, 1/(2*pi) + Or a vector of those. + The value returned should be the encoding of this constant. */ + +int +gcn_inline_fp_constant_p (rtx x, bool allow_vector) +{ + machine_mode mode = GET_MODE (x); + + if ((mode == V64HFmode || mode == V64SFmode || mode == V64DFmode) + && allow_vector) + { + int n; + if (GET_CODE (x) != CONST_VECTOR) + return 0; + n = gcn_inline_fp_constant_p (CONST_VECTOR_ELT (x, 0), false); + if (!n) + return 0; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return 0; + return 1; + } + + if (mode != HFmode && mode != SFmode && mode != DFmode) + return 0; + + const REAL_VALUE_TYPE *r; + + if (x == CONST0_RTX (mode)) + return 128; + if (x == CONST1_RTX (mode)) + return 242; + + r = CONST_DOUBLE_REAL_VALUE (x); + + if (real_identical (r, &dconstm1)) + return 243; + + if (real_identical (r, &dconsthalf)) + return 240; + if (real_identical (r, &dconstm1)) + return 243; + if (real_identical (r, &dconst2)) + return 244; + if (real_identical (r, &dconst4)) + return 246; + if (real_identical (r, &dconst1over2pi)) + return 248; + if (!ext_gcn_constants_init) + init_ext_gcn_constants (); + real_value_negate (r); + if (real_identical (r, &dconsthalf)) + return 241; + if (real_identical (r, &dconst2)) + return 245; + if (real_identical (r, &dconst4)) + return 247; + + /* FIXME: add 4, -4 and 1/(2*PI). */ + + return 0; +} + +/* Return non-zero if X is a constant that can appear as an immediate operand. + This is 0, 0.5, -0.5, 1, -1, 2, -2, 4,-4, 1/(2*pi) + Or a vector of those. + The value returned should be the encoding of this constant. */ + +bool +gcn_fp_constant_p (rtx x, bool allow_vector) +{ + machine_mode mode = GET_MODE (x); + + if ((mode == V64HFmode || mode == V64SFmode || mode == V64DFmode) + && allow_vector) + { + int n; + if (GET_CODE (x) != CONST_VECTOR) + return false; + n = gcn_fp_constant_p (CONST_VECTOR_ELT (x, 0), false); + if (!n) + return false; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return false; + return true; + } + if (mode != HFmode && mode != SFmode && mode != DFmode) + return false; + + if (gcn_inline_fp_constant_p (x, false)) + return true; + /* FIXME: It is not clear how 32bit immediates are interpreted here. */ + return (mode != DFmode); +} + +/* Return true if X is a constant representable as an inline immediate + constant in a 32-bit instruction encoding. */ + +bool +gcn_inline_constant_p (rtx x) +{ + if (GET_CODE (x) == CONST_INT) + return INTVAL (x) >= -16 && INTVAL (x) < 64; + if (GET_CODE (x) == CONST_DOUBLE) + return gcn_inline_fp_constant_p (x, false); + if (GET_CODE (x) == CONST_VECTOR) + { + int n; + if (!vgpr_vector_mode_p (GET_MODE (x))) + return false; + n = gcn_inline_constant_p (CONST_VECTOR_ELT (x, 0)); + if (!n) + return false; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return false; + return 1; + } + return false; +} + +/* Return true if X is a constant representable as an immediate constant + in a 32 or 64-bit instruction encoding. */ + +bool +gcn_constant_p (rtx x) +{ + switch (GET_CODE (x)) + { + case CONST_INT: + return true; + + case CONST_DOUBLE: + return gcn_fp_constant_p (x, false); + + case CONST_VECTOR: + { + int n; + if (!vgpr_vector_mode_p (GET_MODE (x))) + return false; + n = gcn_constant_p (CONST_VECTOR_ELT (x, 0)); + if (!n) + return false; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return false; + return true; + } + + case SYMBOL_REF: + case LABEL_REF: + return true; + + default: + ; + } + + return false; +} + +/* Return true if X is a constant representable as two inline immediate + constants in a 64-bit instruction that is split into two 32-bit + instructions. */ + +bool +gcn_inline_constant64_p (rtx x) +{ + if (GET_CODE (x) == CONST_VECTOR) + { + if (!vgpr_vector_mode_p (GET_MODE (x))) + return false; + if (!gcn_inline_constant64_p (CONST_VECTOR_ELT (x, 0))) + return false; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return false; + + return true; + } + + if (GET_CODE (x) != CONST_INT) + return false; + + rtx val_lo = gcn_operand_part (DImode, x, 0); + rtx val_hi = gcn_operand_part (DImode, x, 1); + return gcn_inline_constant_p (val_lo) && gcn_inline_constant_p (val_hi); +} + +/* Return true if X is a constant representable as an immediate constant + in a 32 or 64-bit instruction encoding where the hardware will + extend the immediate to 64-bits. */ + +bool +gcn_constant64_p (rtx x) +{ + if (!gcn_constant_p (x)) + return false; + + if (GET_CODE (x) != CONST_INT) + return true; + + /* Negative numbers are only allowed if they can be encoded within src0, + because the 32-bit immediates do not get sign-extended. + Unsigned numbers must not be encodable as 32-bit -1..-16, because the + assembler will use a src0 inline immediate and that will get + sign-extended. */ + HOST_WIDE_INT val = INTVAL (x); + return (((val & 0xffffffff) == val /* Positive 32-bit. */ + && (val & 0xfffffff0) != 0xfffffff0) /* Not -1..-16. */ + || gcn_inline_constant_p (x)); /* Src0. */ +} + +/* Implement TARGET_LEGITIMATE_CONSTANT_P. + + Returns true if X is a legitimate constant for a MODE immediate operand. */ + +bool +gcn_legitimate_constant_p (machine_mode, rtx x) +{ + return gcn_constant_p (x); +} + +/* Return true if X is a CONST_VECTOR of single constant. */ + +static bool +single_cst_vector_p (rtx x) +{ + if (GET_CODE (x) != CONST_VECTOR) + return false; + for (int i = 1; i < 64; i++) + if (CONST_VECTOR_ELT (x, i) != CONST_VECTOR_ELT (x, 0)) + return false; + return true; +} + +/* Create a CONST_VECTOR of duplicated value A. */ + +rtx +gcn_vec_constant (machine_mode mode, int a) +{ + /*if (!a) + return CONST0_RTX (mode); + if (a == -1) + return CONSTM1_RTX (mode); + if (a == 1) + return CONST1_RTX (mode); + if (a == 2) + return CONST2_RTX (mode);*/ + + int units = GET_MODE_NUNITS (mode); + rtx tem = gen_int_mode (a, GET_MODE_INNER (mode)); + rtvec v = rtvec_alloc (units); + + for (int i = 0; i < units; ++i) + RTVEC_ELT (v, i) = tem; + + return gen_rtx_CONST_VECTOR (mode, v); +} + +/* Create a CONST_VECTOR of duplicated value A. */ + +rtx +gcn_vec_constant (machine_mode mode, rtx a) +{ + int units = GET_MODE_NUNITS (mode); + rtvec v = rtvec_alloc (units); + + for (int i = 0; i < units; ++i) + RTVEC_ELT (v, i) = a; + + return gen_rtx_CONST_VECTOR (mode, v); +} + +/* Create an undefined vector value, used where an insn operand is + optional. */ + +rtx +gcn_gen_undef (machine_mode mode) +{ + return gen_rtx_UNSPEC (mode, gen_rtvec (1, const0_rtx), UNSPEC_VECTOR); +} + +/* }}} */ +/* {{{ Addresses, pointers and moves. */ + +/* Return true is REG is a valid place to store a pointer, + for instructions that require an SGPR. + FIXME rename. */ + +static bool +gcn_address_register_p (rtx reg, machine_mode mode, bool strict) +{ + if (GET_CODE (reg) == SUBREG) + reg = SUBREG_REG (reg); + + if (!REG_P (reg)) + return false; + + if (GET_MODE (reg) != mode) + return false; + + int regno = REGNO (reg); + + if (regno >= FIRST_PSEUDO_REGISTER) + { + if (!strict) + return true; + + if (!reg_renumber) + return false; + + regno = reg_renumber[regno]; + } + + return (SGPR_REGNO_P (regno) || regno == M0_REG + || regno == ARG_POINTER_REGNUM || regno == FRAME_POINTER_REGNUM); +} + +/* Return true is REG is a valid place to store a pointer, + for instructions that require a VGPR. */ + +static bool +gcn_vec_address_register_p (rtx reg, machine_mode mode, bool strict) +{ + if (GET_CODE (reg) == SUBREG) + reg = SUBREG_REG (reg); + + if (!REG_P (reg)) + return false; + + if (GET_MODE (reg) != mode) + return false; + + int regno = REGNO (reg); + + if (regno >= FIRST_PSEUDO_REGISTER) + { + if (!strict) + return true; + + if (!reg_renumber) + return false; + + regno = reg_renumber[regno]; + } + + return VGPR_REGNO_P (regno); +} + +/* Return true if X would be valid inside a MEM using the Flat address + space. */ + +bool +gcn_flat_address_p (rtx x, machine_mode mode) +{ + bool vec_mode = (GET_MODE_CLASS (mode) == MODE_VECTOR_INT + || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT); + + if (vec_mode && gcn_address_register_p (x, DImode, false)) + return true; + + if (!vec_mode && gcn_vec_address_register_p (x, DImode, false)) + return true; + + if (TARGET_GCN5_PLUS + && GET_CODE (x) == PLUS + && gcn_vec_address_register_p (XEXP (x, 0), DImode, false) + && CONST_INT_P (XEXP (x, 1))) + return true; + + return false; +} + +/* Return true if X would be valid inside a MEM using the Scalar Flat + address space. */ + +bool +gcn_scalar_flat_address_p (rtx x) +{ + if (gcn_address_register_p (x, DImode, false)) + return true; + + if (GET_CODE (x) == PLUS + && gcn_address_register_p (XEXP (x, 0), DImode, false) + && CONST_INT_P (XEXP (x, 1))) + return true; + + return false; +} + +/* Return true if MEM X would be valid for the Scalar Flat address space. */ + +bool +gcn_scalar_flat_mem_p (rtx x) +{ + if (!MEM_P (x)) + return false; + + if (GET_MODE_SIZE (GET_MODE (x)) < 4) + return false; + + return gcn_scalar_flat_address_p (XEXP (x, 0)); +} + +/* Return true if X would be valid inside a MEM using the LDS or GDS + address spaces. */ + +bool +gcn_ds_address_p (rtx x) +{ + if (gcn_vec_address_register_p (x, SImode, false)) + return true; + + if (GET_CODE (x) == PLUS + && gcn_vec_address_register_p (XEXP (x, 0), SImode, false) + && CONST_INT_P (XEXP (x, 1))) + return true; + + return false; +} + +/* Return true if ADDR would be valid inside a MEM using the Global + address space. */ + +bool +gcn_global_address_p (rtx addr) +{ + if (gcn_address_register_p (addr, DImode, false) + || gcn_vec_address_register_p (addr, DImode, false)) + return true; + + if (GET_CODE (addr) == PLUS) + { + rtx base = XEXP (addr, 0); + rtx offset = XEXP (addr, 1); + bool immediate_p = (CONST_INT_P (offset) + && INTVAL (offset) >= -(1 << 12) + && INTVAL (offset) < (1 << 12)); + + if ((gcn_address_register_p (base, DImode, false) + || gcn_vec_address_register_p (base, DImode, false)) + && immediate_p) + /* SGPR + CONST or VGPR + CONST */ + return true; + + if (gcn_address_register_p (base, DImode, false) + && gcn_vgpr_register_operand (offset, SImode)) + /* SPGR + VGPR */ + return true; + + if (GET_CODE (base) == PLUS + && gcn_address_register_p (XEXP (base, 0), DImode, false) + && gcn_vgpr_register_operand (XEXP (base, 1), SImode) + && immediate_p) + /* (SGPR + VGPR) + CONST */ + return true; + } + + return false; +} + +/* Implement TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P. + + Recognizes RTL expressions that are valid memory addresses for an + instruction. The MODE argument is the machine mode for the MEM + expression that wants to use this address. + + It only recognizes address in canonical form. LEGITIMIZE_ADDRESS should + convert common non-canonical forms to canonical form so that they will + be recognized. */ + +static bool +gcn_addr_space_legitimate_address_p (machine_mode mode, rtx x, bool strict, + addr_space_t as) +{ + /* All vector instructions need to work on addresses in registers. */ + if (!TARGET_GCN5_PLUS && (vgpr_vector_mode_p (mode) && !REG_P (x))) + return false; + + if (AS_SCALAR_FLAT_P (as)) + { + if (mode == QImode || mode == HImode) + return 0; + + switch (GET_CODE (x)) + { + case REG: + return gcn_address_register_p (x, DImode, strict); + /* Addresses are in the form BASE+OFFSET + OFFSET is either 20bit unsigned immediate, SGPR or M0. + Writes and atomics do not accept SGPR. */ + case PLUS: + { + rtx x0 = XEXP (x, 0); + rtx x1 = XEXP (x, 1); + if (!gcn_address_register_p (x0, DImode, strict)) + return false; + /* FIXME: This is disabled because of the mode mismatch between + SImode (for the address or m0 register) and the DImode PLUS. + We'll need a zero_extend or similar. + + if (gcn_m0_register_p (x1, SImode, strict) + || gcn_address_register_p (x1, SImode, strict)) + return true; + else*/ + if (GET_CODE (x1) == CONST_INT) + { + if (INTVAL (x1) >= 0 && INTVAL (x1) < (1 << 20) + /* The low bits of the offset are ignored, even when + they're meant to realign the pointer. */ + && !(INTVAL (x1) & 0x3)) + return true; + } + return false; + } + + default: + break; + } + } + else if (AS_SCRATCH_P (as)) + return gcn_address_register_p (x, SImode, strict); + else if (AS_FLAT_P (as) || AS_FLAT_SCRATCH_P (as)) + { + if (TARGET_GCN3 || GET_CODE (x) == REG) + return ((GET_MODE_CLASS (mode) == MODE_VECTOR_INT + || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) + ? gcn_address_register_p (x, DImode, strict) + : gcn_vec_address_register_p (x, DImode, strict)); + else + { + gcc_assert (TARGET_GCN5_PLUS); + + if (GET_CODE (x) == PLUS) + { + rtx x1 = XEXP (x, 1); + + if (VECTOR_MODE_P (mode) + ? !gcn_address_register_p (x, DImode, strict) + : !gcn_vec_address_register_p (x, DImode, strict)) + return false; + + if (GET_CODE (x1) == CONST_INT) + { + if (INTVAL (x1) >= 0 && INTVAL (x1) < (1 << 12) + /* The low bits of the offset are ignored, even when + they're meant to realign the pointer. */ + && !(INTVAL (x1) & 0x3)) + return true; + } + } + return false; + } + } + else if (AS_GLOBAL_P (as)) + { + gcc_assert (TARGET_GCN5_PLUS); + + if (GET_CODE (x) == REG) + return (gcn_address_register_p (x, DImode, strict) + || (!VECTOR_MODE_P (mode) + && gcn_vec_address_register_p (x, DImode, strict))); + else if (GET_CODE (x) == PLUS) + { + rtx base = XEXP (x, 0); + rtx offset = XEXP (x, 1); + + bool immediate_p = (GET_CODE (offset) == CONST_INT + /* Signed 13-bit immediate. */ + && INTVAL (offset) >= -(1 << 12) + && INTVAL (offset) < (1 << 12) + /* The low bits of the offset are ignored, even + when they're meant to realign the pointer. */ + && !(INTVAL (offset) & 0x3)); + + if (!VECTOR_MODE_P (mode)) + { + if ((gcn_address_register_p (base, DImode, strict) + || gcn_vec_address_register_p (base, DImode, strict)) + && immediate_p) + /* SGPR + CONST or VGPR + CONST */ + return true; + + if (gcn_address_register_p (base, DImode, strict) + && gcn_vgpr_register_operand (offset, SImode)) + /* SGPR + VGPR */ + return true; + + if (GET_CODE (base) == PLUS + && gcn_address_register_p (XEXP (base, 0), DImode, strict) + && gcn_vgpr_register_operand (XEXP (base, 1), SImode) + && immediate_p) + /* (SGPR + VGPR) + CONST */ + return true; + } + else + { + if (gcn_address_register_p (base, DImode, strict) + && immediate_p) + /* SGPR + CONST */ + return true; + } + } + else + return false; + } + else if (AS_ANY_DS_P (as)) + switch (GET_CODE (x)) + { + case REG: + return (VECTOR_MODE_P (mode) + ? gcn_address_register_p (x, SImode, strict) + : gcn_vec_address_register_p (x, SImode, strict)); + /* Addresses are in the form BASE+OFFSET + OFFSET is either 20bit unsigned immediate, SGPR or M0. + Writes and atomics do not accept SGPR. */ + case PLUS: + { + rtx x0 = XEXP (x, 0); + rtx x1 = XEXP (x, 1); + if (!gcn_vec_address_register_p (x0, DImode, strict)) + return false; + if (GET_CODE (x1) == REG) + { + if (GET_CODE (x1) != REG + || (REGNO (x1) <= FIRST_PSEUDO_REGISTER + && !gcn_ssrc_register_operand (x1, DImode))) + return false; + } + else if (GET_CODE (x1) == CONST_VECTOR + && GET_CODE (CONST_VECTOR_ELT (x1, 0)) == CONST_INT + && single_cst_vector_p (x1)) + { + x1 = CONST_VECTOR_ELT (x1, 0); + if (INTVAL (x1) >= 0 && INTVAL (x1) < (1 << 20)) + return true; + } + return false; + } + + default: + break; + } + else + gcc_unreachable (); + return false; +} + +/* Implement TARGET_ADDR_SPACE_POINTER_MODE. + + Return the appropriate mode for a named address pointer. */ + +static scalar_int_mode +gcn_addr_space_pointer_mode (addr_space_t addrspace) +{ + switch (addrspace) + { + case ADDR_SPACE_SCRATCH: + case ADDR_SPACE_LDS: + case ADDR_SPACE_GDS: + return SImode; + case ADDR_SPACE_DEFAULT: + case ADDR_SPACE_FLAT: + case ADDR_SPACE_FLAT_SCRATCH: + case ADDR_SPACE_SCALAR_FLAT: + return DImode; + default: + gcc_unreachable (); + } +} + +/* Implement TARGET_ADDR_SPACE_ADDRESS_MODE. + + Return the appropriate mode for a named address space address. */ + +static scalar_int_mode +gcn_addr_space_address_mode (addr_space_t addrspace) +{ + return gcn_addr_space_pointer_mode (addrspace); +} + +/* Implement TARGET_ADDR_SPACE_SUBSET_P. + + Determine if one named address space is a subset of another. */ + +static bool +gcn_addr_space_subset_p (addr_space_t subset, addr_space_t superset) +{ + if (subset == superset) + return true; + /* FIXME is this true? */ + if (AS_FLAT_P (superset) || AS_SCALAR_FLAT_P (superset)) + return true; + return false; +} + +/* Convert from one address space to another. */ + +static rtx +gcn_addr_space_convert (rtx op, tree from_type, tree to_type) +{ + gcc_assert (POINTER_TYPE_P (from_type)); + gcc_assert (POINTER_TYPE_P (to_type)); + + addr_space_t as_from = TYPE_ADDR_SPACE (TREE_TYPE (from_type)); + addr_space_t as_to = TYPE_ADDR_SPACE (TREE_TYPE (to_type)); + + if (AS_LDS_P (as_from) && AS_FLAT_P (as_to)) + { + rtx queue = gen_rtx_REG (DImode, + cfun->machine->args.reg[QUEUE_PTR_ARG]); + rtx group_seg_aperture_hi = gen_rtx_MEM (SImode, + gen_rtx_PLUS (DImode, queue, + gen_int_mode (64, SImode))); + rtx tmp = gen_reg_rtx (DImode); + + emit_move_insn (gen_lowpart (SImode, tmp), op); + emit_move_insn (gen_highpart_mode (SImode, DImode, tmp), + group_seg_aperture_hi); + + return tmp; + } + else if (as_from == as_to) + return op; + else + gcc_unreachable (); +} + + +/* Implement REGNO_MODE_CODE_OK_FOR_BASE_P via gcn.h + + Retun true if REGNO is OK for memory adressing. */ + +bool +gcn_regno_mode_code_ok_for_base_p (int regno, + machine_mode, addr_space_t as, int, int) +{ + if (regno >= FIRST_PSEUDO_REGISTER) + { + if (reg_renumber) + regno = reg_renumber[regno]; + else + return true; + } + if (AS_FLAT_P (as)) + return (VGPR_REGNO_P (regno) + || regno == ARG_POINTER_REGNUM || regno == FRAME_POINTER_REGNUM); + else if (AS_SCALAR_FLAT_P (as)) + return (SGPR_REGNO_P (regno) + || regno == ARG_POINTER_REGNUM || regno == FRAME_POINTER_REGNUM); + else if (AS_GLOBAL_P (as)) + { + return (SGPR_REGNO_P (regno) + || VGPR_REGNO_P (regno) + || regno == ARG_POINTER_REGNUM + || regno == FRAME_POINTER_REGNUM); + } + else + /* For now. */ + return false; +} + +/* Implement MODE_CODE_BASE_REG_CLASS via gcn.h. + + Return a suitable register class for memory addressing. */ + +reg_class +gcn_mode_code_base_reg_class (machine_mode mode, addr_space_t as, int oc, + int ic) +{ + switch (as) + { + case ADDR_SPACE_DEFAULT: + return gcn_mode_code_base_reg_class (mode, DEFAULT_ADDR_SPACE, oc, ic); + case ADDR_SPACE_SCALAR_FLAT: + case ADDR_SPACE_SCRATCH: + return SGPR_REGS; + break; + case ADDR_SPACE_FLAT: + case ADDR_SPACE_FLAT_SCRATCH: + case ADDR_SPACE_LDS: + case ADDR_SPACE_GDS: + return ((GET_MODE_CLASS (mode) == MODE_VECTOR_INT + || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) + ? SGPR_REGS : VGPR_REGS); + case ADDR_SPACE_GLOBAL: + return ((GET_MODE_CLASS (mode) == MODE_VECTOR_INT + || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) + ? SGPR_REGS : ALL_GPR_REGS); + } + gcc_unreachable (); +} + +/* Implement REGNO_OK_FOR_INDEX_P via gcn.h. + + Return true if REGNO is OK for index of memory addressing. */ + +bool +regno_ok_for_index_p (int regno) +{ + if (regno >= FIRST_PSEUDO_REGISTER) + { + if (reg_renumber) + regno = reg_renumber[regno]; + else + return true; + } + return regno == M0_REG || VGPR_REGNO_P (regno); +} + +/* Generate move which uses the exec flags. If EXEC is NULL, then it is + assumed that all lanes normally relevant to the mode of the move are + affected. If PREV is NULL, then a sensible default is supplied for + the inactive lanes. */ + +static rtx +gen_mov_with_exec (rtx op0, rtx op1, rtx exec = NULL, rtx prev = NULL) +{ + machine_mode mode = GET_MODE (op0); + + if (vgpr_vector_mode_p (mode)) + { + if (exec && exec != CONSTM1_RTX (DImode)) + { + if (!prev) + prev = op0; + } + else + { + if (!prev) + prev = gcn_gen_undef (mode); + exec = gcn_full_exec_reg (); + } + + rtx set = gen_rtx_SET (op0, gen_rtx_VEC_MERGE (mode, op1, prev, exec)); + + return gen_rtx_PARALLEL (VOIDmode, + gen_rtvec (2, set, + gen_rtx_CLOBBER (VOIDmode, + gen_rtx_SCRATCH (V64DImode)))); + } + + return (gen_rtx_PARALLEL + (VOIDmode, + gen_rtvec (2, gen_rtx_SET (op0, op1), + gen_rtx_USE (VOIDmode, + exec ? exec : gcn_scalar_exec ())))); +} + +/* Generate masked move. */ + +static rtx +gen_duplicate_load (rtx op0, rtx op1, rtx op2 = NULL, rtx exec = NULL) +{ + if (exec) + return (gen_rtx_SET (op0, + gen_rtx_VEC_MERGE (GET_MODE (op0), + gen_rtx_VEC_DUPLICATE (GET_MODE + (op0), op1), + op2, exec))); + else + return (gen_rtx_SET (op0, gen_rtx_VEC_DUPLICATE (GET_MODE (op0), op1))); +} + +/* Expand vector init of OP0 by VEC. + Implements vec_init instruction pattern. */ + +void +gcn_expand_vector_init (rtx op0, rtx vec) +{ + int64_t initialized_mask = 0; + int64_t curr_mask = 1; + machine_mode mode = GET_MODE (op0); + + rtx val = XVECEXP (vec, 0, 0); + + for (int i = 1; i < 64; i++) + if (rtx_equal_p (val, XVECEXP (vec, 0, i))) + curr_mask |= (int64_t) 1 << i; + + if (gcn_constant_p (val)) + emit_move_insn (op0, gcn_vec_constant (mode, val)); + else + { + val = force_reg (GET_MODE_INNER (mode), val); + emit_insn (gen_duplicate_load (op0, val)); + } + initialized_mask |= curr_mask; + for (int i = 1; i < 64; i++) + if (!(initialized_mask & ((int64_t) 1 << i))) + { + curr_mask = (int64_t) 1 << i; + rtx val = XVECEXP (vec, 0, i); + + for (int j = i + 1; j < 64; j++) + if (rtx_equal_p (val, XVECEXP (vec, 0, j))) + curr_mask |= (int64_t) 1 << j; + if (gcn_constant_p (val)) + emit_insn (gen_mov_with_exec (op0, gcn_vec_constant (mode, val), + get_exec (curr_mask))); + else + { + val = force_reg (GET_MODE_INNER (mode), val); + emit_insn (gen_duplicate_load (op0, val, op0, + get_exec (curr_mask))); + } + initialized_mask |= curr_mask; + } +} + +/* Load vector constant where n-th lane contains BASE+n*VAL. */ + +static rtx +strided_constant (machine_mode mode, int base, int val) +{ + rtx x = gen_reg_rtx (mode); + emit_move_insn (x, gcn_vec_constant (mode, base)); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 32), + x, get_exec (0xffffffff00000000))); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 16), + x, get_exec (0xffff0000ffff0000))); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 8), + x, get_exec (0xff00ff00ff00ff00))); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 4), + x, get_exec (0xf0f0f0f0f0f0f0f0))); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 2), + x, get_exec (0xcccccccccccccccc))); + emit_insn (gen_addv64si3_exec (x, x, gcn_vec_constant (mode, val * 1), + x, get_exec (0xaaaaaaaaaaaaaaaa))); + return x; +} + +/* Implement TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS. */ + +static rtx +gcn_addr_space_legitimize_address (rtx x, rtx old, machine_mode mode, + addr_space_t as) +{ + switch (as) + { + case ADDR_SPACE_DEFAULT: + return gcn_addr_space_legitimize_address (x, old, mode, + DEFAULT_ADDR_SPACE); + case ADDR_SPACE_SCALAR_FLAT: + case ADDR_SPACE_SCRATCH: + /* Instructions working on vectors need the address to be in + a register. */ + if (vgpr_vector_mode_p (mode)) + return force_reg (GET_MODE (x), x); + + return x; + case ADDR_SPACE_FLAT: + case ADDR_SPACE_FLAT_SCRATCH: + case ADDR_SPACE_GLOBAL: + return TARGET_GCN3 ? force_reg (DImode, x) : x; + case ADDR_SPACE_LDS: + case ADDR_SPACE_GDS: + /* FIXME: LDS support offsets, handle them!. */ + if (vgpr_vector_mode_p (mode) && GET_MODE (x) != V64SImode) + { + rtx addrs = gen_reg_rtx (V64SImode); + rtx base = force_reg (SImode, x); + rtx offsets = strided_constant (V64SImode, 0, + GET_MODE_UNIT_SIZE (mode)); + + emit_insn (gen_vec_duplicatev64si (addrs, base)); + emit_insn (gen_addv64si3 (addrs, offsets, addrs)); + return addrs; + } + return x; + } + gcc_unreachable (); +} + +/* Convert a (mem: (reg:DI)) to (mem: (reg:V64DI)) with the + proper vector of stepped addresses. + + MEM will be a DImode address of a vector in an SGPR. + TMP will be a V64DImode VGPR pair or (scratch:V64DI). */ + +rtx +gcn_expand_scalar_to_vector_address (machine_mode mode, rtx exec, rtx mem, + rtx tmp) +{ + gcc_assert (MEM_P (mem)); + rtx mem_base = XEXP (mem, 0); + rtx mem_index = NULL_RTX; + + if (!TARGET_GCN5_PLUS) + { + /* gcn_addr_space_legitimize_address should have put the address in a + register. If not, it is too late to do anything about it. */ + gcc_assert (REG_P (mem_base)); + } + + if (GET_CODE (mem_base) == PLUS) + { + mem_index = XEXP (mem_base, 1); + mem_base = XEXP (mem_base, 0); + } + + /* RF and RM base registers for vector modes should be always an SGPR. */ + gcc_assert (SGPR_REGNO_P (REGNO (mem_base)) + || REGNO (mem_base) >= FIRST_PSEUDO_REGISTER); + + machine_mode inner = GET_MODE_INNER (mode); + int shift = exact_log2 (GET_MODE_SIZE (inner)); + rtx ramp = gen_rtx_REG (V64SImode, VGPR_REGNO (1)); + rtx undef_v64si = gcn_gen_undef (V64SImode); + rtx new_base = NULL_RTX; + addr_space_t as = MEM_ADDR_SPACE (mem); + + rtx tmplo = (REG_P (tmp) + ? gcn_operand_part (V64DImode, tmp, 0) + : gen_reg_rtx (V64SImode)); + + /* tmplo[:] = ramp[:] << shift */ + if (exec) + emit_insn (gen_ashlv64si3_exec (tmplo, ramp, + gen_int_mode (shift, SImode), + undef_v64si, exec)); + else + emit_insn (gen_ashlv64si3 (tmplo, ramp, gen_int_mode (shift, SImode))); + + if (AS_FLAT_P (as)) + { + if (REG_P (tmp)) + { + rtx vcc = gen_rtx_REG (DImode, CC_SAVE_REG); + rtx mem_base_lo = gcn_operand_part (DImode, mem_base, 0); + rtx mem_base_hi = gcn_operand_part (DImode, mem_base, 1); + rtx tmphi = gcn_operand_part (V64DImode, tmp, 1); + + /* tmphi[:] = mem_base_hi */ + if (exec) + emit_insn (gen_vec_duplicatev64si_exec (tmphi, mem_base_hi, + undef_v64si, exec)); + else + emit_insn (gen_vec_duplicatev64si (tmphi, mem_base_hi)); + + /* tmp[:] += zext (mem_base) */ + if (exec) + { + rtx undef_di = gcn_gen_undef (DImode); + emit_insn (gen_addv64si3_vcc_dup_exec (tmplo, mem_base_lo, tmplo, + vcc, undef_v64si, exec)); + emit_insn (gen_addcv64si3_exec (tmphi, tmphi, const0_rtx, + vcc, vcc, undef_v64si, exec)); + } + else + emit_insn (gen_addv64di3_zext_dup (tmp, mem_base_lo, tmp)); + } + else + { + tmp = gen_reg_rtx (V64DImode); + if (exec) + emit_insn (gen_addv64di3_zext_dup2_exec (tmp, tmplo, mem_base, + gcn_gen_undef (V64DImode), + exec)); + else + emit_insn (gen_addv64di3_zext_dup2 (tmp, tmplo, mem_base)); + } + + new_base = tmp; + } + else if (AS_ANY_DS_P (as)) + { + if (!exec) + emit_insn (gen_addv64si3_dup (tmplo, tmplo, mem_base)); + else + emit_insn (gen_addv64si3_dup_exec (tmplo, tmplo, mem_base, + gcn_gen_undef (V64SImode), exec)); + new_base = tmplo; + } + else + { + mem_base = gen_rtx_VEC_DUPLICATE (V64DImode, mem_base); + new_base = gen_rtx_PLUS (V64DImode, mem_base, + gen_rtx_SIGN_EXTEND (V64DImode, tmplo)); + } + + return gen_rtx_PLUS (GET_MODE (new_base), new_base, + gen_rtx_VEC_DUPLICATE (GET_MODE (new_base), + (mem_index ? mem_index + : const0_rtx))); +} + +/* Convert a BASE address, a vector of OFFSETS, and a SCALE, to addresses + suitable for the given address space. This is indented for use in + gather/scatter patterns. + + The offsets may be signed or unsigned, according to UNSIGNED_P. + If EXEC is set then _exec patterns will be used, otherwise plain. + + Return values. + ADDR_SPACE_FLAT - return V64DImode vector of absolute addresses. + ADDR_SPACE_GLOBAL - return V64SImode vector of offsets. */ + +rtx +gcn_expand_scaled_offsets (addr_space_t as, rtx base, rtx offsets, rtx scale, + bool unsigned_p, rtx exec) +{ + /* Convert the offsets to V64SImode. + TODO: more conversions will be needed when more types are vectorized. */ + if (GET_MODE (offsets) == V64DImode) + { + rtx tmp = gen_reg_rtx (V64SImode); + emit_insn (gen_vec_truncatev64div64si (tmp, offsets)); + offsets = tmp; + } + + rtx tmpsi = gen_reg_rtx (V64SImode); + rtx tmpdi = gen_reg_rtx (V64DImode); + rtx undefsi = exec ? gcn_gen_undef (V64SImode) : NULL; + rtx undefdi = exec ? gcn_gen_undef (V64DImode) : NULL; + + if (CONST_INT_P (scale) + && INTVAL (scale) > 0 + && exact_log2 (INTVAL (scale)) >= 0) + emit_insn (gen_ashlv64si3 (tmpsi, offsets, + GEN_INT (exact_log2 (INTVAL (scale))))); + else + (exec + ? emit_insn (gen_mulv64si3_dup_exec (tmpsi, offsets, scale, undefsi, + exec)) + : emit_insn (gen_mulv64si3_dup (tmpsi, offsets, scale))); + + /* "Global" instructions do not support negative register offsets. */ + if (as == ADDR_SPACE_FLAT || !unsigned_p) + { + if (unsigned_p) + (exec + ? emit_insn (gen_addv64di3_zext_dup2_exec (tmpdi, tmpsi, base, + undefdi, exec)) + : emit_insn (gen_addv64di3_zext_dup2 (tmpdi, tmpsi, base))); + else + (exec + ? emit_insn (gen_addv64di3_sext_dup2_exec (tmpdi, tmpsi, base, + undefdi, exec)) + : emit_insn (gen_addv64di3_sext_dup2 (tmpdi, tmpsi, base))); + return tmpdi; + } + else if (as == ADDR_SPACE_GLOBAL) + return tmpsi; + + gcc_unreachable (); +} + +/* Return true if move from OP0 to OP1 is known to be executed in vector + unit. */ + +bool +gcn_vgpr_move_p (rtx op0, rtx op1) +{ + if (MEM_P (op0) && AS_SCALAR_FLAT_P (MEM_ADDR_SPACE (op0))) + return true; + if (MEM_P (op1) && AS_SCALAR_FLAT_P (MEM_ADDR_SPACE (op1))) + return true; + return ((REG_P (op0) && VGPR_REGNO_P (REGNO (op0))) + || (REG_P (op1) && VGPR_REGNO_P (REGNO (op1))) + || vgpr_vector_mode_p (GET_MODE (op0))); +} + +/* Return true if move from OP0 to OP1 is known to be executed in scalar + unit. Used in the machine description. */ + +bool +gcn_sgpr_move_p (rtx op0, rtx op1) +{ + if (MEM_P (op0) && AS_SCALAR_FLAT_P (MEM_ADDR_SPACE (op0))) + return true; + if (MEM_P (op1) && AS_SCALAR_FLAT_P (MEM_ADDR_SPACE (op1))) + return true; + if (!REG_P (op0) || REGNO (op0) >= FIRST_PSEUDO_REGISTER + || VGPR_REGNO_P (REGNO (op0))) + return false; + if (REG_P (op1) + && REGNO (op1) < FIRST_PSEUDO_REGISTER + && !VGPR_REGNO_P (REGNO (op1))) + return true; + return immediate_operand (op1, VOIDmode) || memory_operand (op1, VOIDmode); +} + +/* Implement TARGET_SECONDARY_RELOAD. + + The address space determines which registers can be used for loads and + stores. */ + +static reg_class_t +gcn_secondary_reload (bool in_p, rtx x, reg_class_t rclass, + machine_mode reload_mode, secondary_reload_info *sri) +{ + reg_class_t result = NO_REGS; + bool spilled_pseudo = + (REG_P (x) || GET_CODE (x) == SUBREG) && true_regnum (x) == -1; + + if (dump_file && (dump_flags & TDF_DETAILS)) + { + fprintf (dump_file, "gcn_secondary_reload: "); + dump_value_slim (dump_file, x, 1); + fprintf (dump_file, " %s %s:%s", (in_p ? "->" : "<-"), + reg_class_names[rclass], GET_MODE_NAME (reload_mode)); + if (REG_P (x) || GET_CODE (x) == SUBREG) + fprintf (dump_file, " (true regnum: %d \"%s\")", true_regnum (x), + (true_regnum (x) >= 0 + && true_regnum (x) < FIRST_PSEUDO_REGISTER + ? reg_names[true_regnum (x)] + : (spilled_pseudo ? "stack spill" : "??"))); + fprintf (dump_file, "\n"); + } + + /* Some callers don't use or initialize icode. */ + sri->icode = CODE_FOR_nothing; + + if (MEM_P (x) || spilled_pseudo) + { + addr_space_t as = DEFAULT_ADDR_SPACE; + + /* If we have a spilled pseudo, we can't find the address space + directly, but we know it's in ADDR_SPACE_FLAT space for GCN3 or + ADDR_SPACE_GLOBAL for GCN5. */ + if (MEM_P (x)) + as = MEM_ADDR_SPACE (x); + + if (as == ADDR_SPACE_DEFAULT) + as = DEFAULT_ADDR_SPACE; + + switch (as) + { + case ADDR_SPACE_SCALAR_FLAT: + result = + ((!MEM_P (x) || rclass == SGPR_REGS) ? NO_REGS : SGPR_REGS); + break; + case ADDR_SPACE_FLAT: + case ADDR_SPACE_FLAT_SCRATCH: + case ADDR_SPACE_GLOBAL: + if (GET_MODE_CLASS (reload_mode) == MODE_VECTOR_INT + || GET_MODE_CLASS (reload_mode) == MODE_VECTOR_FLOAT) + { + if (in_p) + switch (reload_mode) + { + case E_V64SImode: + sri->icode = CODE_FOR_reload_inv64si; + break; + case E_V64SFmode: + sri->icode = CODE_FOR_reload_inv64sf; + break; + case E_V64HImode: + sri->icode = CODE_FOR_reload_inv64hi; + break; + case E_V64HFmode: + sri->icode = CODE_FOR_reload_inv64hf; + break; + case E_V64QImode: + sri->icode = CODE_FOR_reload_inv64qi; + break; + case E_V64DImode: + sri->icode = CODE_FOR_reload_inv64di; + break; + case E_V64DFmode: + sri->icode = CODE_FOR_reload_inv64df; + break; + default: + gcc_unreachable (); + } + else + switch (reload_mode) + { + case E_V64SImode: + sri->icode = CODE_FOR_reload_outv64si; + break; + case E_V64SFmode: + sri->icode = CODE_FOR_reload_outv64sf; + break; + case E_V64HImode: + sri->icode = CODE_FOR_reload_outv64hi; + break; + case E_V64HFmode: + sri->icode = CODE_FOR_reload_outv64hf; + break; + case E_V64QImode: + sri->icode = CODE_FOR_reload_outv64qi; + break; + case E_V64DImode: + sri->icode = CODE_FOR_reload_outv64di; + break; + case E_V64DFmode: + sri->icode = CODE_FOR_reload_outv64df; + break; + default: + gcc_unreachable (); + } + break; + } + /* Fallthrough. */ + case ADDR_SPACE_LDS: + case ADDR_SPACE_GDS: + case ADDR_SPACE_SCRATCH: + result = (rclass == VGPR_REGS ? NO_REGS : VGPR_REGS); + break; + } + } + + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, " <= %s (icode: %s)\n", reg_class_names[result], + get_insn_name (sri->icode)); + + return result; +} + +/* Update register usage after having seen the compiler flags and kernel + attributes. We typically want to fix registers that contain values + set by the HSA runtime. */ + +static void +gcn_conditional_register_usage (void) +{ + int i; + + /* FIXME: Do we need to reset fixed_regs? */ + +/* Limit ourselves to 1/16 the register file for maximimum sized workgroups. + There are enough SGPRs not to limit those. + TODO: Adjust this more dynamically. */ + for (i = FIRST_VGPR_REG + 64; i <= LAST_VGPR_REG; i++) + fixed_regs[i] = 1, call_used_regs[i] = 1; + + if (!cfun || !cfun->machine || cfun->machine->normal_function) + { + /* Normal functions can't know what kernel argument registers are + live, so just fix the bottom 16 SGPRs, and bottom 3 VGPRs. */ + for (i = 0; i < 16; i++) + fixed_regs[FIRST_SGPR_REG + i] = 1; + for (i = 0; i < 3; i++) + fixed_regs[FIRST_VGPR_REG + i] = 1; + return; + } + + /* Fix the runtime argument register containing values that may be + needed later. DISPATCH_PTR_ARG and FLAT_SCRATCH_* should not be + needed after the prologue so there's no need to fix them. */ + if (cfun->machine->args.reg[PRIVATE_SEGMENT_WAVE_OFFSET_ARG] >= 0) + fixed_regs[cfun->machine->args.reg[PRIVATE_SEGMENT_WAVE_OFFSET_ARG]] = 1; + if (cfun->machine->args.reg[PRIVATE_SEGMENT_BUFFER_ARG] >= 0) + { + fixed_regs[cfun->machine->args.reg[PRIVATE_SEGMENT_BUFFER_ARG]] = 1; + fixed_regs[cfun->machine->args.reg[PRIVATE_SEGMENT_BUFFER_ARG] + 1] = 1; + fixed_regs[cfun->machine->args.reg[PRIVATE_SEGMENT_BUFFER_ARG] + 2] = 1; + fixed_regs[cfun->machine->args.reg[PRIVATE_SEGMENT_BUFFER_ARG] + 3] = 1; + } + if (cfun->machine->args.reg[KERNARG_SEGMENT_PTR_ARG] >= 0) + { + fixed_regs[cfun->machine->args.reg[KERNARG_SEGMENT_PTR_ARG]] = 1; + fixed_regs[cfun->machine->args.reg[KERNARG_SEGMENT_PTR_ARG] + 1] = 1; + } + if (cfun->machine->args.reg[DISPATCH_PTR_ARG] >= 0) + { + fixed_regs[cfun->machine->args.reg[DISPATCH_PTR_ARG]] = 1; + fixed_regs[cfun->machine->args.reg[DISPATCH_PTR_ARG] + 1] = 1; + } + if (cfun->machine->args.reg[WORKGROUP_ID_X_ARG] >= 0) + fixed_regs[cfun->machine->args.reg[WORKGROUP_ID_X_ARG]] = 1; + if (cfun->machine->args.reg[WORK_ITEM_ID_X_ARG] >= 0) + fixed_regs[cfun->machine->args.reg[WORK_ITEM_ID_X_ARG]] = 1; + if (cfun->machine->args.reg[WORK_ITEM_ID_Y_ARG] >= 0) + fixed_regs[cfun->machine->args.reg[WORK_ITEM_ID_Y_ARG]] = 1; + if (cfun->machine->args.reg[WORK_ITEM_ID_Z_ARG] >= 0) + fixed_regs[cfun->machine->args.reg[WORK_ITEM_ID_Z_ARG]] = 1; + + if (TARGET_GCN5_PLUS) + /* v0 is always zero, for global nul-offsets. */ + fixed_regs[VGPR_REGNO (0)] = 1; +} + +/* Determine if a load or store is valid, according to the register classes + and address space. Used primarily by the machine description to decide + when to split a move into two steps. */ + +bool +gcn_valid_move_p (machine_mode mode, rtx dest, rtx src) +{ + if (!MEM_P (dest) && !MEM_P (src)) + return true; + + if (MEM_P (dest) + && AS_FLAT_P (MEM_ADDR_SPACE (dest)) + && (gcn_flat_address_p (XEXP (dest, 0), mode) + || GET_CODE (XEXP (dest, 0)) == SYMBOL_REF + || GET_CODE (XEXP (dest, 0)) == LABEL_REF) + && gcn_vgpr_register_operand (src, mode)) + return true; + else if (MEM_P (src) + && AS_FLAT_P (MEM_ADDR_SPACE (src)) + && (gcn_flat_address_p (XEXP (src, 0), mode) + || GET_CODE (XEXP (src, 0)) == SYMBOL_REF + || GET_CODE (XEXP (src, 0)) == LABEL_REF) + && gcn_vgpr_register_operand (dest, mode)) + return true; + + if (MEM_P (dest) + && AS_GLOBAL_P (MEM_ADDR_SPACE (dest)) + && (gcn_global_address_p (XEXP (dest, 0)) + || GET_CODE (XEXP (dest, 0)) == SYMBOL_REF + || GET_CODE (XEXP (dest, 0)) == LABEL_REF) + && gcn_vgpr_register_operand (src, mode)) + return true; + else if (MEM_P (src) + && AS_GLOBAL_P (MEM_ADDR_SPACE (src)) + && (gcn_global_address_p (XEXP (src, 0)) + || GET_CODE (XEXP (src, 0)) == SYMBOL_REF + || GET_CODE (XEXP (src, 0)) == LABEL_REF) + && gcn_vgpr_register_operand (dest, mode)) + return true; + + if (MEM_P (dest) + && MEM_ADDR_SPACE (dest) == ADDR_SPACE_SCALAR_FLAT + && (gcn_scalar_flat_address_p (XEXP (dest, 0)) + || GET_CODE (XEXP (dest, 0)) == SYMBOL_REF + || GET_CODE (XEXP (dest, 0)) == LABEL_REF) + && gcn_ssrc_register_operand (src, mode)) + return true; + else if (MEM_P (src) + && MEM_ADDR_SPACE (src) == ADDR_SPACE_SCALAR_FLAT + && (gcn_scalar_flat_address_p (XEXP (src, 0)) + || GET_CODE (XEXP (src, 0)) == SYMBOL_REF + || GET_CODE (XEXP (src, 0)) == LABEL_REF) + && gcn_sdst_register_operand (dest, mode)) + return true; + + if (MEM_P (dest) + && AS_ANY_DS_P (MEM_ADDR_SPACE (dest)) + && gcn_ds_address_p (XEXP (dest, 0)) + && gcn_vgpr_register_operand (src, mode)) + return true; + else if (MEM_P (src) + && AS_ANY_DS_P (MEM_ADDR_SPACE (src)) + && gcn_ds_address_p (XEXP (src, 0)) + && gcn_vgpr_register_operand (dest, mode)) + return true; + + return false; +} + +/* }}} */ +/* {{{ Functions and ABI. */ + +/* Implement TARGET_FUNCTION_VALUE. + + Define how to find the value returned by a function. + The register location is always the same, but the mode depends on + VALTYPE. */ + +static rtx +gcn_function_value (const_tree valtype, const_tree, bool) +{ + machine_mode mode = TYPE_MODE (valtype); + + if (INTEGRAL_TYPE_P (valtype) + && GET_MODE_CLASS (mode) == MODE_INT + && GET_MODE_SIZE (mode) < 4) + mode = SImode; + + return gen_rtx_REG (mode, SGPR_REGNO (RETURN_VALUE_REG)); +} + +/* Implement TARGET_FUNCTION_VALUE_REGNO_P. + + Return true if N is a possible register number for the function return + value. */ + +static bool +gcn_function_value_regno_p (const unsigned int n) +{ + return n == RETURN_VALUE_REG; +} + +/* Calculate the number of registers required to hold a function argument + of MODE and TYPE. */ + +static int +num_arg_regs (machine_mode mode, const_tree type) +{ + int size; + + if (targetm.calls.must_pass_in_stack (mode, type)) + return 0; + + if (type && mode == BLKmode) + size = int_size_in_bytes (type); + else + size = GET_MODE_SIZE (mode); + + return (size + UNITS_PER_WORD - 1) / UNITS_PER_WORD; +} + +/* Implement TARGET_STRICT_ARGUMENT_NAMING. + + Return true if the location where a function argument is passed + depends on whether or not it is a named argument + + For gcn, we know how to handle functions declared as stdarg: by + passing an extra pointer to the unnamed arguments. However, the + Fortran frontend can produce a different situation, where a + function pointer is declared with no arguments, but the actual + function and calls to it take more arguments. In that case, we + want to ensure the call matches the definition of the function. */ + +static bool +gcn_strict_argument_naming (cumulative_args_t cum_v) +{ + CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); + + return cum->fntype == NULL_TREE || stdarg_p (cum->fntype); +} + +/* Implement TARGET_PRETEND_OUTGOING_VARARGS_NAMED. + + See comment on gcn_strict_argument_naming. */ + +static bool +gcn_pretend_outgoing_varargs_named (cumulative_args_t cum_v) +{ + return !gcn_strict_argument_naming (cum_v); +} + +/* Implement TARGET_FUNCTION_ARG. + + Return an RTX indicating whether a function argument is passed in a register + and if so, which register. */ + +static rtx +gcn_function_arg (cumulative_args_t cum_v, machine_mode mode, const_tree type, + bool named) +{ + CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); + if (cum->normal_function) + { + if (!named || mode == VOIDmode) + return 0; + + if (targetm.calls.must_pass_in_stack (mode, type)) + return 0; + + int reg_num = FIRST_PARM_REG + cum->num; + int num_regs = num_arg_regs (mode, type); + if (num_regs > 0) + while (reg_num % num_regs != 0) + reg_num++; + if (reg_num + num_regs <= FIRST_PARM_REG + NUM_PARM_REGS) + return gen_rtx_REG (mode, reg_num); + } + else + { + if (cum->num >= cum->args.nargs) + { + cum->offset = (cum->offset + TYPE_ALIGN (type) / 8 - 1) + & -(TYPE_ALIGN (type) / 8); + cfun->machine->kernarg_segment_alignment + = MAX ((unsigned) cfun->machine->kernarg_segment_alignment, + TYPE_ALIGN (type) / 8); + rtx addr = gen_rtx_REG (DImode, + cum->args.reg[KERNARG_SEGMENT_PTR_ARG]); + if (cum->offset) + addr = gen_rtx_PLUS (DImode, addr, + gen_int_mode (cum->offset, DImode)); + rtx mem = gen_rtx_MEM (mode, addr); + set_mem_attributes (mem, const_cast(type), 1); + set_mem_addr_space (mem, ADDR_SPACE_SCALAR_FLAT); + MEM_READONLY_P (mem) = 1; + return mem; + } + + int a = cum->args.order[cum->num]; + if (mode != gcn_kernel_arg_types[a].mode) + { + error ("wrong type of argument %s", gcn_kernel_arg_types[a].name); + return 0; + } + return gen_rtx_REG ((machine_mode) gcn_kernel_arg_types[a].mode, + cum->args.reg[a]); + } + return 0; +} + +/* Implement TARGET_FUNCTION_ARG_ADVANCE. + + Updates the summarizer variable pointed to by CUM_V to advance past an + argument in the argument list. */ + +static void +gcn_function_arg_advance (cumulative_args_t cum_v, machine_mode mode, + const_tree type, bool named) +{ + CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); + + if (cum->normal_function) + { + if (!named) + return; + + int num_regs = num_arg_regs (mode, type); + if (num_regs > 0) + while ((FIRST_PARM_REG + cum->num) % num_regs != 0) + cum->num++; + cum->num += num_regs; + } + else + { + if (cum->num < cum->args.nargs) + cum->num++; + else + { + cum->offset += tree_to_uhwi (TYPE_SIZE_UNIT (type)); + cfun->machine->kernarg_segment_byte_size = cum->offset; + } + } +} + +/* Implement TARGET_ARG_PARTIAL_BYTES. + + Returns the number of bytes at the beginning of an argument that must be put + in registers. The value must be zero for arguments that are passed entirely + in registers or that are entirely pushed on the stack. */ + +static int +gcn_arg_partial_bytes (cumulative_args_t cum_v, machine_mode mode, tree type, + bool named) +{ + CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); + + if (!named) + return 0; + + if (targetm.calls.must_pass_in_stack (mode, type)) + return 0; + + if (cum->num >= NUM_PARM_REGS) + return 0; + + /* If the argument fits entirely in registers, return 0. */ + if (cum->num + num_arg_regs (mode, type) <= NUM_PARM_REGS) + return 0; + + return (NUM_PARM_REGS - cum->num) * UNITS_PER_WORD; +} + +/* A normal function which takes a pointer argument (to a scalar) may be + passed a pointer to LDS space (via a high-bits-set aperture), and that only + works with FLAT addressing, not GLOBAL. Force FLAT addressing if the + function has an incoming pointer-to-scalar parameter. */ + +static void +gcn_detect_incoming_pointer_arg (tree fndecl) +{ + gcc_assert (cfun && cfun->machine); + + for (tree arg = TYPE_ARG_TYPES (TREE_TYPE (fndecl)); + arg; + arg = TREE_CHAIN (arg)) + if (POINTER_TYPE_P (TREE_VALUE (arg)) + && !AGGREGATE_TYPE_P (TREE_TYPE (TREE_VALUE (arg)))) + cfun->machine->use_flat_addressing = true; +} + +/* Implement INIT_CUMULATIVE_ARGS, via gcn.h. + + Initialize a variable CUM of type CUMULATIVE_ARGS for a call to a function + whose data type is FNTYPE. For a library call, FNTYPE is 0. */ + +void +gcn_init_cumulative_args (CUMULATIVE_ARGS *cum /* Argument info to init */ , + tree fntype /* tree ptr for function decl */ , + rtx libname /* SYMBOL_REF of library name or 0 */ , + tree fndecl, int caller) +{ + memset (cum, 0, sizeof (*cum)); + cum->fntype = fntype; + if (libname) + { + gcc_assert (cfun && cfun->machine); + cum->normal_function = true; + if (!caller) + { + cfun->machine->normal_function = true; + gcn_detect_incoming_pointer_arg (fndecl); + } + return; + } + tree attr = NULL; + if (fndecl) + attr = lookup_attribute ("amdgpu_hsa_kernel", DECL_ATTRIBUTES (fndecl)); + if (fndecl && !attr) + attr = lookup_attribute ("amdgpu_hsa_kernel", + TYPE_ATTRIBUTES (TREE_TYPE (fndecl))); + if (!attr && fntype) + attr = lookup_attribute ("amdgpu_hsa_kernel", TYPE_ATTRIBUTES (fntype)); + /* Handle main () as kernel, so we can run testsuite. + Handle OpenACC kernels similarly to main. */ + if (!attr && !caller && fndecl + && (MAIN_NAME_P (DECL_NAME (fndecl)) + || lookup_attribute ("omp target entrypoint", + DECL_ATTRIBUTES (fndecl)) != NULL_TREE)) + gcn_parse_amdgpu_hsa_kernel_attribute (&cum->args, NULL_TREE); + else + { + if (!attr || caller) + { + gcc_assert (cfun && cfun->machine); + cum->normal_function = true; + if (!caller) + cfun->machine->normal_function = true; + } + gcn_parse_amdgpu_hsa_kernel_attribute + (&cum->args, attr ? TREE_VALUE (attr) : NULL_TREE); + } + cfun->machine->args = cum->args; + if (!caller && cfun->machine->normal_function) + gcn_detect_incoming_pointer_arg (fndecl); +} + +static bool +gcn_return_in_memory (const_tree type, const_tree ARG_UNUSED (fntype)) +{ + machine_mode mode = TYPE_MODE (type); + HOST_WIDE_INT size = int_size_in_bytes (type); + + if (AGGREGATE_TYPE_P (type)) + return true; + + if (mode == BLKmode) + return true; + + if (size > 2 * UNITS_PER_WORD) + return true; + + return false; +} + +/* Implement TARGET_PROMOTE_FUNCTION_MODE. + + Return the mode to use for outgoing function arguments. */ + +machine_mode +gcn_promote_function_mode (const_tree ARG_UNUSED (type), machine_mode mode, + int *ARG_UNUSED (punsignedp), + const_tree ARG_UNUSED (funtype), + int ARG_UNUSED (for_return)) +{ + if (GET_MODE_CLASS (mode) == MODE_INT && GET_MODE_SIZE (mode) < 4) + return SImode; + + return mode; +} + +/* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. + + Derived from hppa_gimplify_va_arg_expr. The generic routine doesn't handle + ARGS_GROW_DOWNWARDS. */ + +static tree +gcn_gimplify_va_arg_expr (tree valist, tree type, + gimple_seq *ARG_UNUSED (pre_p), + gimple_seq *ARG_UNUSED (post_p)) +{ + tree ptr = build_pointer_type (type); + tree valist_type; + tree t, u; + bool indirect; + + indirect = pass_by_reference (NULL, TYPE_MODE (type), type, 0); + if (indirect) + { + type = ptr; + ptr = build_pointer_type (type); + } + valist_type = TREE_TYPE (valist); + + /* Args grow down. Not handled by generic routines. */ + + u = fold_convert (sizetype, size_in_bytes (type)); + u = fold_build1 (NEGATE_EXPR, sizetype, u); + t = fold_build_pointer_plus (valist, u); + + /* Align to 8 byte boundary. */ + + u = build_int_cst (TREE_TYPE (t), -8); + t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, u); + t = fold_convert (valist_type, t); + + t = build2 (MODIFY_EXPR, valist_type, valist, t); + + t = fold_convert (ptr, t); + t = build_va_arg_indirect_ref (t); + + if (indirect) + t = build_va_arg_indirect_ref (t); + + return t; +} + +/* Calculate stack offsets needed to create prologues and epilogues. */ + +static struct machine_function * +gcn_compute_frame_offsets (void) +{ + machine_function *offsets = cfun->machine; + + if (reload_completed) + return offsets; + + offsets->need_frame_pointer = frame_pointer_needed; + + offsets->outgoing_args_size = crtl->outgoing_args_size; + offsets->pretend_size = crtl->args.pretend_args_size; + + offsets->local_vars = get_frame_size (); + + offsets->lr_needs_saving = (!leaf_function_p () + || df_regs_ever_live_p (LR_REGNUM) + || df_regs_ever_live_p (LR_REGNUM + 1)); + + offsets->callee_saves = offsets->lr_needs_saving ? 8 : 0; + + for (int regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) + if ((df_regs_ever_live_p (regno) && !call_used_regs[regno]) + || ((regno & ~1) == HARD_FRAME_POINTER_REGNUM + && frame_pointer_needed)) + offsets->callee_saves += (VGPR_REGNO_P (regno) ? 256 : 4); + + /* Round up to 64-bit boundary to maintain stack alignment. */ + offsets->callee_saves = (offsets->callee_saves + 7) & ~7; + + return offsets; +} + +/* Insert code into the prologue or epilogue to store or load any + callee-save register to/from the stack. + + Helper function for gcn_expand_prologue and gcn_expand_epilogue. */ + +static void +move_callee_saved_registers (rtx sp, machine_function *offsets, + bool prologue) +{ + int regno, offset, saved_scalars; + rtx exec = gen_rtx_REG (DImode, EXEC_REG); + rtx vcc = gen_rtx_REG (DImode, VCC_LO_REG); + rtx offreg = gen_rtx_REG (SImode, SGPR_REGNO (22)); + rtx as = gen_rtx_CONST_INT (VOIDmode, STACK_ADDR_SPACE); + HOST_WIDE_INT exec_set = 0; + int offreg_set = 0; + + start_sequence (); + + /* Move scalars into two vector registers. */ + for (regno = 0, saved_scalars = 0; regno < FIRST_VGPR_REG; regno++) + if ((df_regs_ever_live_p (regno) && !call_used_regs[regno]) + || ((regno & ~1) == LINK_REGNUM && offsets->lr_needs_saving) + || ((regno & ~1) == HARD_FRAME_POINTER_REGNUM + && offsets->need_frame_pointer)) + { + rtx reg = gen_rtx_REG (SImode, regno); + rtx vreg = gen_rtx_REG (V64SImode, + VGPR_REGNO (6 + (saved_scalars / 64))); + int lane = saved_scalars % 64; + + if (prologue) + emit_insn (gen_vec_setv64si (vreg, reg, GEN_INT (lane))); + else + emit_insn (gen_vec_extractv64sisi (reg, vreg, GEN_INT (lane))); + + saved_scalars++; + } + + rtx move_scalars = get_insns (); + end_sequence (); + start_sequence (); + + /* Ensure that all vector lanes are moved. */ + exec_set = -1; + emit_move_insn (exec, GEN_INT (exec_set)); + + /* Set up a vector stack pointer. */ + rtx _0_1_2_3 = gen_rtx_REG (V64SImode, VGPR_REGNO (1)); + rtx _0_4_8_12 = gen_rtx_REG (V64SImode, VGPR_REGNO (3)); + emit_insn (gen_ashlv64si3_exec (_0_4_8_12, _0_1_2_3, GEN_INT (2), + gcn_gen_undef (V64SImode), exec)); + rtx vsp = gen_rtx_REG (V64DImode, VGPR_REGNO (4)); + emit_insn (gen_vec_duplicatev64di_exec (vsp, sp, gcn_gen_undef (V64DImode), + exec)); + emit_insn (gen_addv64si3_vcc_exec (gcn_operand_part (V64SImode, vsp, 0), + gcn_operand_part (V64SImode, vsp, 0), + _0_4_8_12, vcc, gcn_gen_undef (V64SImode), + exec)); + emit_insn (gen_addcv64si3_exec (gcn_operand_part (V64SImode, vsp, 1), + gcn_operand_part (V64SImode, vsp, 1), + const0_rtx, vcc, vcc, + gcn_gen_undef (V64SImode), exec)); + + /* Move vectors. */ + for (regno = FIRST_VGPR_REG, offset = offsets->pretend_size; + regno < FIRST_PSEUDO_REGISTER; regno++) + if ((df_regs_ever_live_p (regno) && !call_used_regs[regno]) + || (regno == VGPR_REGNO (6) && saved_scalars > 0) + || (regno == VGPR_REGNO (7) && saved_scalars > 63)) + { + rtx reg = gen_rtx_REG (V64SImode, regno); + int size = 256; + + if (regno == VGPR_REGNO (6) && saved_scalars < 64) + size = saved_scalars * 4; + else if (regno == VGPR_REGNO (7) && saved_scalars < 128) + size = (saved_scalars - 64) * 4; + + if (size != 256 || exec_set != -1) + { + exec_set = ((unsigned HOST_WIDE_INT) 1 << (size / 4)) - 1; + emit_move_insn (exec, gen_int_mode (exec_set, DImode)); + } + + if (prologue) + emit_insn (gen_scatterv64si_insn_1offset_exec (vsp, const0_rtx, reg, + as, const0_rtx, exec)); + else + emit_insn (gen_gatherv64si_insn_1offset_exec + (reg, vsp, const0_rtx, as, const0_rtx, + gcn_gen_undef (V64SImode), exec)); + + /* Move our VSP to the next stack entry. */ + if (offreg_set != size) + { + offreg_set = size; + emit_move_insn (offreg, GEN_INT (size)); + } + if (exec_set != -1) + { + exec_set = -1; + emit_move_insn (exec, GEN_INT (exec_set)); + } + emit_insn (gen_addv64si3_vcc_dup_exec + (gcn_operand_part (V64SImode, vsp, 0), + offreg, gcn_operand_part (V64SImode, vsp, 0), + vcc, gcn_gen_undef (V64SImode), exec)); + emit_insn (gen_addcv64si3_exec + (gcn_operand_part (V64SImode, vsp, 1), + gcn_operand_part (V64SImode, vsp, 1), + const0_rtx, vcc, vcc, gcn_gen_undef (V64SImode), exec)); + + offset += size; + } + + rtx move_vectors = get_insns (); + end_sequence (); + + if (prologue) + { + emit_insn (move_scalars); + emit_insn (move_vectors); + } + else + { + emit_insn (move_vectors); + emit_insn (move_scalars); + } +} + +/* Generate prologue. Called from gen_prologue during pro_and_epilogue pass. + + For a non-kernel function, the stack layout looks like this (interim), + growing *upwards*: + + hi | + ... + |__________________| <-- current SP + | outgoing args | + |__________________| + | (alloca space) | + |__________________| + | local vars | + |__________________| <-- FP/hard FP + | callee-save regs | + |__________________| <-- soft arg pointer + | pretend args | + |__________________| <-- incoming SP + | incoming args | + lo |..................| + + This implies arguments (beyond the first N in registers) must grow + downwards (as, apparently, PA has them do). + + For a kernel function we have the simpler: + + hi | + ... + |__________________| <-- current SP + | outgoing args | + |__________________| + | (alloca space) | + |__________________| + | local vars | + lo |__________________| <-- FP/hard FP + +*/ + +void +gcn_expand_prologue () +{ + machine_function *offsets = gcn_compute_frame_offsets (); + + if (!cfun || !cfun->machine || cfun->machine->normal_function) + { + rtx sp = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); + rtx fp = gen_rtx_REG (Pmode, HARD_FRAME_POINTER_REGNUM); + + start_sequence (); + + if (offsets->pretend_size > 0) + { + /* FIXME: Do the actual saving of register pretend args to the stack. + Register order needs consideration. */ + } + + /* Save callee-save regs. */ + move_callee_saved_registers (sp, offsets, true); + + HOST_WIDE_INT sp_adjust = offsets->pretend_size + + offsets->callee_saves + + offsets->local_vars + offsets->outgoing_args_size; + if (sp_adjust > 0) + emit_insn (gen_adddi3_scc (sp, sp, gen_int_mode (sp_adjust, DImode))); + + if (offsets->need_frame_pointer) + emit_insn (gen_adddi3_scc (fp, sp, + gen_int_mode + (-(offsets->local_vars + + offsets->outgoing_args_size), + DImode))); + + rtx_insn *seq = get_insns (); + end_sequence (); + + /* FIXME: Prologue insns should have this flag set for debug output, etc. + but it causes issues for now. + for (insn = seq; insn; insn = NEXT_INSN (insn)) + if (INSN_P (insn)) + RTX_FRAME_RELATED_P (insn) = 1;*/ + + emit_insn (seq); + } + else + { + rtx wave_offset = gen_rtx_REG (SImode, + cfun->machine->args. + reg[PRIVATE_SEGMENT_WAVE_OFFSET_ARG]); + + if (TARGET_GCN5_PLUS) + { + /* v0 is reserved for constant zero so that "global" + memory instructions can have a nul-offset without + causing reloads. */ + emit_insn (gen_vec_duplicatev64si + (gen_rtx_REG (V64SImode, VGPR_REGNO (0)), const0_rtx)); + } + + if (cfun->machine->args.requested & (1 << FLAT_SCRATCH_INIT_ARG)) + { + rtx fs_init_lo = + gen_rtx_REG (SImode, + cfun->machine->args.reg[FLAT_SCRATCH_INIT_ARG]); + rtx fs_init_hi = + gen_rtx_REG (SImode, + cfun->machine->args.reg[FLAT_SCRATCH_INIT_ARG] + 1); + rtx fs_reg_lo = gen_rtx_REG (SImode, FLAT_SCRATCH_REG); + rtx fs_reg_hi = gen_rtx_REG (SImode, FLAT_SCRATCH_REG + 1); + + /*rtx queue = gen_rtx_REG(DImode, + cfun->machine->args.reg[QUEUE_PTR_ARG]); + rtx aperture = gen_rtx_MEM (SImode, + gen_rtx_PLUS (DImode, queue, + gen_int_mode (68, SImode))); + set_mem_addr_space (aperture, ADDR_SPACE_SCALAR_FLAT);*/ + + /* Set up flat_scratch. */ + emit_insn (gen_addsi3_scc (fs_reg_hi, fs_init_lo, wave_offset)); + emit_insn (gen_lshrsi3_scc (fs_reg_hi, fs_reg_hi, + gen_int_mode (8, SImode))); + emit_move_insn (fs_reg_lo, fs_init_hi); + } + + /* Set up frame pointer and stack pointer. */ + rtx sp = gen_rtx_REG (DImode, STACK_POINTER_REGNUM); + rtx fp = gen_rtx_REG (DImode, HARD_FRAME_POINTER_REGNUM); + rtx fp_hi = simplify_gen_subreg (SImode, fp, DImode, 4); + rtx fp_lo = simplify_gen_subreg (SImode, fp, DImode, 0); + + HOST_WIDE_INT sp_adjust = (offsets->local_vars + + offsets->outgoing_args_size); + + /* Initialise FP and SP from the buffer descriptor in s[0:3]. */ + emit_move_insn (fp_lo, gen_rtx_REG (SImode, 0)); + emit_insn (gen_andsi3_scc (fp_hi, gen_rtx_REG (SImode, 1), + gen_int_mode (0xffff, SImode))); + emit_insn (gen_addsi3_scc (fp_lo, fp_lo, wave_offset)); + emit_insn (gen_addcsi3_scalar_zero (fp_hi, fp_hi, + gen_rtx_REG (BImode, SCC_REG))); + + if (sp_adjust > 0) + emit_insn (gen_adddi3_scc (sp, fp, gen_int_mode (sp_adjust, DImode))); + else + emit_move_insn (sp, fp); + + /* Make sure the flat scratch reg doesn't get optimised away. */ + emit_insn (gen_prologue_use (gen_rtx_REG (DImode, FLAT_SCRATCH_REG))); + } + + /* Ensure that the scheduler doesn't do anything unexpected. */ + emit_insn (gen_blockage ()); + + emit_move_insn (gen_rtx_REG (SImode, M0_REG), + gen_int_mode (LDS_SIZE, SImode)); + + emit_insn (gen_prologue_use (gen_rtx_REG (SImode, M0_REG))); + if (TARGET_GCN5_PLUS) + emit_insn (gen_prologue_use (gen_rtx_REG (SImode, VGPR_REGNO (0)))); + + if (cfun && cfun->machine && !cfun->machine->normal_function && flag_openmp) + { + /* OpenMP kernels have an implicit call to gomp_gcn_enter_kernel. */ + rtx fn_reg = gen_rtx_REG (Pmode, FIRST_PARM_REG); + emit_move_insn (fn_reg, gen_rtx_SYMBOL_REF (Pmode, + "gomp_gcn_enter_kernel")); + emit_call_insn (gen_gcn_indirect_call (fn_reg, const0_rtx)); + } +} + +/* Generate epilogue. Called from gen_epilogue during pro_and_epilogue pass. + + See gcn_expand_prologue for stack details. */ + +void +gcn_expand_epilogue (void) +{ + /* Ensure that the scheduler doesn't do anything unexpected. */ + emit_insn (gen_blockage ()); + + if (!cfun || !cfun->machine || cfun->machine->normal_function) + { + machine_function *offsets = gcn_compute_frame_offsets (); + rtx sp = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); + rtx fp = gen_rtx_REG (Pmode, HARD_FRAME_POINTER_REGNUM); + + HOST_WIDE_INT sp_adjust = offsets->callee_saves + offsets->pretend_size; + + if (offsets->need_frame_pointer) + { + /* Restore old SP from the frame pointer. */ + if (sp_adjust > 0) + emit_insn (gen_subdi3 (sp, fp, gen_int_mode (sp_adjust, DImode))); + else + emit_move_insn (sp, fp); + } + else + { + /* Restore old SP from current SP. */ + sp_adjust += offsets->outgoing_args_size + offsets->local_vars; + + if (sp_adjust > 0) + emit_insn (gen_subdi3 (sp, sp, gen_int_mode (sp_adjust, DImode))); + } + + move_callee_saved_registers (sp, offsets, false); + + /* There's no explicit use of the link register on the return insn. Emit + one here instead. */ + if (offsets->lr_needs_saving) + emit_use (gen_rtx_REG (DImode, LINK_REGNUM)); + + /* Similar for frame pointer. */ + if (offsets->need_frame_pointer) + emit_use (gen_rtx_REG (DImode, HARD_FRAME_POINTER_REGNUM)); + } + else if (flag_openmp) + { + /* OpenMP kernels have an implicit call to gomp_gcn_exit_kernel. */ + rtx fn_reg = gen_rtx_REG (Pmode, FIRST_PARM_REG); + emit_move_insn (fn_reg, + gen_rtx_SYMBOL_REF (Pmode, "gomp_gcn_exit_kernel")); + emit_call_insn (gen_gcn_indirect_call (fn_reg, const0_rtx)); + } + else if (TREE_CODE (TREE_TYPE (DECL_RESULT (cfun->decl))) != VOID_TYPE) + { + /* Assume that an exit value compatible with gcn-run is expected. + That is, the third input parameter is an int*. + + We can't allocate any new registers, but the kernarg_reg is + dead after this, so we'll use that. */ + rtx kernarg_reg = gen_rtx_REG (DImode, cfun->machine->args.reg + [KERNARG_SEGMENT_PTR_ARG]); + rtx retptr_mem = gen_rtx_MEM (DImode, + gen_rtx_PLUS (DImode, kernarg_reg, + GEN_INT (16))); + set_mem_addr_space (retptr_mem, ADDR_SPACE_SCALAR_FLAT); + emit_move_insn (kernarg_reg, retptr_mem); + + rtx retval_mem = gen_rtx_MEM (SImode, kernarg_reg); + set_mem_addr_space (retval_mem, ADDR_SPACE_SCALAR_FLAT); + emit_move_insn (retval_mem, + gen_rtx_REG (SImode, SGPR_REGNO (RETURN_VALUE_REG))); + } + + emit_jump_insn (gen_gcn_return ()); +} + +/* Implement TARGET_CAN_ELIMINATE. + + Return true if the compiler is allowed to try to replace register number + FROM_REG with register number TO_REG. + + FIXME: is the default "true" not enough? Should this be a negative set? */ + +bool +gcn_can_eliminate_p (int /*from_reg */ , int to_reg) +{ + return (to_reg == HARD_FRAME_POINTER_REGNUM + || to_reg == STACK_POINTER_REGNUM); +} + +/* Implement INITIAL_ELIMINATION_OFFSET. + + Returns the initial difference between the specified pair of registers, in + terms of stack position. */ + +HOST_WIDE_INT +gcn_initial_elimination_offset (int from, int to) +{ + machine_function *offsets = gcn_compute_frame_offsets (); + + switch (from) + { + case ARG_POINTER_REGNUM: + if (to == STACK_POINTER_REGNUM) + return -(offsets->callee_saves + offsets->local_vars + + offsets->outgoing_args_size); + else if (to == FRAME_POINTER_REGNUM || to == HARD_FRAME_POINTER_REGNUM) + return -offsets->callee_saves; + else + gcc_unreachable (); + break; + + case FRAME_POINTER_REGNUM: + if (to == STACK_POINTER_REGNUM) + return -(offsets->local_vars + offsets->outgoing_args_size); + else if (to == HARD_FRAME_POINTER_REGNUM) + return 0; + else + gcc_unreachable (); + break; + + default: + gcc_unreachable (); + } +} + +/* Implement HARD_REGNO_RENAME_OK. + + Return true if it is permissible to rename a hard register from + FROM_REG to TO_REG. */ + +bool +gcn_hard_regno_rename_ok (unsigned int from_reg, unsigned int to_reg) +{ + if (from_reg == SCC_REG + || from_reg == VCC_LO_REG || from_reg == VCC_HI_REG + || from_reg == EXEC_LO_REG || from_reg == EXEC_HI_REG + || to_reg == SCC_REG + || to_reg == VCC_LO_REG || to_reg == VCC_HI_REG + || to_reg == EXEC_LO_REG || to_reg == EXEC_HI_REG) + return false; + + /* Allow the link register to be used if it was saved. */ + if ((to_reg & ~1) == LINK_REGNUM) + return !cfun || cfun->machine->lr_needs_saving; + + /* Allow the registers used for the static chain to be used if the chain is + not in active use. */ + if ((to_reg & ~1) == STATIC_CHAIN_REGNUM) + return !cfun + || !(cfun->static_chain_decl + && df_regs_ever_live_p (STATIC_CHAIN_REGNUM) + && df_regs_ever_live_p (STATIC_CHAIN_REGNUM + 1)); + + return true; +} + +/* Implement HARD_REGNO_CALLER_SAVE_MODE. + + Which mode is required for saving NREGS of a pseudo-register in + call-clobbered hard register REGNO. */ + +machine_mode +gcn_hard_regno_caller_save_mode (unsigned int regno, unsigned int nregs, + machine_mode regmode) +{ + machine_mode result = choose_hard_reg_mode (regno, nregs, false); + + if (VECTOR_MODE_P (result) && !VECTOR_MODE_P (regmode)) + result = (nregs == 1 ? SImode : DImode); + + return result; +} + +/* Implement TARGET_ASM_TRAMPOLINE_TEMPLATE. + + Output assembler code for a block containing the constant parts + of a trampoline, leaving space for the variable parts. */ + +static void +gcn_asm_trampoline_template (FILE *f) +{ + /* The source operand of the move instructions must be a 32-bit + constant following the opcode. */ + asm_fprintf (f, "\ts_mov_b32\ts%i, 0xffff\n", STATIC_CHAIN_REGNUM); + asm_fprintf (f, "\ts_mov_b32\ts%i, 0xffff\n", STATIC_CHAIN_REGNUM + 1); + asm_fprintf (f, "\ts_mov_b32\ts%i, 0xffff\n", CC_SAVE_REG); + asm_fprintf (f, "\ts_mov_b32\ts%i, 0xffff\n", CC_SAVE_REG + 1); + asm_fprintf (f, "\ts_setpc_b64\ts[%i:%i]\n", CC_SAVE_REG, CC_SAVE_REG + 1); +} + +/* Implement TARGET_TRAMPOLINE_INIT. + + Emit RTL insns to initialize the variable parts of a trampoline. + FNDECL is the decl of the target address, M_TRAMP is a MEM for + the trampoline, and CHAIN_VALUE is an RTX for the static chain + to be passed to the target function. */ + +static void +gcn_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) +{ + emit_block_move (m_tramp, assemble_trampoline_template (), + GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL); + + rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); + rtx chain_value_reg = copy_to_reg (chain_value); + rtx fnaddr_reg = copy_to_reg (fnaddr); + + for (int i = 0; i < 4; i++) + { + rtx mem = adjust_address (m_tramp, SImode, i * 8 + 4); + rtx reg = i < 2 ? chain_value_reg : fnaddr_reg; + emit_move_insn (mem, gen_rtx_SUBREG (SImode, reg, (i % 2) * 4)); + } + + rtx tramp_addr = XEXP (m_tramp, 0); + emit_insn (gen_clear_icache (tramp_addr, + plus_constant (ptr_mode, tramp_addr, + TRAMPOLINE_SIZE))); +} + +/* }}} */ +/* {{{ Miscellaneous. */ + +/* Implement TARGET_CANNOT_COPY_INSN_P. + + Return true if INSN must not be duplicated. */ + +static bool +gcn_cannot_copy_insn_p (rtx_insn *insn) +{ + if (recog_memoized (insn) == CODE_FOR_gcn_wavefront_barrier) + return true; + + return false; +} + +/* Implement TARGET_DEBUG_UNWIND_INFO. + + Defines the mechanism that will be used for describing frame unwind + information to the debugger. */ + +static enum unwind_info_type +gcn_debug_unwind_info () +{ + /* No support for debug info, yet. */ + return UI_NONE; +} + +/* Determine if there is a suitable hardware conversion instruction. + Used primarily by the machine description. */ + +bool +gcn_valid_cvt_p (machine_mode from, machine_mode to, enum gcn_cvt_t op) +{ + if (VECTOR_MODE_P (from) != VECTOR_MODE_P (to)) + return false; + + if (VECTOR_MODE_P (from)) + { + from = GET_MODE_INNER (from); + to = GET_MODE_INNER (to); + } + + switch (op) + { + case fix_trunc_cvt: + case fixuns_trunc_cvt: + if (GET_MODE_CLASS (from) != MODE_FLOAT + || GET_MODE_CLASS (to) != MODE_INT) + return false; + break; + case float_cvt: + case floatuns_cvt: + if (GET_MODE_CLASS (from) != MODE_INT + || GET_MODE_CLASS (to) != MODE_FLOAT) + return false; + break; + case extend_cvt: + if (GET_MODE_CLASS (from) != MODE_FLOAT + || GET_MODE_CLASS (to) != MODE_FLOAT + || GET_MODE_SIZE (from) >= GET_MODE_SIZE (to)) + return false; + break; + case trunc_cvt: + if (GET_MODE_CLASS (from) != MODE_FLOAT + || GET_MODE_CLASS (to) != MODE_FLOAT + || GET_MODE_SIZE (from) <= GET_MODE_SIZE (to)) + return false; + break; + } + + return ((to == HImode && from == HFmode) + || (to == SImode && (from == SFmode || from == DFmode)) + || (to == HFmode && (from == HImode || from == SFmode)) + || (to == SFmode && (from == SImode || from == HFmode + || from == DFmode)) + || (to == DFmode && (from == SImode || from == SFmode))); +} + +/* Implement both TARGET_ASM_CONSTRUCTOR and TARGET_ASM_DESTRUCTOR. + + The current loader does not support running code outside "main". This + hook implementation can be replaced or removed when that changes. */ + +void +gcn_disable_constructors (rtx symbol, int priority __attribute__ ((unused))) +{ + tree d = SYMBOL_REF_DECL (symbol); + location_t l = d ? DECL_SOURCE_LOCATION (d) : UNKNOWN_LOCATION; + + sorry_at (l, "GCN does not support static constructors or destructors"); +} + +/* }}} */ +/* {{{ Costs. */ + +/* Implement TARGET_RTX_COSTS. + + Compute a (partial) cost for rtx X. Return true if the complete + cost has been computed, and false if subexpressions should be + scanned. In either case, *TOTAL contains the cost result. */ + +static bool +gcn_rtx_costs (rtx x, machine_mode, int, int, int *total, bool) +{ + enum rtx_code code = GET_CODE (x); + switch (code) + { + case CONST: + case CONST_DOUBLE: + case CONST_VECTOR: + case CONST_INT: + if (gcn_inline_constant_p (x)) + *total = 0; + else if (code == CONST_INT + && ((unsigned HOST_WIDE_INT) INTVAL (x) + 0x8000) < 0x10000) + *total = 1; + else if (gcn_constant_p (x)) + *total = 2; + else + *total = vgpr_vector_mode_p (GET_MODE (x)) ? 64 : 4; + return true; + + case DIV: + *total = 100; + return false; + + default: + *total = 3; + return false; + } +} + +/* Implement TARGET_MEMORY_MOVE_COST. + + Return the cost of moving data of mode M between a + register and memory. A value of 2 is the default; this cost is + relative to those in `REGISTER_MOVE_COST'. + + This function is used extensively by register_move_cost that is used to + build tables at startup. Make it inline in this case. + When IN is 2, return maximum of in and out move cost. + + If moving between registers and memory is more expensive than + between two registers, you should define this macro to express the + relative cost. + + Model also increased moving costs of QImode registers in non + Q_REGS classes. */ + +#define LOAD_COST 32 +#define STORE_COST 32 +static int +gcn_memory_move_cost (machine_mode mode, reg_class_t regclass, bool in) +{ + int nregs = CEIL (GET_MODE_SIZE (mode), 4); + switch (regclass) + { + case SCC_CONDITIONAL_REG: + case VCCZ_CONDITIONAL_REG: + case VCC_CONDITIONAL_REG: + case EXECZ_CONDITIONAL_REG: + case ALL_CONDITIONAL_REGS: + case SGPR_REGS: + case SGPR_EXEC_REGS: + case EXEC_MASK_REG: + case SGPR_VOP_SRC_REGS: + case SGPR_MEM_SRC_REGS: + case SGPR_SRC_REGS: + case SGPR_DST_REGS: + case GENERAL_REGS: + case AFP_REGS: + if (!in) + return (STORE_COST + 2) * nregs; + return LOAD_COST * nregs; + case VGPR_REGS: + if (in) + return (LOAD_COST + 2) * nregs; + return STORE_COST * nregs; + case ALL_REGS: + case ALL_GPR_REGS: + case SRCDST_REGS: + if (in) + return (LOAD_COST + 2) * nregs; + return (STORE_COST + 2) * nregs; + default: + gcc_unreachable (); + } +} + +/* Implement TARGET_REGISTER_MOVE_COST. + + Return the cost of moving data from a register in class CLASS1 to + one in class CLASS2. Base value is 2. */ + +static int +gcn_register_move_cost (machine_mode, reg_class_t dst, reg_class_t src) +{ + /* Increase cost of moving from and to vector registers. While this is + fast in hardware (I think), it has hidden cost of setting up the exec + flags. */ + if ((src < VGPR_REGS) != (dst < VGPR_REGS)) + return 4; + return 2; +} + +/* }}} */ +/* {{{ Builtins. */ + +/* Type codes used by GCN built-in definitions. */ + +enum gcn_builtin_type_index +{ + GCN_BTI_END_OF_PARAMS, + + GCN_BTI_VOID, + GCN_BTI_BOOL, + GCN_BTI_INT, + GCN_BTI_UINT, + GCN_BTI_SIZE_T, + GCN_BTI_LLINT, + GCN_BTI_LLUINT, + GCN_BTI_EXEC, + + GCN_BTI_SF, + GCN_BTI_V64SI, + GCN_BTI_V64SF, + GCN_BTI_V64PTR, + GCN_BTI_SIPTR, + GCN_BTI_SFPTR, + GCN_BTI_VOIDPTR, + + GCN_BTI_LDS_VOIDPTR, + + GCN_BTI_MAX +}; + +static GTY(()) tree gcn_builtin_types[GCN_BTI_MAX]; + +#define exec_type_node (gcn_builtin_types[GCN_BTI_EXEC]) +#define sf_type_node (gcn_builtin_types[GCN_BTI_SF]) +#define v64si_type_node (gcn_builtin_types[GCN_BTI_V64SI]) +#define v64sf_type_node (gcn_builtin_types[GCN_BTI_V64SF]) +#define v64ptr_type_node (gcn_builtin_types[GCN_BTI_V64PTR]) +#define siptr_type_node (gcn_builtin_types[GCN_BTI_SIPTR]) +#define sfptr_type_node (gcn_builtin_types[GCN_BTI_SFPTR]) +#define voidptr_type_node (gcn_builtin_types[GCN_BTI_VOIDPTR]) +#define size_t_type_node (gcn_builtin_types[GCN_BTI_SIZE_T]) + +static rtx gcn_expand_builtin_1 (tree, rtx, rtx, machine_mode, int, + struct gcn_builtin_description *); +static rtx gcn_expand_builtin_binop (tree, rtx, rtx, machine_mode, int, + struct gcn_builtin_description *); + +struct gcn_builtin_description; +typedef rtx (*gcn_builtin_expander) (tree, rtx, rtx, machine_mode, int, + struct gcn_builtin_description *); + +enum gcn_builtin_type +{ + B_UNIMPLEMENTED, /* Sorry out */ + B_INSN, /* Emit a pattern */ + B_OVERLOAD /* Placeholder for an overloaded function */ +}; + +struct gcn_builtin_description +{ + int fcode; + int icode; + const char *name; + enum gcn_builtin_type type; + /* The first element of parm is always the return type. The rest + are a zero terminated list of parameters. */ + int parm[6]; + gcn_builtin_expander expander; +}; + +/* Read in the GCN builtins from gcn-builtins.def. */ + +extern GTY(()) struct gcn_builtin_description gcn_builtins[GCN_BUILTIN_MAX]; + +struct gcn_builtin_description gcn_builtins[] = { +#define DEF_BUILTIN(fcode, icode, name, type, params, expander) \ + {GCN_BUILTIN_ ## fcode, icode, name, type, params, expander}, + +#define DEF_BUILTIN_BINOP_INT_FP(fcode, ic, name) \ + {GCN_BUILTIN_ ## fcode ## _V64SI, \ + CODE_FOR_ ## ic ##v64si3_exec, name "_v64int", B_INSN, \ + {GCN_BTI_V64SI, GCN_BTI_EXEC, GCN_BTI_V64SI, GCN_BTI_V64SI, \ + GCN_BTI_V64SI, GCN_BTI_END_OF_PARAMS}, gcn_expand_builtin_binop}, \ + {GCN_BUILTIN_ ## fcode ## _V64SI_unspec, \ + CODE_FOR_ ## ic ##v64si3_exec, name "_v64int_unspec", B_INSN, \ + {GCN_BTI_V64SI, GCN_BTI_EXEC, GCN_BTI_V64SI, GCN_BTI_V64SI, \ + GCN_BTI_END_OF_PARAMS}, gcn_expand_builtin_binop}, + +#include "gcn-builtins.def" +#undef DEF_BUILTIN_BINOP_INT_FP +#undef DEF_BUILTIN +}; + +static GTY(()) tree gcn_builtin_decls[GCN_BUILTIN_MAX]; + +/* Implement TARGET_BUILTIN_DECL. + + Return the GCN builtin for CODE. */ + +tree +gcn_builtin_decl (unsigned code, bool ARG_UNUSED (initialize_p)) +{ + if (code >= GCN_BUILTIN_MAX) + return error_mark_node; + + return gcn_builtin_decls[code]; +} + +/* Helper function for gcn_init_builtins. */ + +static void +gcn_init_builtin_types (void) +{ + gcn_builtin_types[GCN_BTI_VOID] = void_type_node; + gcn_builtin_types[GCN_BTI_BOOL] = boolean_type_node; + gcn_builtin_types[GCN_BTI_INT] = intSI_type_node; + gcn_builtin_types[GCN_BTI_UINT] = unsigned_type_for (intSI_type_node); + gcn_builtin_types[GCN_BTI_SIZE_T] = size_type_node; + gcn_builtin_types[GCN_BTI_LLINT] = intDI_type_node; + gcn_builtin_types[GCN_BTI_LLUINT] = unsigned_type_for (intDI_type_node); + + exec_type_node = unsigned_intDI_type_node; + sf_type_node = float32_type_node; + v64si_type_node = build_vector_type (intSI_type_node, 64); + v64sf_type_node = build_vector_type (float_type_node, 64); + v64ptr_type_node = build_vector_type (unsigned_intDI_type_node + /*build_pointer_type + (integer_type_node) */ + , 64); + tree tmp = build_distinct_type_copy (intSI_type_node); + TYPE_ADDR_SPACE (tmp) = ADDR_SPACE_FLAT; + siptr_type_node = build_pointer_type (tmp); + + tmp = build_distinct_type_copy (float_type_node); + TYPE_ADDR_SPACE (tmp) = ADDR_SPACE_FLAT; + sfptr_type_node = build_pointer_type (tmp); + + tmp = build_distinct_type_copy (void_type_node); + TYPE_ADDR_SPACE (tmp) = ADDR_SPACE_FLAT; + voidptr_type_node = build_pointer_type (tmp); + + tmp = build_distinct_type_copy (void_type_node); + TYPE_ADDR_SPACE (tmp) = ADDR_SPACE_LDS; + gcn_builtin_types[GCN_BTI_LDS_VOIDPTR] = build_pointer_type (tmp); +} + +/* Implement TARGET_INIT_BUILTINS. + + Set up all builtin functions for this target. */ + +static void +gcn_init_builtins (void) +{ + gcn_init_builtin_types (); + + struct gcn_builtin_description *d; + unsigned int i; + for (i = 0, d = gcn_builtins; i < GCN_BUILTIN_MAX; i++, d++) + { + tree p; + char name[64]; /* build_function will make a copy. */ + int parm; + + /* FIXME: Is this necessary/useful? */ + if (d->name == 0) + continue; + + /* Find last parm. */ + for (parm = 1; d->parm[parm] != GCN_BTI_END_OF_PARAMS; parm++) + ; + + p = void_list_node; + while (parm > 1) + p = tree_cons (NULL_TREE, gcn_builtin_types[d->parm[--parm]], p); + + p = build_function_type (gcn_builtin_types[d->parm[0]], p); + + sprintf (name, "__builtin_gcn_%s", d->name); + gcn_builtin_decls[i] + = add_builtin_function (name, p, i, BUILT_IN_MD, NULL, NULL_TREE); + + /* These builtins don't throw. */ + TREE_NOTHROW (gcn_builtin_decls[i]) = 1; + } + +/* FIXME: remove the ifdef once OpenACC support is merged upstream. */ +#ifdef BUILT_IN_GOACC_SINGLE_START + /* These builtins need to take/return an LDS pointer: override the generic + versions here. */ + + set_builtin_decl (BUILT_IN_GOACC_SINGLE_START, + gcn_builtin_decls[GCN_BUILTIN_ACC_SINGLE_START], false); + + set_builtin_decl (BUILT_IN_GOACC_SINGLE_COPY_START, + gcn_builtin_decls[GCN_BUILTIN_ACC_SINGLE_COPY_START], + false); + + set_builtin_decl (BUILT_IN_GOACC_SINGLE_COPY_END, + gcn_builtin_decls[GCN_BUILTIN_ACC_SINGLE_COPY_END], + false); + + set_builtin_decl (BUILT_IN_GOACC_BARRIER, + gcn_builtin_decls[GCN_BUILTIN_ACC_BARRIER], false); +#endif +} + +/* Expand the CMP_SWAP GCN builtins. We have our own versions that do + not require taking the address of any object, other than the memory + cell being operated on. + + Helper function for gcn_expand_builtin_1. */ + +static rtx +gcn_expand_cmp_swap (tree exp, rtx target) +{ + machine_mode mode = TYPE_MODE (TREE_TYPE (exp)); + addr_space_t as + = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (CALL_EXPR_ARG (exp, 0)))); + machine_mode as_mode = gcn_addr_space_address_mode (as); + + if (!target) + target = gen_reg_rtx (mode); + + rtx addr = expand_expr (CALL_EXPR_ARG (exp, 0), + NULL_RTX, as_mode, EXPAND_NORMAL); + rtx cmp = expand_expr (CALL_EXPR_ARG (exp, 1), + NULL_RTX, mode, EXPAND_NORMAL); + rtx src = expand_expr (CALL_EXPR_ARG (exp, 2), + NULL_RTX, mode, EXPAND_NORMAL); + rtx pat; + + rtx mem = gen_rtx_MEM (mode, force_reg (as_mode, addr)); + set_mem_addr_space (mem, as); + + if (!REG_P (cmp)) + cmp = copy_to_mode_reg (mode, cmp); + if (!REG_P (src)) + src = copy_to_mode_reg (mode, src); + + if (mode == SImode) + pat = gen_sync_compare_and_swapsi (target, mem, cmp, src); + else + pat = gen_sync_compare_and_swapdi (target, mem, cmp, src); + + emit_insn (pat); + + return target; +} + +/* Expand many different builtins. + + Intended for use in gcn-builtins.def. */ + +static rtx +gcn_expand_builtin_1 (tree exp, rtx target, rtx /*subtarget */ , + machine_mode /*mode */ , int ignore, + struct gcn_builtin_description *) +{ + tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); + switch (DECL_FUNCTION_CODE (fndecl)) + { + case GCN_BUILTIN_FLAT_LOAD_INT32: + { + if (ignore) + return target; + /*rtx exec = */ + force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, DImode, + EXPAND_NORMAL)); + /*rtx ptr = */ + force_reg (V64DImode, + expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, V64DImode, + EXPAND_NORMAL)); + /*emit_insn (gen_vector_flat_loadv64si + (target, gcn_gen_undef (V64SImode), ptr, exec)); */ + return target; + } + case GCN_BUILTIN_FLAT_LOAD_PTR_INT32: + case GCN_BUILTIN_FLAT_LOAD_PTR_FLOAT: + { + if (ignore) + return target; + rtx exec = force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, + DImode, + EXPAND_NORMAL)); + rtx ptr = force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, + V64DImode, + EXPAND_NORMAL)); + rtx offsets = force_reg (V64SImode, + expand_expr (CALL_EXPR_ARG (exp, 2), + NULL_RTX, V64DImode, + EXPAND_NORMAL)); + rtx addrs = gen_reg_rtx (V64DImode); + rtx tmp = gen_reg_rtx (V64SImode); + emit_insn (gen_ashlv64si3_exec (tmp, offsets, + GEN_INT (2), + gcn_gen_undef (V64SImode), exec)); + emit_insn (gen_addv64di3_zext_dup2_exec (addrs, tmp, ptr, + gcn_gen_undef (V64DImode), + exec)); + rtx mem = gen_rtx_MEM (GET_MODE (target), addrs); + /*set_mem_addr_space (mem, ADDR_SPACE_FLAT); */ + /* FIXME: set attributes. */ + emit_insn (gen_mov_with_exec (target, mem, exec)); + return target; + } + case GCN_BUILTIN_FLAT_STORE_PTR_INT32: + case GCN_BUILTIN_FLAT_STORE_PTR_FLOAT: + { + rtx exec = force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, + DImode, + EXPAND_NORMAL)); + rtx ptr = force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, + V64DImode, + EXPAND_NORMAL)); + rtx offsets = force_reg (V64SImode, + expand_expr (CALL_EXPR_ARG (exp, 2), + NULL_RTX, V64DImode, + EXPAND_NORMAL)); + machine_mode vmode = TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, + 3))); + rtx val = force_reg (vmode, + expand_expr (CALL_EXPR_ARG (exp, 3), NULL_RTX, + vmode, + EXPAND_NORMAL)); + rtx addrs = gen_reg_rtx (V64DImode); + rtx tmp = gen_reg_rtx (V64SImode); + emit_insn (gen_ashlv64si3_exec (tmp, offsets, + GEN_INT (2), + gcn_gen_undef (V64SImode), exec)); + emit_insn (gen_addv64di3_zext_dup2_exec (addrs, tmp, ptr, + gcn_gen_undef (V64DImode), + exec)); + rtx mem = gen_rtx_MEM (vmode, addrs); + /*set_mem_addr_space (mem, ADDR_SPACE_FLAT); */ + /* FIXME: set attributes. */ + emit_insn (gen_mov_with_exec (mem, val, exec)); + return target; + } + case GCN_BUILTIN_SQRTVF: + { + if (ignore) + return target; + rtx exec = gcn_full_exec_reg (); + rtx arg = force_reg (V64SFmode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, + V64SFmode, + EXPAND_NORMAL)); + emit_insn (gen_sqrtv64sf2_exec + (target, arg, gcn_gen_undef (V64SFmode), exec)); + return target; + } + case GCN_BUILTIN_SQRTF: + { + if (ignore) + return target; + rtx arg = force_reg (SFmode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, + SFmode, + EXPAND_NORMAL)); + emit_insn (gen_sqrtsf2 (target, arg)); + return target; + } + case GCN_BUILTIN_OMP_DIM_SIZE: + { + if (ignore) + return target; + emit_insn (gen_oacc_dim_size (target, + expand_expr (CALL_EXPR_ARG (exp, 0), + NULL_RTX, SImode, + EXPAND_NORMAL))); + return target; + } + case GCN_BUILTIN_OMP_DIM_POS: + { + if (ignore) + return target; + emit_insn (gen_oacc_dim_pos (target, + expand_expr (CALL_EXPR_ARG (exp, 0), + NULL_RTX, SImode, + EXPAND_NORMAL))); + return target; + } + case GCN_BUILTIN_CMP_SWAP: + case GCN_BUILTIN_CMP_SWAPLL: + return gcn_expand_cmp_swap (exp, target); + + case GCN_BUILTIN_ACC_SINGLE_START: + { + if (ignore) + return target; + + rtx wavefront = gcn_oacc_dim_pos (1); + rtx cond = gen_rtx_EQ (VOIDmode, wavefront, const0_rtx); + rtx cc = (target && REG_P (target)) ? target : gen_reg_rtx (BImode); + emit_insn (gen_cstoresi4 (cc, cond, wavefront, const0_rtx)); + return cc; + } + + case GCN_BUILTIN_ACC_SINGLE_COPY_START: + { + rtx blk = force_reg (SImode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, + SImode, EXPAND_NORMAL)); + rtx wavefront = gcn_oacc_dim_pos (1); + rtx cond = gen_rtx_NE (VOIDmode, wavefront, const0_rtx); + rtx not_zero = gen_label_rtx (); + emit_insn (gen_cbranchsi4 (cond, wavefront, const0_rtx, not_zero)); + emit_move_insn (blk, const0_rtx); + emit_label (not_zero); + return blk; + } + + case GCN_BUILTIN_ACC_SINGLE_COPY_END: + return target; + + case GCN_BUILTIN_ACC_BARRIER: + emit_insn (gen_gcn_wavefront_barrier ()); + return target; + + default: + gcc_unreachable (); + } +} + +/* Expansion of simple arithmetic and bit binary operation builtins. + + Intended for use with gcn_builtins table. */ + +static rtx +gcn_expand_builtin_binop (tree exp, rtx target, rtx /*subtarget */ , + machine_mode /*mode */ , int ignore, + struct gcn_builtin_description *d) +{ + int icode = d->icode; + if (ignore) + return target; + + rtx exec = force_reg (DImode, + expand_expr (CALL_EXPR_ARG (exp, 0), NULL_RTX, DImode, + EXPAND_NORMAL)); + + machine_mode m1 = insn_data[icode].operand[1].mode; + rtx arg1 = expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, m1, + EXPAND_NORMAL); + if (!insn_data[icode].operand[1].predicate (arg1, m1)) + arg1 = force_reg (m1, arg1); + + machine_mode m2 = insn_data[icode].operand[2].mode; + rtx arg2 = expand_expr (CALL_EXPR_ARG (exp, 2), NULL_RTX, m2, + EXPAND_NORMAL); + if (!insn_data[icode].operand[2].predicate (arg2, m2)) + arg2 = force_reg (m2, arg2); + + rtx arg_prev; + if (call_expr_nargs (exp) == 4) + { + machine_mode m_prev = insn_data[icode].operand[4].mode; + arg_prev = force_reg (m_prev, + expand_expr (CALL_EXPR_ARG (exp, 3), NULL_RTX, + m_prev, EXPAND_NORMAL)); + } + else + arg_prev = gcn_gen_undef (GET_MODE (target)); + + rtx pat = GEN_FCN (icode) (target, arg1, arg2, exec, arg_prev); + emit_insn (pat); + return target; +} + +/* Implement TARGET_EXPAND_BUILTIN. + + Expand an expression EXP that calls a built-in function, with result going + to TARGET if that's convenient (and in mode MODE if that's convenient). + SUBTARGET may be used as the target for computing one of EXP's operands. + IGNORE is nonzero if the value is to be ignored. */ + +rtx +gcn_expand_builtin (tree exp, rtx target, rtx subtarget, machine_mode mode, + int ignore) +{ + tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); + unsigned int fcode = DECL_FUNCTION_CODE (fndecl); + struct gcn_builtin_description *d; + + gcc_assert (fcode < GCN_BUILTIN_MAX); + d = &gcn_builtins[fcode]; + + if (d->type == B_UNIMPLEMENTED) + sorry ("Builtin not implemented"); + + return d->expander (exp, target, subtarget, mode, ignore, d); +} + +/* }}} */ +/* {{{ Vectorization. */ + +/* Implement TARGET_VECTORIZE_GET_MASK_MODE. + + A vector mask is a value that holds one boolean result for every element in + a vector. */ + +opt_machine_mode +gcn_vectorize_get_mask_mode (poly_uint64 ARG_UNUSED (nunits), + poly_uint64 ARG_UNUSED (length)) +{ + /* GCN uses a DImode bit-mask. */ + return DImode; +} + +/* Return an RTX that references a vector with the i-th lane containing + PERM[i]*4. + + Helper function for gcn_vectorize_vec_perm_const. */ + +static rtx +gcn_make_vec_perm_address (unsigned int *perm) +{ + rtx x = gen_reg_rtx (V64SImode); + emit_move_insn (x, gcn_vec_constant (V64SImode, 0)); + + /* Permutation addresses use byte addressing. With each vector lane being + 4 bytes wide, and with 64 lanes in total, only bits 2..7 are significant, + so only set those. + + The permutation given to the vec_perm* patterns range from 0 to 2N-1 to + select between lanes in two vectors, but as the DS_BPERMUTE* instructions + only take one source vector, the most-significant bit can be ignored + here. Instead, we can use EXEC masking to select the relevant part of + each source vector after they are permuted separately. */ + uint64_t bit_mask = 1 << 2; + for (int i = 2; i < 8; i++, bit_mask <<= 1) + { + uint64_t exec_mask = 0; + uint64_t lane_mask = 1; + for (int j = 0; j < 64; j++, lane_mask <<= 1) + if ((perm[j] * 4) & bit_mask) + exec_mask |= lane_mask; + + if (exec_mask) + emit_insn (gen_addv64si3_exec (x, x, + gcn_vec_constant (V64SImode, + bit_mask), + x, get_exec (exec_mask))); + } + + return x; +} + +/* Implement TARGET_VECTORIZE_VEC_PERM_CONST. + + Return true if permutation with SEL is possible. + + If DST/SRC0/SRC1 are non-null, emit the instructions to perform the + permutations. */ + +static bool +gcn_vectorize_vec_perm_const (machine_mode vmode, rtx dst, + rtx src0, rtx src1, + const vec_perm_indices & sel) +{ + unsigned int nelt = GET_MODE_NUNITS (vmode); + + gcc_assert (VECTOR_MODE_P (vmode)); + gcc_assert (nelt <= 64); + gcc_assert (sel.length () == nelt); + + if (!dst) + { + /* All vector permutations are possible on this architecture, + with varying degrees of efficiency depending on the permutation. */ + return true; + } + + unsigned int perm[64]; + for (unsigned int i = 0; i < nelt; ++i) + perm[i] = sel[i] & (2 * nelt - 1); + + /* Make life a bit easier by swapping operands if necessary so that + the first element always comes from src0. */ + if (perm[0] >= nelt) + { + rtx temp = src0; + src0 = src1; + src1 = temp; + + for (unsigned int i = 0; i < nelt; ++i) + if (perm[i] < nelt) + perm[i] += nelt; + else + perm[i] -= nelt; + } + + /* TODO: There are more efficient ways to implement certain permutations + using ds_swizzle_b32 and/or DPP. Test for and expand them here, before + this more inefficient generic approach is used. */ + + int64_t src1_lanes = 0; + int64_t lane_bit = 1; + + for (unsigned int i = 0; i < nelt; ++i, lane_bit <<= 1) + { + /* Set the bits for lanes from src1. */ + if (perm[i] >= nelt) + src1_lanes |= lane_bit; + } + + rtx addr = gcn_make_vec_perm_address (perm); + rtx (*ds_bpermute) (rtx, rtx, rtx, rtx); + + switch (vmode) + { + case E_V64QImode: + ds_bpermute = gen_ds_bpermutev64qi; + break; + case E_V64HImode: + ds_bpermute = gen_ds_bpermutev64hi; + break; + case E_V64SImode: + ds_bpermute = gen_ds_bpermutev64si; + break; + case E_V64HFmode: + ds_bpermute = gen_ds_bpermutev64hf; + break; + case E_V64SFmode: + ds_bpermute = gen_ds_bpermutev64sf; + break; + case E_V64DImode: + ds_bpermute = gen_ds_bpermutev64di; + break; + case E_V64DFmode: + ds_bpermute = gen_ds_bpermutev64df; + break; + default: + gcc_assert (false); + } + + /* Load elements from src0 to dst. */ + gcc_assert (~src1_lanes); + emit_insn (ds_bpermute (dst, addr, src0, gcn_full_exec_reg ())); + + /* Load elements from src1 to dst. */ + if (src1_lanes) + { + /* Masking a lane masks both the destination and source lanes for + DS_BPERMUTE, so we need to have all lanes enabled for the permute, + then add an extra masked move to merge the results of permuting + the two source vectors together. + */ + rtx tmp = gen_reg_rtx (vmode); + emit_insn (ds_bpermute (tmp, addr, src1, gcn_full_exec_reg ())); + emit_insn (gen_mov_with_exec (dst, tmp, get_exec (src1_lanes))); + } + + return true; +} + +/* Implements TARGET_VECTOR_MODE_SUPPORTED_P. + + Return nonzero if vector MODE is supported with at least move + instructions. */ + +static bool +gcn_vector_mode_supported_p (machine_mode mode) +{ + /* FIXME: Enable V64QImode and V64HImode. + We should support these modes, but vector operations are usually + assumed to automatically truncate types, and GCN does not. We + need to add explicit truncates and/or use SDWA for QI/HI insns. */ + return (/* mode == V64QImode || mode == V64HImode + ||*/ mode == V64SImode || mode == V64DImode + || mode == V64SFmode || mode == V64DFmode); +} + +/* Implement TARGET_VECTORIZE_PREFERRED_SIMD_MODE. + + Enables autovectorization for all supported modes. */ + +static machine_mode +gcn_vectorize_preferred_simd_mode (scalar_mode mode) +{ + switch (mode) + { + case E_QImode: + return V64QImode; + case E_HImode: + return V64HImode; + case E_SImode: + return V64SImode; + case E_DImode: + return V64DImode; + case E_SFmode: + return V64SFmode; + case E_DFmode: + return V64DFmode; + default: + return word_mode; + } +} + +/* Implement TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT. + + Returns the preferred alignment in bits for accesses to vectors of type type + in vectorized code. This might be less than or greater than the ABI-defined + value returned by TARGET_VECTOR_ALIGNMENT. It can be equal to the alignment + of a single element, in which case the vectorizer will not try to optimize + for alignment. */ + +static poly_uint64 +gcn_preferred_vector_alignment (const_tree type) +{ + return TYPE_ALIGN (TREE_TYPE (type)); +} + +/* Implement TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT. + + Return true if the target supports misaligned vector store/load of a + specific factor denoted in the misalignment parameter. */ + +static bool +gcn_vectorize_support_vector_misalignment (machine_mode ARG_UNUSED (mode), + const_tree type, int misalignment, + bool is_packed) +{ + if (is_packed) + return false; + + /* If the misalignment is unknown, we should be able to handle the access + so long as it is not to a member of a packed data structure. */ + if (misalignment == -1) + return true; + + /* Return true if the misalignment is a multiple of the natural alignment + of the vector's element type. This is probably always going to be + true in practice, since we've already established that this isn't a + packed access. */ + return misalignment % TYPE_ALIGN_UNIT (type) == 0; +} + +/* Implement TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. + + Return true if vector alignment is reachable (by peeling N iterations) for + the given scalar type TYPE. */ + +static bool +gcn_vector_alignment_reachable (const_tree ARG_UNUSED (type), bool is_packed) +{ + /* Vectors which aren't in packed structures will not be less aligned than + the natural alignment of their element type, so this is safe. */ + return !is_packed; +} + +/* Generate DPP instructions used for vector reductions. + + The opcode is given by INSN. + The first operand of the operation is shifted right by SHIFT vector lanes. + SHIFT must be a power of 2. If SHIFT is 16, the 15th lane of each row is + broadcast the next row (thereby acting like a shift of 16 for the end of + each row). If SHIFT is 32, lane 31 is broadcast to all the + following lanes (thereby acting like a shift of 32 for lane 63). */ + +char * +gcn_expand_dpp_shr_insn (machine_mode mode, const char *insn, + int unspec, int shift) +{ + static char buf[64]; + const char *dpp; + const char *vcc_in = ""; + const char *vcc_out = ""; + + /* Add the vcc operand if needed. */ + if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT) + { + if (unspec == UNSPEC_PLUS_CARRY_IN_DPP_SHR) + vcc_in = ", vcc"; + + if (unspec == UNSPEC_PLUS_CARRY_DPP_SHR + || unspec == UNSPEC_PLUS_CARRY_IN_DPP_SHR) + vcc_out = ", vcc"; + } + + /* Add the DPP modifiers. */ + switch (shift) + { + case 1: + dpp = "row_shr:1 bound_ctrl:0"; + break; + case 2: + dpp = "row_shr:2 bound_ctrl:0"; + break; + case 4: + dpp = "row_shr:4 bank_mask:0xe"; + break; + case 8: + dpp = "row_shr:8 bank_mask:0xc"; + break; + case 16: + dpp = "row_bcast:15 row_mask:0xa"; + break; + case 32: + dpp = "row_bcast:31 row_mask:0xc"; + break; + default: + gcc_unreachable (); + } + + sprintf (buf, "%s\t%%0%s, %%1, %%2%s %s", insn, vcc_out, vcc_in, dpp); + + return buf; +} + +/* Generate vector reductions in terms of DPP instructions. + + The vector register SRC of mode MODE is reduced using the operation given + by UNSPEC, and the scalar result is returned in lane 63 of a vector + register. */ + +rtx +gcn_expand_reduc_scalar (machine_mode mode, rtx src, int unspec) +{ + rtx tmp = gen_reg_rtx (mode); + bool use_plus_carry = unspec == UNSPEC_PLUS_DPP_SHR + && GET_MODE_CLASS (mode) == MODE_VECTOR_INT + && (TARGET_GCN3 || mode == V64DImode); + + if (use_plus_carry) + unspec = UNSPEC_PLUS_CARRY_DPP_SHR; + + /* Perform reduction by first performing the reduction operation on every + pair of lanes, then on every pair of results from the previous + iteration (thereby effectively reducing every 4 lanes) and so on until + all lanes are reduced. */ + for (int i = 0, shift = 1; i < 6; i++, shift <<= 1) + { + rtx shift_val = gen_rtx_CONST_INT (VOIDmode, shift); + rtx insn = gen_rtx_SET (tmp, + gen_rtx_UNSPEC (mode, + gen_rtvec (3, + src, src, shift_val), + unspec)); + + /* Add clobber for instructions that set the carry flags. */ + if (use_plus_carry) + { + rtx clobber = gen_rtx_CLOBBER (VOIDmode, + gen_rtx_REG (DImode, VCC_REG)); + insn = gen_rtx_PARALLEL (VOIDmode, + gen_rtvec (2, insn, clobber)); + } + + emit_insn (insn); + + /* The source operands for every iteration after the first + should be TMP. */ + src = tmp; + } + + return tmp; +} + +/* Implement TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST. */ + +int +gcn_vectorization_cost (enum vect_cost_for_stmt ARG_UNUSED (type_of_cost), + tree ARG_UNUSED (vectype), int ARG_UNUSED (misalign)) +{ + /* Always vectorize. */ + return 1; +} + +/* }}} */ +/* {{{ md_reorg pass. */ + +/* Identify VMEM instructions from their "type" attribute. */ + +static bool +gcn_vmem_insn_p (attr_type type) +{ + switch (type) + { + case TYPE_MUBUF: + case TYPE_MTBUF: + case TYPE_FLAT: + return true; + case TYPE_UNKNOWN: + case TYPE_SOP1: + case TYPE_SOP2: + case TYPE_SOPK: + case TYPE_SOPC: + case TYPE_SOPP: + case TYPE_SMEM: + case TYPE_DS: + case TYPE_VOP2: + case TYPE_VOP1: + case TYPE_VOPC: + case TYPE_VOP3A: + case TYPE_VOP3B: + case TYPE_VOP_SDWA: + case TYPE_VOP_DPP: + case TYPE_MULT: + case TYPE_VMULT: + return false; + } + gcc_unreachable (); + return false; +} + +/* If INSN sets the EXEC register to a constant value, return the value, + otherwise return zero. */ + +static int64_t +gcn_insn_exec_value (rtx_insn *insn) +{ + if (!NONDEBUG_INSN_P (insn)) + return 0; + + rtx pattern = PATTERN (insn); + + if (GET_CODE (pattern) == SET) + { + rtx dest = XEXP (pattern, 0); + rtx src = XEXP (pattern, 1); + + if (GET_MODE (dest) == DImode + && REG_P (dest) && REGNO (dest) == EXEC_REG + && CONST_INT_P (src)) + return INTVAL (src); + } + + return 0; +} + +/* Sets the EXEC register before INSN to the value that it had after + LAST_EXEC_DEF. The constant value of the EXEC register is returned if + known, otherwise it returns zero. */ + +static int64_t +gcn_restore_exec (rtx_insn *insn, rtx_insn *last_exec_def, int64_t curr_exec, + bool curr_exec_known, bool &last_exec_def_saved) +{ + rtx exec_reg = gen_rtx_REG (DImode, EXEC_REG); + rtx exec; + + int64_t exec_value = gcn_insn_exec_value (last_exec_def); + + if (exec_value) + { + /* If the EXEC value is a constant and it happens to be the same as the + current EXEC value, the restore can be skipped. */ + if (curr_exec_known && exec_value == curr_exec) + return exec_value; + + exec = GEN_INT (exec_value); + } + else + { + /* If the EXEC value is not a constant, save it in a register after the + point of definition. */ + rtx exec_save_reg = gen_rtx_REG (DImode, EXEC_SAVE_REG); + + if (!last_exec_def_saved) + { + start_sequence (); + emit_move_insn (exec_save_reg, exec_reg); + rtx_insn *seq = get_insns (); + end_sequence (); + + emit_insn_after (seq, last_exec_def); + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Saving EXEC after insn %d.\n", + INSN_UID (last_exec_def)); + + last_exec_def_saved = true; + } + + exec = exec_save_reg; + } + + /* Restore EXEC register before the usage. */ + start_sequence (); + emit_move_insn (exec_reg, exec); + rtx_insn *seq = get_insns (); + end_sequence (); + emit_insn_before (seq, insn); + + if (dump_file && (dump_flags & TDF_DETAILS)) + { + if (exec_value) + fprintf (dump_file, "Restoring EXEC to %ld before insn %d.\n", + exec_value, INSN_UID (insn)); + else + fprintf (dump_file, + "Restoring EXEC from saved value before insn %d.\n", + INSN_UID (insn)); + } + + return exec_value; +} + +/* Implement TARGET_MACHINE_DEPENDENT_REORG. + + Ensure that pipeline dependencies and lane masking are set correctly. */ + +static void +gcn_md_reorg (void) +{ + basic_block bb; + rtx exec_reg = gen_rtx_REG (DImode, EXEC_REG); + rtx exec_lo_reg = gen_rtx_REG (SImode, EXEC_LO_REG); + rtx exec_hi_reg = gen_rtx_REG (SImode, EXEC_HI_REG); + regset_head live; + + INIT_REG_SET (&live); + + compute_bb_for_insn (); + + if (!optimize) + { + split_all_insns (); + if (dump_file && (dump_flags & TDF_DETAILS)) + { + fprintf (dump_file, "After split:\n"); + print_rtl_with_bb (dump_file, get_insns (), dump_flags); + } + + /* Update data-flow information for split instructions. */ + df_insn_rescan_all (); + } + + df_analyze (); + + /* This pass ensures that the EXEC register is set correctly, according + to the "exec" attribute. However, care must be taken so that the + value that reaches explicit uses of the EXEC register remains the + same as before. + */ + + FOR_EACH_BB_FN (bb, cfun) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "BB %d:\n", bb->index); + + rtx_insn *insn, *curr; + rtx_insn *last_exec_def = BB_HEAD (bb); + bool last_exec_def_saved = false; + bool curr_exec_explicit = true; + bool curr_exec_known = true; + int64_t curr_exec = 0; /* 0 here means 'the value is that of EXEC + after last_exec_def is executed'. */ + + FOR_BB_INSNS_SAFE (bb, insn, curr) + { + if (!NONDEBUG_INSN_P (insn)) + continue; + + if (GET_CODE (PATTERN (insn)) == USE + || GET_CODE (PATTERN (insn)) == CLOBBER) + continue; + + HARD_REG_SET defs, uses; + CLEAR_HARD_REG_SET (defs); + CLEAR_HARD_REG_SET (uses); + note_stores (PATTERN (insn), record_hard_reg_sets, &defs); + note_uses (&PATTERN (insn), record_hard_reg_uses, &uses); + + bool exec_lo_def_p = TEST_HARD_REG_BIT (defs, EXEC_LO_REG); + bool exec_hi_def_p = TEST_HARD_REG_BIT (defs, EXEC_HI_REG); + bool exec_used = (hard_reg_set_intersect_p + (uses, reg_class_contents[(int) EXEC_MASK_REG]) + || TEST_HARD_REG_BIT (uses, EXECZ_REG)); + + /* Check the instruction for implicit setting of EXEC via an + attribute. */ + attr_exec exec_attr = get_attr_exec (insn); + int64_t new_exec; + + switch (exec_attr) + { + case EXEC_NONE: + new_exec = 0; + break; + + case EXEC_SINGLE: + /* Instructions that do not involve memory accesses only require + bit 0 of EXEC to be set. */ + if (gcn_vmem_insn_p (get_attr_type (insn)) + || get_attr_type (insn) == TYPE_DS) + new_exec = 1; + else + new_exec = curr_exec | 1; + break; + + case EXEC_FULL: + new_exec = -1; + break; + + default: /* Auto-detect what setting is appropriate. */ + { + new_exec = 0; + + /* If EXEC is referenced explicitly then we don't need to do + anything to set it, so we're done. */ + if (exec_used) + break; + + /* Scan the insn for VGPRs defs or uses. The mode determines + what kind of exec is needed. */ + subrtx_iterator::array_type array; + FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) + { + const_rtx x = *iter; + if (REG_P (x) && VGPR_REGNO_P (REGNO (x))) + { + if (VECTOR_MODE_P (GET_MODE (x))) + { + new_exec = -1; + break; + } + else + new_exec = 1; + } + } + } + break; + } + + if (new_exec && (!curr_exec_known || new_exec != curr_exec)) + { + start_sequence (); + emit_move_insn (exec_reg, GEN_INT (new_exec)); + rtx_insn *seq = get_insns (); + end_sequence (); + emit_insn_before (seq, insn); + + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Setting EXEC to %ld before insn %d.\n", + new_exec, INSN_UID (insn)); + + curr_exec = new_exec; + curr_exec_explicit = false; + curr_exec_known = true; + } + else if (new_exec && dump_file && (dump_flags & TDF_DETAILS)) + { + fprintf (dump_file, "Exec already is %ld before insn %d.\n", + new_exec, INSN_UID (insn)); + } + + /* The state of the EXEC register is unknown after a + function call. */ + if (CALL_P (insn)) + curr_exec_known = false; + + /* Handle explicit uses of EXEC. If the instruction is a partial + explicit definition of EXEC, then treat it as an explicit use of + EXEC as well. */ + if (exec_used || exec_lo_def_p != exec_hi_def_p) + { + /* An instruction that explicitly uses EXEC should not also + implicitly define it. */ + gcc_assert (!exec_used || !new_exec); + + if (!curr_exec_known || !curr_exec_explicit) + { + /* Restore the previous explicitly defined value. */ + curr_exec = gcn_restore_exec (insn, last_exec_def, + curr_exec, curr_exec_known, + last_exec_def_saved); + curr_exec_explicit = true; + curr_exec_known = true; + } + } + + /* Handle explicit definitions of EXEC. */ + if (exec_lo_def_p || exec_hi_def_p) + { + last_exec_def = insn; + last_exec_def_saved = false; + curr_exec = gcn_insn_exec_value (insn); + curr_exec_explicit = true; + curr_exec_known = true; + + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Found %s definition of EXEC at insn %d.\n", + exec_lo_def_p == exec_hi_def_p ? "full" : "partial", + INSN_UID (insn)); + } + } + + COPY_REG_SET (&live, DF_LR_OUT (bb)); + df_simulate_initialize_backwards (bb, &live); + + /* If EXEC is live after the basic block, restore the value of EXEC + at the end of the block. */ + if ((REGNO_REG_SET_P (&live, EXEC_LO_REG) + || REGNO_REG_SET_P (&live, EXEC_HI_REG)) + && (!curr_exec_known || !curr_exec_explicit)) + { + rtx_insn *end_insn = BB_END (bb); + + /* If the instruction is not a jump instruction, do the restore + after the last instruction in the basic block. */ + if (NONJUMP_INSN_P (end_insn)) + end_insn = NEXT_INSN (end_insn); + + gcn_restore_exec (end_insn, last_exec_def, curr_exec, + curr_exec_known, last_exec_def_saved); + } + } + + CLEAR_REG_SET (&live); + + /* "Manually Inserted Wait States (NOPs)." + + GCN hardware detects most kinds of register dependencies, but there + are some exceptions documented in the ISA manual. This pass + detects the missed cases, and inserts the documented number of NOPs + required for correct execution. */ + + const int max_waits = 5; + struct ilist + { + rtx_insn *insn; + attr_unit unit; + HARD_REG_SET writes; + int age; + } back[max_waits]; + int oldest = 0; + for (int i = 0; i < max_waits; i++) + back[i].insn = NULL; + + rtx_insn *insn, *last_insn = NULL; + for (insn = get_insns (); insn != 0; insn = NEXT_INSN (insn)) + { + if (!NONDEBUG_INSN_P (insn)) + continue; + + if (GET_CODE (PATTERN (insn)) == USE + || GET_CODE (PATTERN (insn)) == CLOBBER) + continue; + + attr_type itype = get_attr_type (insn); + attr_unit iunit = get_attr_unit (insn); + HARD_REG_SET ireads, iwrites; + CLEAR_HARD_REG_SET (ireads); + CLEAR_HARD_REG_SET (iwrites); + note_stores (PATTERN (insn), record_hard_reg_sets, &iwrites); + note_uses (&PATTERN (insn), record_hard_reg_uses, &ireads); + + /* Scan recent previous instructions for dependencies not handled in + hardware. */ + int nops_rqd = 0; + for (int i = oldest; i < oldest + max_waits; i++) + { + struct ilist *prev_insn = &back[i % max_waits]; + + if (!prev_insn->insn) + continue; + + /* VALU writes SGPR followed by VMEM reading the same SGPR + requires 5 wait states. */ + if ((prev_insn->age + nops_rqd) < 5 + && prev_insn->unit == UNIT_VECTOR + && gcn_vmem_insn_p (itype)) + { + HARD_REG_SET regs; + COPY_HARD_REG_SET (regs, prev_insn->writes); + AND_HARD_REG_SET (regs, ireads); + if (hard_reg_set_intersect_p + (regs, reg_class_contents[(int) SGPR_REGS])) + nops_rqd = 5 - prev_insn->age; + } + + /* VALU sets VCC/EXEC followed by VALU uses VCCZ/EXECZ + requires 5 wait states. */ + if ((prev_insn->age + nops_rqd) < 5 + && prev_insn->unit == UNIT_VECTOR + && iunit == UNIT_VECTOR + && ((hard_reg_set_intersect_p + (prev_insn->writes, + reg_class_contents[(int) EXEC_MASK_REG]) + && TEST_HARD_REG_BIT (ireads, EXECZ_REG)) + || + (hard_reg_set_intersect_p + (prev_insn->writes, + reg_class_contents[(int) VCC_CONDITIONAL_REG]) + && TEST_HARD_REG_BIT (ireads, VCCZ_REG)))) + nops_rqd = 5 - prev_insn->age; + + /* VALU writes SGPR/VCC followed by v_{read,write}lane using + SGPR/VCC as lane select requires 4 wait states. */ + if ((prev_insn->age + nops_rqd) < 4 + && prev_insn->unit == UNIT_VECTOR + && get_attr_laneselect (insn) == LANESELECT_YES) + { + HARD_REG_SET regs; + COPY_HARD_REG_SET (regs, prev_insn->writes); + AND_HARD_REG_SET (regs, ireads); + if (hard_reg_set_intersect_p + (regs, reg_class_contents[(int) SGPR_REGS]) + || hard_reg_set_intersect_p + (regs, reg_class_contents[(int) VCC_CONDITIONAL_REG])) + nops_rqd = 4 - prev_insn->age; + } + + /* VALU writes VGPR followed by VALU_DPP reading that VGPR + requires 2 wait states. */ + if ((prev_insn->age + nops_rqd) < 2 + && prev_insn->unit == UNIT_VECTOR + && itype == TYPE_VOP_DPP) + { + HARD_REG_SET regs; + COPY_HARD_REG_SET (regs, prev_insn->writes); + AND_HARD_REG_SET (regs, ireads); + if (hard_reg_set_intersect_p + (regs, reg_class_contents[(int) VGPR_REGS])) + nops_rqd = 2 - prev_insn->age; + } + } + + /* Insert the required number of NOPs. */ + for (int i = nops_rqd; i > 0; i--) + emit_insn_after (gen_nop (), last_insn); + + /* Age the previous instructions. We can also ignore writes to + registers subsequently overwritten. */ + HARD_REG_SET written; + CLEAR_HARD_REG_SET (written); + for (int i = oldest + max_waits - 1; i > oldest; i--) + { + struct ilist *prev_insn = &back[i % max_waits]; + + /* Assume all instructions are equivalent to one "wait", the same + as s_nop. This is probably true for SALU, but not VALU (which + may take longer), so this is not optimal. However, AMD do + not publish the cycle times for instructions. */ + prev_insn->age += 1 + nops_rqd; + + IOR_HARD_REG_SET (written, iwrites); + AND_COMPL_HARD_REG_SET (prev_insn->writes, written); + } + + /* Track the current instruction as a previous instruction. */ + back[oldest].insn = insn; + back[oldest].unit = iunit; + COPY_HARD_REG_SET (back[oldest].writes, iwrites); + back[oldest].age = 0; + oldest = (oldest + 1) % max_waits; + + last_insn = insn; + } +} + +/* }}} */ +/* {{{ OpenACC / OpenMP. */ + +#define GCN_DEFAULT_GANGS 0 /* Choose at runtime. */ +#define GCN_DEFAULT_WORKERS 0 /* Choose at runtime. */ +#define GCN_DEFAULT_VECTORS 1 /* Use autovectorization only, for now. */ + +/* Implement TARGET_GOACC_VALIDATE_DIMS. + + Check the launch dimensions provided for an OpenACC compute + region, or routine. */ + +static bool +gcn_goacc_validate_dims (tree decl, int dims[], int fn_level, + unsigned /*used*/) +{ + bool changed = false; + + /* FIXME: remove -facc-experimental-workers when they're ready. */ + int max_workers = flag_worker_partitioning ? 16 : 1; + + /* The vector size must appear to be 64, to the user, unless this is a + SEQ routine. The real, internal value is always 1, which means use + autovectorization, but the user should not see that. */ + if (fn_level <= GOMP_DIM_VECTOR && fn_level >= -1 + && dims[GOMP_DIM_VECTOR] >= 0) + { + if (fn_level < 0 && dims[GOMP_DIM_VECTOR] >= 0 + && dims[GOMP_DIM_VECTOR] != 64) + warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, + OPT_Wopenacc_dims, + (dims[GOMP_DIM_VECTOR] + ? G_("using vector_length (64), ignoring %d") + : G_("using vector_length (64), " + "ignoring runtime setting")), + dims[GOMP_DIM_VECTOR]); + dims[GOMP_DIM_VECTOR] = 1; + changed = true; + } + + /* Check the num workers is not too large. */ + if (dims[GOMP_DIM_WORKER] > max_workers) + { + warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, + OPT_Wopenacc_dims, + "using num_workers (%d), ignoring %d", + max_workers, dims[GOMP_DIM_WORKER]); + dims[GOMP_DIM_WORKER] = max_workers; + changed = true; + } + + /* Set global defaults. */ + if (!decl) + { + dims[GOMP_DIM_VECTOR] = GCN_DEFAULT_VECTORS; + if (dims[GOMP_DIM_WORKER] < 0) + dims[GOMP_DIM_WORKER] = (flag_worker_partitioning + ? GCN_DEFAULT_WORKERS : 1); + if (dims[GOMP_DIM_GANG] < 0) + dims[GOMP_DIM_GANG] = GCN_DEFAULT_GANGS; + changed = true; + } + + return changed; +} + +/* Helper function for oacc_dim_size instruction. + Also used for OpenMP, via builtin_gcn_dim_size, and the omp_gcn pass. */ + +rtx +gcn_oacc_dim_size (int dim) +{ + if (dim < 0 || dim > 2) + error ("offload dimension out of range (%d)", dim); + + /* Vectors are a special case. */ + if (dim == 2) + return const1_rtx; /* Think of this as 1 times 64. */ + + static int offset[] = { + /* Offsets into dispatch packet. */ + 12, /* X dim = Gang / Team / Work-group. */ + 20, /* Z dim = Worker / Thread / Wavefront. */ + 16 /* Y dim = Vector / SIMD / Work-item. */ + }; + rtx addr = gen_rtx_PLUS (DImode, + gen_rtx_REG (DImode, + cfun->machine->args. + reg[DISPATCH_PTR_ARG]), + GEN_INT (offset[dim])); + return gen_rtx_MEM (SImode, addr); +} + +/* Helper function for oacc_dim_pos instruction. + Also used for OpenMP, via builtin_gcn_dim_pos, and the omp_gcn pass. */ + +rtx +gcn_oacc_dim_pos (int dim) +{ + if (dim < 0 || dim > 2) + error ("offload dimension out of range (%d)", dim); + + static const int reg[] = { + WORKGROUP_ID_X_ARG, /* Gang / Team / Work-group. */ + WORK_ITEM_ID_Z_ARG, /* Worker / Thread / Wavefront. */ + WORK_ITEM_ID_Y_ARG /* Vector / SIMD / Work-item. */ + }; + + int reg_num = cfun->machine->args.reg[reg[dim]]; + + /* The information must have been requested by the kernel. */ + gcc_assert (reg_num >= 0); + + return gen_rtx_REG (SImode, reg_num); +} + +/* Implement TARGET_GOACC_FORK_JOIN. */ + +static bool +gcn_fork_join (gcall *ARG_UNUSED (call), const int *ARG_UNUSED (dims), + bool ARG_UNUSED (is_fork)) +{ + /* GCN does not use the fork/join concept invented for NVPTX. + Instead we use standard autovectorization. */ + return false; +} + +/* Implement ??????? + FIXME make this a real hook. + + Adjust FNDECL such that options inherited from the host compiler + are made appropriate for the accelerator compiler. */ + +void +gcn_fixup_accel_lto_options (tree fndecl) +{ + tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); + if (!func_optimize) + return; + + tree old_optimize = build_optimization_node (&global_options); + tree new_optimize; + + /* If the function changed the optimization levels as well as + setting target options, start with the optimizations + specified. */ + if (func_optimize != old_optimize) + cl_optimization_restore (&global_options, + TREE_OPTIMIZATION (func_optimize)); + + gcn_option_override (); + + /* The target attributes may also change some optimization flags, + so update the optimization options if necessary. */ + new_optimize = build_optimization_node (&global_options); + + if (old_optimize != new_optimize) + { + DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize; + cl_optimization_restore (&global_options, + TREE_OPTIMIZATION (old_optimize)); + } +} + +/* }}} */ +/* {{{ ASM Output. */ + +/* Implement TARGET_ASM_FILE_START. + + Print assembler file header text. */ + +static void +output_file_start (void) +{ + fprintf (asm_out_file, "\t.text\n"); + fprintf (asm_out_file, "\t.hsa_code_object_version 2,0\n"); + fprintf (asm_out_file, "\t.hsa_code_object_isa\n"); /* Autodetect. */ + fprintf (asm_out_file, "\t.section\t.AMDGPU.config\n"); + fprintf (asm_out_file, "\t.text\n"); +} + +/* Implement ASM_DECLARE_FUNCTION_NAME via gcn-hsa.h. + + Print the initial definition of a function name. + + For GCN kernel entry points this includes all the HSA meta-data, special + alignment constraints that don't apply to regular functions, and magic + comments that pass information to mkoffload. */ + +void +gcn_hsa_declare_function_name (FILE *file, const char *name, tree) +{ + int sgpr, vgpr; + bool xnack_enabled = false; + int extra_regs = 0; + + if (cfun && cfun->machine && cfun->machine->normal_function) + { + fputs ("\t.type\t", file); + assemble_name (file, name); + fputs (",@function\n", file); + assemble_name (file, name); + fputs (":\n", file); + return; + } + + /* Determine count of sgpr/vgpr registers by looking for last + one used. */ + for (sgpr = 101; sgpr >= 0; sgpr--) + if (df_regs_ever_live_p (FIRST_SGPR_REG + sgpr)) + break; + sgpr++; + for (vgpr = 255; vgpr >= 0; vgpr--) + if (df_regs_ever_live_p (FIRST_VGPR_REG + vgpr)) + break; + vgpr++; + + if (xnack_enabled) + extra_regs = 6; + if (df_regs_ever_live_p (FLAT_SCRATCH_LO_REG) + || df_regs_ever_live_p (FLAT_SCRATCH_HI_REG)) + extra_regs = 4; + else if (df_regs_ever_live_p (VCC_LO_REG) + || df_regs_ever_live_p (VCC_HI_REG)) + extra_regs = 2; + + if (!leaf_function_p ()) + { + /* We can't know how many registers function calls might use. */ + if (vgpr < 64) + vgpr = 64; + if (sgpr + extra_regs < 102) + sgpr = 102 - extra_regs; + } + + fputs ("\t.align\t256\n", file); + fputs ("\t.type\t", file); + assemble_name (file, name); + fputs (",@function\n\t.amdgpu_hsa_kernel\t", file); + assemble_name (file, name); + fputs ("\n", file); + assemble_name (file, name); + fputs (":\n", file); + fprintf (file, "\t.amd_kernel_code_t\n" + "\t\tkernel_code_version_major = 1\n" + "\t\tkernel_code_version_minor = 0\n" "\t\tmachine_kind = 1\n" + /* "\t\tmachine_version_major = 8\n" + "\t\tmachine_version_minor = 0\n" + "\t\tmachine_version_stepping = 1\n" */ + "\t\tkernel_code_entry_byte_offset = 256\n" + "\t\tkernel_code_prefetch_byte_size = 0\n" + "\t\tmax_scratch_backing_memory_byte_size = 0\n" + "\t\tcompute_pgm_rsrc1_vgprs = %i\n" + "\t\tcompute_pgm_rsrc1_sgprs = %i\n" + "\t\tcompute_pgm_rsrc1_priority = 0\n" + "\t\tcompute_pgm_rsrc1_float_mode = 192\n" + "\t\tcompute_pgm_rsrc1_priv = 0\n" + "\t\tcompute_pgm_rsrc1_dx10_clamp = 1\n" + "\t\tcompute_pgm_rsrc1_debug_mode = 0\n" + "\t\tcompute_pgm_rsrc1_ieee_mode = 1\n" + /* We enable scratch memory. */ + "\t\tcompute_pgm_rsrc2_scratch_en = 1\n" + "\t\tcompute_pgm_rsrc2_user_sgpr = %i\n" + "\t\tcompute_pgm_rsrc2_tgid_x_en = 1\n" + "\t\tcompute_pgm_rsrc2_tgid_y_en = 0\n" + "\t\tcompute_pgm_rsrc2_tgid_z_en = 0\n" + "\t\tcompute_pgm_rsrc2_tg_size_en = 0\n" + "\t\tcompute_pgm_rsrc2_tidig_comp_cnt = 0\n" + "\t\tcompute_pgm_rsrc2_excp_en_msb = 0\n" + "\t\tcompute_pgm_rsrc2_lds_size = 0\n" /* Set at runtime. */ + "\t\tcompute_pgm_rsrc2_excp_en = 0\n", + (vgpr - 1) / 4, + /* Must match wavefront_sgpr_count */ + (sgpr + extra_regs + 7) / 8 - 1, + /* The total number of SGPR user data registers requested. This + number must match the number of user data registers enabled. */ + cfun->machine->args.nsgprs); + int reg = FIRST_SGPR_REG; + for (int a = 0; a < GCN_KERNEL_ARG_TYPES; a++) + { + int reg_first = -1; + int reg_last; + if ((cfun->machine->args.requested & (1 << a)) + && (gcn_kernel_arg_types[a].fixed_regno < 0)) + { + reg_first = reg; + reg_last = (reg_first + + (GET_MODE_SIZE (gcn_kernel_arg_types[a].mode) + / UNITS_PER_WORD) - 1); + reg = reg_last + 1; + } + + if (gcn_kernel_arg_types[a].header_pseudo) + { + fprintf (file, "\t\t%s = %i", + gcn_kernel_arg_types[a].header_pseudo, + (cfun->machine->args.requested & (1 << a)) != 0); + if (reg_first != -1) + { + fprintf (file, " ; ("); + for (int i = reg_first; i <= reg_last; ++i) + { + if (i != reg_first) + fprintf (file, ", "); + fprintf (file, "%s", reg_names[i]); + } + fprintf (file, ")"); + } + fprintf (file, "\n"); + } + else if (gcn_kernel_arg_types[a].fixed_regno >= 0 + && cfun->machine->args.requested & (1 << a)) + fprintf (file, "\t\t; %s = %i (%s)\n", + gcn_kernel_arg_types[a].name, + (cfun->machine->args.requested & (1 << a)) != 0, + reg_names[gcn_kernel_arg_types[a].fixed_regno]); + } + fprintf (file, "\t\tenable_vgpr_workitem_id = %i\n", + (cfun->machine->args.requested & (1 << WORK_ITEM_ID_Z_ARG)) + ? 2 + : cfun->machine->args.requested & (1 << WORK_ITEM_ID_Y_ARG) + ? 1 : 0); + fprintf (file, "\t\tenable_ordered_append_gds = 0\n" + "\t\tprivate_element_size = 1\n" + "\t\tis_ptr64 = 1\n" + "\t\tis_dynamic_callstack = 0\n" + "\t\tis_debug_enabled = 0\n" + "\t\tis_xnack_enabled = %i\n" + "\t\tworkitem_private_segment_byte_size = %i\n" + "\t\tworkgroup_group_segment_byte_size = %u\n" + "\t\tgds_segment_byte_size = 0\n" + "\t\tkernarg_segment_byte_size = %i\n" + "\t\tworkgroup_fbarrier_count = 0\n" + "\t\twavefront_sgpr_count = %i\n" + "\t\tworkitem_vgpr_count = %i\n" + "\t\treserved_vgpr_first = 0\n" + "\t\treserved_vgpr_count = 0\n" + "\t\treserved_sgpr_first = 0\n" + "\t\treserved_sgpr_count = 0\n" + "\t\tdebug_wavefront_private_segment_offset_sgpr = 0\n" + "\t\tdebug_private_segment_buffer_sgpr = 0\n" + "\t\tkernarg_segment_alignment = %i\n" + "\t\tgroup_segment_alignment = 4\n" + "\t\tprivate_segment_alignment = %i\n" + "\t\twavefront_size = 6\n" + "\t\tcall_convention = 0\n" + "\t\truntime_loader_kernel_symbol = 0\n" + "\t.end_amd_kernel_code_t\n", xnack_enabled, + /* workitem_private_segment_bytes_size needs to be + one 64th the wave-front stack size. */ + stack_size_opt / 64, + LDS_SIZE, cfun->machine->kernarg_segment_byte_size, + /* Number of scalar registers used by a wavefront. This + includes the special SGPRs for VCC, Flat Scratch (Base, + Size) and XNACK (for GFX8 (VI)+). It does not include the + 16 SGPR added if a trap handler is enabled. Must match + compute_pgm_rsrc1.sgprs. */ + sgpr + extra_regs, vgpr, + cfun->machine->kernarg_segment_alignment, + crtl->stack_alignment_needed / 8); + + /* This comment is read by mkoffload. */ + if (flag_openacc) + fprintf (file, "\t;; OPENACC-DIMS: %d, %d, %d : %s\n", + oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_GANG), + oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_WORKER), + oacc_get_fn_dim_size (cfun->decl, GOMP_DIM_VECTOR), name); +} + +/* Implement TARGET_ASM_SELECT_SECTION. + + Return the section into which EXP should be placed. */ + +static section * +gcn_asm_select_section (tree exp, int reloc, unsigned HOST_WIDE_INT align) +{ + if (TREE_TYPE (exp) != error_mark_node + && TYPE_ADDR_SPACE (TREE_TYPE (exp)) == ADDR_SPACE_LDS) + { + if (!DECL_P (exp)) + return get_section (".lds_bss", + SECTION_WRITE | SECTION_BSS | SECTION_DEBUG, + NULL); + + return get_named_section (exp, ".lds_bss", reloc); + } + + return default_elf_select_section (exp, reloc, align); +} + +/* Implement TARGET_ASM_FUNCTION_PROLOGUE. + + Emits custom text into the assembler file at the head of each function. */ + +static void +gcn_target_asm_function_prologue (FILE *file) +{ + machine_function *offsets = gcn_compute_frame_offsets (); + + asm_fprintf (file, "\t; using %s addressing in function\n", + offsets->use_flat_addressing ? "flat" : "global"); + + if (offsets->normal_function) + { + asm_fprintf (file, "\t; frame pointer needed: %s\n", + offsets->need_frame_pointer ? "true" : "false"); + asm_fprintf (file, "\t; lr needs saving: %s\n", + offsets->lr_needs_saving ? "true" : "false"); + asm_fprintf (file, "\t; outgoing args size: %wd\n", + offsets->outgoing_args_size); + asm_fprintf (file, "\t; pretend size: %wd\n", offsets->pretend_size); + asm_fprintf (file, "\t; local vars size: %wd\n", offsets->local_vars); + asm_fprintf (file, "\t; callee save size: %wd\n", + offsets->callee_saves); + } + else + { + asm_fprintf (file, "\t; HSA kernel entry point\n"); + asm_fprintf (file, "\t; local vars size: %wd\n", offsets->local_vars); + asm_fprintf (file, "\t; outgoing args size: %wd\n", + offsets->outgoing_args_size); + + /* Enable denorms. */ + asm_fprintf (file, "\n\t; Set MODE[FP_DENORM]: allow single and double" + " input and output denorms\n"); + asm_fprintf (file, "\ts_setreg_imm32_b32\thwreg(1, 4, 4), 0xf\n\n"); + } +} + +/* Helper function for print_operand and print_operand_address. + + Print a register as the assembler requires, according to mode and name. */ + +static void +print_reg (FILE *file, rtx x) +{ + machine_mode mode = GET_MODE (x); + if (mode == BImode || mode == QImode || mode == HImode || mode == SImode + || mode == HFmode || mode == SFmode + || mode == V64SFmode || mode == V64SImode + || mode == V64QImode || mode == V64HImode) + fprintf (file, "%s", reg_names[REGNO (x)]); + else if (mode == DImode || mode == V64DImode + || mode == DFmode || mode == V64DFmode) + { + if (SGPR_REGNO_P (REGNO (x))) + fprintf (file, "s[%i:%i]", REGNO (x) - FIRST_SGPR_REG, + REGNO (x) - FIRST_SGPR_REG + 1); + else if (VGPR_REGNO_P (REGNO (x))) + fprintf (file, "v[%i:%i]", REGNO (x) - FIRST_VGPR_REG, + REGNO (x) - FIRST_VGPR_REG + 1); + else if (REGNO (x) == FLAT_SCRATCH_REG) + fprintf (file, "flat_scratch"); + else if (REGNO (x) == EXEC_REG) + fprintf (file, "exec"); + else if (REGNO (x) == VCC_LO_REG) + fprintf (file, "vcc"); + else + fprintf (file, "[%s:%s]", + reg_names[REGNO (x)], reg_names[REGNO (x) + 1]); + } + else if (mode == TImode) + { + if (SGPR_REGNO_P (REGNO (x))) + fprintf (file, "s[%i:%i]", REGNO (x) - FIRST_SGPR_REG, + REGNO (x) - FIRST_SGPR_REG + 3); + else if (VGPR_REGNO_P (REGNO (x))) + fprintf (file, "v[%i:%i]", REGNO (x) - FIRST_VGPR_REG, + REGNO (x) - FIRST_VGPR_REG + 3); + else + gcc_unreachable (); + } + else + gcc_unreachable (); +} + +/* Implement TARGET_SECTION_TYPE_FLAGS. + + Return a set of section attributes for use by TARGET_ASM_NAMED_SECTION. */ + +static unsigned int +gcn_section_type_flags (tree decl, const char *name, int reloc) +{ + if (strcmp (name, ".lds_bss") == 0) + return SECTION_WRITE | SECTION_BSS | SECTION_DEBUG; + + return default_section_type_flags (decl, name, reloc); +} + +/* Helper function for gcn_asm_output_symbol_ref. + + FIXME: If we want to have propagation blocks allocated separately and + statically like this, it would be better done via symbol refs and the + assembler/linker. This is a temporary hack. */ + +static void +gcn_print_lds_decl (FILE *f, tree var) +{ + int *offset; + machine_function *machfun = cfun->machine; + + if ((offset = machfun->lds_allocs->get (var))) + fprintf (f, "%u", (unsigned) *offset); + else + { + unsigned HOST_WIDE_INT align = DECL_ALIGN_UNIT (var); + tree type = TREE_TYPE (var); + unsigned HOST_WIDE_INT size = tree_to_uhwi (TYPE_SIZE_UNIT (type)); + if (size > align && size > 4 && align < 8) + align = 8; + + machfun->lds_allocated = ((machfun->lds_allocated + align - 1) + & ~(align - 1)); + + machfun->lds_allocs->put (var, machfun->lds_allocated); + fprintf (f, "%u", machfun->lds_allocated); + machfun->lds_allocated += size; + if (machfun->lds_allocated > LDS_SIZE) + error ("local data-share memory exhausted"); + } +} + +/* Implement ASM_OUTPUT_SYMBOL_REF via gcn-hsa.h. */ + +void +gcn_asm_output_symbol_ref (FILE *file, rtx x) +{ + tree decl; + if ((decl = SYMBOL_REF_DECL (x)) != 0 + && TREE_CODE (decl) == VAR_DECL + && AS_LDS_P (TYPE_ADDR_SPACE (TREE_TYPE (decl)))) + { + /* LDS symbols (emitted using this hook) are only used at present + to propagate worker values from an active thread to neutered + threads. Use the same offset for each such block, but don't + use zero because null pointers are used to identify the active + thread in GOACC_single_copy_start calls. */ + gcn_print_lds_decl (file, decl); + } + else + { + assemble_name (file, XSTR (x, 0)); + /* FIXME: See above -- this condition is unreachable. */ + if ((decl = SYMBOL_REF_DECL (x)) != 0 + && TREE_CODE (decl) == VAR_DECL + && AS_LDS_P (TYPE_ADDR_SPACE (TREE_TYPE (decl)))) + fputs ("@abs32", file); + } +} + +/* Implement TARGET_CONSTANT_ALIGNMENT. + + Returns the alignment in bits of a constant that is being placed in memory. + CONSTANT is the constant and BASIC_ALIGN is the alignment that the object + would ordinarily have. */ + +static HOST_WIDE_INT +gcn_constant_alignment (const_tree ARG_UNUSED (constant), + HOST_WIDE_INT basic_align) +{ + return basic_align > 128 ? basic_align : 128; +} + +/* Implement PRINT_OPERAND_ADDRESS via gcn.h. */ + +void +print_operand_address (FILE *file, rtx mem) +{ + gcc_assert (MEM_P (mem)); + + rtx reg; + rtx offset; + addr_space_t as = MEM_ADDR_SPACE (mem); + rtx addr = XEXP (mem, 0); + gcc_assert (REG_P (addr) || GET_CODE (addr) == PLUS); + + if (AS_SCRATCH_P (as)) + switch (GET_CODE (addr)) + { + case REG: + print_reg (file, addr); + break; + + case PLUS: + reg = XEXP (addr, 0); + offset = XEXP (addr, 1); + print_reg (file, reg); + if (GET_CODE (offset) == CONST_INT) + fprintf (file, " offset:" HOST_WIDE_INT_PRINT_DEC, INTVAL (offset)); + else + abort (); + break; + + default: + debug_rtx (addr); + abort (); + } + else if (AS_ANY_FLAT_P (as)) + { + if (GET_CODE (addr) == REG) + print_reg (file, addr); + else + { + gcc_assert (TARGET_GCN5_PLUS); + print_reg (file, XEXP (addr, 0)); + } + } + else if (AS_GLOBAL_P (as)) + { + gcc_assert (TARGET_GCN5_PLUS); + + rtx base = addr; + rtx vgpr_offset = NULL_RTX; + + if (GET_CODE (addr) == PLUS) + { + base = XEXP (addr, 0); + + if (GET_CODE (base) == PLUS) + { + /* (SGPR + VGPR) + CONST */ + vgpr_offset = XEXP (base, 1); + base = XEXP (base, 0); + } + else + { + rtx offset = XEXP (addr, 1); + + if (REG_P (offset)) + /* SGPR + VGPR */ + vgpr_offset = offset; + else if (CONST_INT_P (offset)) + /* VGPR + CONST or SGPR + CONST */ + ; + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + } + } + + if (REG_P (base)) + { + if (VGPR_REGNO_P (REGNO (base))) + print_reg (file, base); + else if (SGPR_REGNO_P (REGNO (base))) + { + /* The assembler requires a 64-bit VGPR pair here, even though + the offset should be only 32-bit. */ + if (vgpr_offset == NULL_RTX) + /* In this case, the vector offset is zero, so we use v0, + which is initialized by the kernel prologue to zero. */ + fprintf (file, "v[0:1]"); + else if (REG_P (vgpr_offset) + && VGPR_REGNO_P (REGNO (vgpr_offset))) + { + fprintf (file, "v[%d:%d]", + REGNO (vgpr_offset) - FIRST_VGPR_REG, + REGNO (vgpr_offset) - FIRST_VGPR_REG + 1); + } + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + } + } + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + } + else if (AS_ANY_DS_P (as)) + switch (GET_CODE (addr)) + { + case REG: + print_reg (file, addr); + break; + + case PLUS: + reg = XEXP (addr, 0); + print_reg (file, reg); + break; + + default: + debug_rtx (addr); + abort (); + } + else + switch (GET_CODE (addr)) + { + case REG: + print_reg (file, addr); + fprintf (file, ", 0"); + break; + + case PLUS: + reg = XEXP (addr, 0); + offset = XEXP (addr, 1); + print_reg (file, reg); + fprintf (file, ", "); + if (GET_CODE (offset) == REG) + print_reg (file, reg); + else if (GET_CODE (offset) == CONST_INT) + fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset)); + else + abort (); + break; + + default: + debug_rtx (addr); + abort (); + } +} + +/* Implement PRINT_OPERAND via gcn.h. + + b - print operand size as untyped operand (b8/b16/b32/b64) + B - print operand size as SI/DI untyped operand (b32/b32/b32/b64) + i - print operand size as untyped operand (i16/b32/i64) + u - print operand size as untyped operand (u16/u32/u64) + o - print operand size as memory access size for loads + (ubyte/ushort/dword/dwordx2/wordx3/dwordx4) + s - print operand size as memory access size for stores + (byte/short/dword/dwordx2/wordx3/dwordx4) + C - print conditional code for s_cbranch (_sccz/_sccnz/_vccz/_vccnz...) + c - print inverse conditional code for s_cbranch + D - print conditional code for s_cmp (eq_u64/lg_u64...) + E - print conditional code for v_cmp (eq_u64/ne_u64...) + A - print address in formatting suitable for given address space. + O - print offset:n for data share operations. + ^ - print "_co" suffix for GCN5 mnemonics + g - print "glc", if appropriate for given MEM + */ + +void +print_operand (FILE *file, rtx x, int code) +{ + int xcode = x ? GET_CODE (x) : 0; + bool invert = false; + switch (code) + { + /* Instructions have the following suffixes. + If there are two suffixes, the first is the destination type, + and the second is the source type. + + B32 Bitfield (untyped data) 32-bit + B64 Bitfield (untyped data) 64-bit + F16 floating-point 16-bit + F32 floating-point 32-bit (IEEE 754 single-precision float) + F64 floating-point 64-bit (IEEE 754 double-precision float) + I16 signed 32-bit integer + I32 signed 32-bit integer + I64 signed 64-bit integer + U16 unsigned 32-bit integer + U32 unsigned 32-bit integer + U64 unsigned 64-bit integer */ + + /* Print operand size as untyped suffix. */ + case 'b': + { + const char *s = ""; + machine_mode mode = GET_MODE (x); + if (VECTOR_MODE_P (mode)) + mode = GET_MODE_INNER (mode); + switch (GET_MODE_SIZE (mode)) + { + case 1: + s = "_b8"; + break; + case 2: + s = "_b16"; + break; + case 4: + s = "_b32"; + break; + case 8: + s = "_b64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + } + return; + case 'B': + { + const char *s = ""; + machine_mode mode = GET_MODE (x); + if (VECTOR_MODE_P (mode)) + mode = GET_MODE_INNER (mode); + switch (GET_MODE_SIZE (mode)) + { + case 1: + case 2: + case 4: + s = "_b32"; + break; + case 8: + s = "_b64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + } + return; + case 'e': + fputs ("sext(", file); + print_operand (file, x, 0); + fputs (")", file); + return; + case 'i': + case 'u': + { + bool signed_p = code == 'i'; + const char *s = ""; + machine_mode mode = GET_MODE (x); + if (VECTOR_MODE_P (mode)) + mode = GET_MODE_INNER (mode); + if (mode == VOIDmode) + switch (GET_CODE (x)) + { + case CONST_INT: + s = signed_p ? "_i32" : "_u32"; + break; + case CONST_DOUBLE: + s = "_f64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + else if (FLOAT_MODE_P (mode)) + switch (GET_MODE_SIZE (mode)) + { + case 2: + s = "_f16"; + break; + case 4: + s = "_f32"; + break; + case 8: + s = "_f64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + else + switch (GET_MODE_SIZE (mode)) + { + case 1: + s = signed_p ? "_i8" : "_u8"; + break; + case 2: + s = signed_p ? "_i16" : "_u16"; + break; + case 4: + s = signed_p ? "_i32" : "_u32"; + break; + case 8: + s = signed_p ? "_i64" : "_u64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + } + return; + /* Print operand size as untyped suffix. */ + case 'o': + { + const char *s = 0; + switch (GET_MODE_SIZE (GET_MODE (x))) + { + case 1: + s = "_ubyte"; + break; + case 2: + s = "_ushort"; + break; + /* The following are full-vector variants. */ + case 64: + s = "_ubyte"; + break; + case 128: + s = "_ushort"; + break; + } + + if (s) + { + fputs (s, file); + return; + } + + /* Fall-through - the other cases for 'o' are the same as for 's'. */ + gcc_fallthrough(); + } + case 's': + { + const char *s = ""; + switch (GET_MODE_SIZE (GET_MODE (x))) + { + case 1: + s = "_byte"; + break; + case 2: + s = "_short"; + break; + case 4: + s = "_dword"; + break; + case 8: + s = "_dwordx2"; + break; + case 12: + s = "_dwordx3"; + break; + case 16: + s = "_dwordx4"; + break; + case 32: + s = "_dwordx8"; + break; + case 64: + s = VECTOR_MODE_P (GET_MODE (x)) ? "_byte" : "_dwordx16"; + break; + /* The following are full-vector variants. */ + case 128: + s = "_short"; + break; + case 256: + s = "_dword"; + break; + case 512: + s = "_dwordx2"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + } + return; + case 'A': + if (xcode != MEM) + { + output_operand_lossage ("invalid %%xn code"); + return; + } + print_operand_address (file, x); + return; + case 'O': + { + if (xcode != MEM) + { + output_operand_lossage ("invalid %%xn code"); + return; + } + if (AS_GDS_P (MEM_ADDR_SPACE (x))) + fprintf (file, " gds"); + + rtx x0 = XEXP (x, 0); + if (AS_GLOBAL_P (MEM_ADDR_SPACE (x))) + { + gcc_assert (TARGET_GCN5_PLUS); + + fprintf (file, ", "); + + rtx base = x0; + rtx const_offset = NULL_RTX; + + if (GET_CODE (base) == PLUS) + { + rtx offset = XEXP (x0, 1); + base = XEXP (x0, 0); + + if (GET_CODE (base) == PLUS) + /* (SGPR + VGPR) + CONST */ + /* Ignore the VGPR offset for this operand. */ + base = XEXP (base, 0); + + if (CONST_INT_P (offset)) + const_offset = XEXP (x0, 1); + else if (REG_P (offset)) + /* SGPR + VGPR */ + /* Ignore the VGPR offset for this operand. */ + ; + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + } + + if (REG_P (base)) + { + if (VGPR_REGNO_P (REGNO (base))) + /* The VGPR address is specified in the %A operand. */ + fprintf (file, "off"); + else if (SGPR_REGNO_P (REGNO (base))) + print_reg (file, base); + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + } + else + output_operand_lossage ("bad ADDR_SPACE_GLOBAL address"); + + if (const_offset != NULL_RTX) + fprintf (file, " offset:" HOST_WIDE_INT_PRINT_DEC, + INTVAL (const_offset)); + + return; + } + + if (GET_CODE (x0) == REG) + return; + if (GET_CODE (x0) != PLUS) + { + output_operand_lossage ("invalid %%xn code"); + return; + } + rtx val = XEXP (x0, 1); + if (GET_CODE (val) == CONST_VECTOR) + val = CONST_VECTOR_ELT (val, 0); + if (GET_CODE (val) != CONST_INT) + { + output_operand_lossage ("invalid %%xn code"); + return; + } + fprintf (file, " offset:" HOST_WIDE_INT_PRINT_DEC, INTVAL (val)); + + } + return; + case 'c': + invert = true; + /* Fall through. */ + case 'C': + { + const char *s; + bool num = false; + if ((xcode != EQ && xcode != NE) || !REG_P (XEXP (x, 0))) + { + output_operand_lossage ("invalid %%xn code"); + return; + } + switch (REGNO (XEXP (x, 0))) + { + case VCC_REG: + case VCCZ_REG: + s = "_vcc"; + break; + case SCC_REG: + /* For some reason llvm-mc insists on scc0 instead of sccz. */ + num = true; + s = "_scc"; + break; + case EXECZ_REG: + s = "_exec"; + break; + default: + output_operand_lossage ("invalid %%xn code"); + return; + } + fputs (s, file); + if (xcode == (invert ? NE : EQ)) + fputc (num ? '0' : 'z', file); + else + fputs (num ? "1" : "nz", file); + return; + } + case 'D': + { + const char *s; + bool cmp_signed = false; + switch (xcode) + { + case EQ: + s = "_eq_"; + break; + case NE: + s = "_lg_"; + break; + case LT: + s = "_lt_"; + cmp_signed = true; + break; + case LE: + s = "_le_"; + cmp_signed = true; + break; + case GT: + s = "_gt_"; + cmp_signed = true; + break; + case GE: + s = "_ge_"; + cmp_signed = true; + break; + case LTU: + s = "_lt_"; + break; + case LEU: + s = "_le_"; + break; + case GTU: + s = "_gt_"; + break; + case GEU: + s = "_ge_"; + break; + default: + output_operand_lossage ("invalid %%xn code"); + return; + } + fputs (s, file); + fputc (cmp_signed ? 'i' : 'u', file); + + machine_mode mode = GET_MODE (XEXP (x, 0)); + + if (mode == VOIDmode) + mode = GET_MODE (XEXP (x, 1)); + + /* If both sides are constants, then assume the instruction is in + SImode since s_cmp can only do integer compares. */ + if (mode == VOIDmode) + mode = SImode; + + switch (GET_MODE_SIZE (mode)) + { + case 4: + s = "32"; + break; + case 8: + s = "64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + return; + } + case 'E': + { + const char *s; + bool cmp_signed = false; + machine_mode mode = GET_MODE (XEXP (x, 0)); + + if (mode == VOIDmode) + mode = GET_MODE (XEXP (x, 1)); + + /* If both sides are constants, assume the instruction is in SFmode + if either operand is floating point, otherwise assume SImode. */ + if (mode == VOIDmode) + { + if (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE + || GET_CODE (XEXP (x, 1)) == CONST_DOUBLE) + mode = SFmode; + else + mode = SImode; + } + + /* Use the same format code for vector comparisons. */ + if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT + || GET_MODE_CLASS (mode) == MODE_VECTOR_INT) + mode = GET_MODE_INNER (mode); + + bool float_p = GET_MODE_CLASS (mode) == MODE_FLOAT; + + switch (xcode) + { + case EQ: + s = "_eq_"; + break; + case NE: + s = float_p ? "_neq_" : "_ne_"; + break; + case LT: + s = "_lt_"; + cmp_signed = true; + break; + case LE: + s = "_le_"; + cmp_signed = true; + break; + case GT: + s = "_gt_"; + cmp_signed = true; + break; + case GE: + s = "_ge_"; + cmp_signed = true; + break; + case LTU: + s = "_lt_"; + break; + case LEU: + s = "_le_"; + break; + case GTU: + s = "_gt_"; + break; + case GEU: + s = "_ge_"; + break; + case ORDERED: + s = "_o_"; + break; + case UNORDERED: + s = "_u_"; + break; + default: + output_operand_lossage ("invalid %%xn code"); + return; + } + fputs (s, file); + fputc (float_p ? 'f' : cmp_signed ? 'i' : 'u', file); + + switch (GET_MODE_SIZE (mode)) + { + case 1: + s = "32"; + break; + case 2: + s = float_p ? "16" : "32"; + break; + case 4: + s = "32"; + break; + case 8: + s = "64"; + break; + default: + output_operand_lossage ("invalid operand %%xn code"); + return; + } + fputs (s, file); + return; + } + case 'L': + print_operand (file, gcn_operand_part (GET_MODE (x), x, 0), 0); + return; + case 'H': + print_operand (file, gcn_operand_part (GET_MODE (x), x, 1), 0); + return; + case 'R': + /* Print a scalar register number as an integer. Temporary hack. */ + gcc_assert (REG_P (x)); + fprintf (file, "%u", (int) REGNO (x)); + return; + case 'V': + /* Print a vector register number as an integer. Temporary hack. */ + gcc_assert (REG_P (x)); + fprintf (file, "%u", (int) REGNO (x) - FIRST_VGPR_REG); + return; + case 0: + if (xcode == REG) + print_reg (file, x); + else if (xcode == MEM) + output_address (GET_MODE (x), x); + else if (xcode == CONST_INT) + fprintf (file, "%i", (int) INTVAL (x)); + else if (xcode == CONST_VECTOR) + print_operand (file, CONST_VECTOR_ELT (x, 0), code); + else if (xcode == CONST_DOUBLE) + { + const char *str; + switch (gcn_inline_fp_constant_p (x, false)) + { + case 240: + str = "0.5"; + break; + case 241: + str = "-0.5"; + break; + case 242: + str = "1.0"; + break; + case 243: + str = "-1.0"; + break; + case 244: + str = "2.0"; + break; + case 245: + str = "-2.0"; + break; + case 246: + str = "4.0"; + break; + case 247: + str = "-4.0"; + break; + case 248: + str = "1/pi"; + break; + default: + rtx ix = simplify_gen_subreg (GET_MODE (x) == DFmode + ? DImode : SImode, + x, GET_MODE (x), 0); + if (x) + print_operand (file, ix, code); + else + output_operand_lossage ("invlaid fp constant"); + return; + break; + } + fprintf (file, str); + return; + } + else + output_addr_const (file, x); + return; + case '^': + if (TARGET_GCN5_PLUS) + fputs ("_co", file); + return; + case 'g': + gcc_assert (xcode == MEM); + if (MEM_VOLATILE_P (x)) + fputs (" glc", file); + return; + default: + output_operand_lossage ("invalid %%xn code"); + } + gcc_unreachable (); +} + +/* }}} */ +/* {{{ TARGET hook overrides. */ + +#undef TARGET_ADDR_SPACE_ADDRESS_MODE +#define TARGET_ADDR_SPACE_ADDRESS_MODE gcn_addr_space_address_mode +#undef TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P +#define TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P \ + gcn_addr_space_legitimate_address_p +#undef TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS +#define TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS gcn_addr_space_legitimize_address +#undef TARGET_ADDR_SPACE_POINTER_MODE +#define TARGET_ADDR_SPACE_POINTER_MODE gcn_addr_space_pointer_mode +#undef TARGET_ADDR_SPACE_SUBSET_P +#define TARGET_ADDR_SPACE_SUBSET_P gcn_addr_space_subset_p +#undef TARGET_ADDR_SPACE_CONVERT +#define TARGET_ADDR_SPACE_CONVERT gcn_addr_space_convert +#undef TARGET_ARG_PARTIAL_BYTES +#define TARGET_ARG_PARTIAL_BYTES gcn_arg_partial_bytes +#undef TARGET_ASM_ALIGNED_DI_OP +#define TARGET_ASM_ALIGNED_DI_OP "\t.8byte\t" +#undef TARGET_ASM_CONSTRUCTOR +#define TARGET_ASM_CONSTRUCTOR gcn_disable_constructors +#undef TARGET_ASM_DESTRUCTOR +#define TARGET_ASM_DESTRUCTOR gcn_disable_constructors +#undef TARGET_ASM_FILE_START +#define TARGET_ASM_FILE_START output_file_start +#undef TARGET_ASM_FUNCTION_PROLOGUE +#define TARGET_ASM_FUNCTION_PROLOGUE gcn_target_asm_function_prologue +#undef TARGET_ASM_SELECT_SECTION +#define TARGET_ASM_SELECT_SECTION gcn_asm_select_section +#undef TARGET_ASM_TRAMPOLINE_TEMPLATE +#define TARGET_ASM_TRAMPOLINE_TEMPLATE gcn_asm_trampoline_template +#undef TARGET_ATTRIBUTE_TABLE +#define TARGET_ATTRIBUTE_TABLE gcn_attribute_table +#undef TARGET_BUILTIN_DECL +#define TARGET_BUILTIN_DECL gcn_builtin_decl +#undef TARGET_CAN_CHANGE_MODE_CLASS +#define TARGET_CAN_CHANGE_MODE_CLASS gcn_can_change_mode_class +#undef TARGET_CAN_ELIMINATE +#define TARGET_CAN_ELIMINATE gcn_can_eliminate_p +#undef TARGET_CANNOT_COPY_INSN_P +#define TARGET_CANNOT_COPY_INSN_P gcn_cannot_copy_insn_p +#undef TARGET_CLASS_LIKELY_SPILLED_P +#define TARGET_CLASS_LIKELY_SPILLED_P gcn_class_likely_spilled_p +#undef TARGET_CLASS_MAX_NREGS +#define TARGET_CLASS_MAX_NREGS gcn_class_max_nregs +#undef TARGET_CONDITIONAL_REGISTER_USAGE +#define TARGET_CONDITIONAL_REGISTER_USAGE gcn_conditional_register_usage +#undef TARGET_CONSTANT_ALIGNMENT +#define TARGET_CONSTANT_ALIGNMENT gcn_constant_alignment +#undef TARGET_DEBUG_UNWIND_INFO +#define TARGET_DEBUG_UNWIND_INFO gcn_debug_unwind_info +#undef TARGET_EXPAND_BUILTIN +#define TARGET_EXPAND_BUILTIN gcn_expand_builtin +#undef TARGET_FUNCTION_ARG +#undef TARGET_FUNCTION_ARG_ADVANCE +#define TARGET_FUNCTION_ARG_ADVANCE gcn_function_arg_advance +#define TARGET_FUNCTION_ARG gcn_function_arg +#undef TARGET_FUNCTION_VALUE +#define TARGET_FUNCTION_VALUE gcn_function_value +#undef TARGET_FUNCTION_VALUE_REGNO_P +#define TARGET_FUNCTION_VALUE_REGNO_P gcn_function_value_regno_p +#undef TARGET_GIMPLIFY_VA_ARG_EXPR +#define TARGET_GIMPLIFY_VA_ARG_EXPR gcn_gimplify_va_arg_expr +#undef TARGET_GOACC_ADJUST_PROPAGATION_RECORD +#define TARGET_GOACC_ADJUST_PROPAGATION_RECORD \ + gcn_goacc_adjust_propagation_record +#undef TARGET_GOACC_ADJUST_GANGPRIVATE_DECL +#define TARGET_GOACC_ADJUST_GANGPRIVATE_DECL gcn_goacc_adjust_gangprivate_decl +#undef TARGET_GOACC_FORK_JOIN +#define TARGET_GOACC_FORK_JOIN gcn_fork_join +#undef TARGET_GOACC_REDUCTION +#define TARGET_GOACC_REDUCTION gcn_goacc_reduction +#undef TARGET_GOACC_VALIDATE_DIMS +#define TARGET_GOACC_VALIDATE_DIMS gcn_goacc_validate_dims +#undef TARGET_GOACC_WORKER_PARTITIONING +#define TARGET_GOACC_WORKER_PARTITIONING true +#undef TARGET_HARD_REGNO_MODE_OK +#define TARGET_HARD_REGNO_MODE_OK gcn_hard_regno_mode_ok +#undef TARGET_HARD_REGNO_NREGS +#define TARGET_HARD_REGNO_NREGS gcn_hard_regno_nregs +#undef TARGET_HAVE_SPECULATION_SAFE_VALUE +#define TARGET_HAVE_SPECULATION_SAFE_VALUE speculation_safe_value_not_needed +#undef TARGET_INIT_BUILTINS +#define TARGET_INIT_BUILTINS gcn_init_builtins +#undef TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS +#define TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS \ + gcn_ira_change_pseudo_allocno_class +#undef TARGET_LEGITIMATE_CONSTANT_P +#define TARGET_LEGITIMATE_CONSTANT_P gcn_legitimate_constant_p +#undef TARGET_LRA_P +#define TARGET_LRA_P hook_bool_void_true +#undef TARGET_MACHINE_DEPENDENT_REORG +#define TARGET_MACHINE_DEPENDENT_REORG gcn_md_reorg +#undef TARGET_MEMORY_MOVE_COST +#define TARGET_MEMORY_MOVE_COST gcn_memory_move_cost +#undef TARGET_MODES_TIEABLE_P +#define TARGET_MODES_TIEABLE_P gcn_modes_tieable_p +#undef TARGET_OPTION_OVERRIDE +#define TARGET_OPTION_OVERRIDE gcn_option_override +#undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED +#define TARGET_PRETEND_OUTGOING_VARARGS_NAMED \ + gcn_pretend_outgoing_varargs_named +#undef TARGET_PROMOTE_FUNCTION_MODE +#define TARGET_PROMOTE_FUNCTION_MODE gcn_promote_function_mode +#undef TARGET_REGISTER_MOVE_COST +#define TARGET_REGISTER_MOVE_COST gcn_register_move_cost +#undef TARGET_RETURN_IN_MEMORY +#define TARGET_RETURN_IN_MEMORY gcn_return_in_memory +#undef TARGET_RTX_COSTS +#define TARGET_RTX_COSTS gcn_rtx_costs +#undef TARGET_SECONDARY_RELOAD +#define TARGET_SECONDARY_RELOAD gcn_secondary_reload +#undef TARGET_SECTION_TYPE_FLAGS +#define TARGET_SECTION_TYPE_FLAGS gcn_section_type_flags +#undef TARGET_SMALL_REGISTER_CLASSES_FOR_MODE_P +#define TARGET_SMALL_REGISTER_CLASSES_FOR_MODE_P \ + gcn_small_register_classes_for_mode_p +#undef TARGET_SPILL_CLASS +#define TARGET_SPILL_CLASS gcn_spill_class +#undef TARGET_STRICT_ARGUMENT_NAMING +#define TARGET_STRICT_ARGUMENT_NAMING gcn_strict_argument_naming +#undef TARGET_TRAMPOLINE_INIT +#define TARGET_TRAMPOLINE_INIT gcn_trampoline_init +#undef TARGET_TRULY_NOOP_TRUNCATION +#define TARGET_TRULY_NOOP_TRUNCATION gcn_truly_noop_truncation +#undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST +#define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST gcn_vectorization_cost +#undef TARGET_VECTORIZE_GET_MASK_MODE +#define TARGET_VECTORIZE_GET_MASK_MODE gcn_vectorize_get_mask_mode +#undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE +#define TARGET_VECTORIZE_PREFERRED_SIMD_MODE gcn_vectorize_preferred_simd_mode +#undef TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT +#define TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT \ + gcn_preferred_vector_alignment +#undef TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT +#define TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT \ + gcn_vectorize_support_vector_misalignment +#undef TARGET_VECTORIZE_VEC_PERM_CONST +#define TARGET_VECTORIZE_VEC_PERM_CONST gcn_vectorize_vec_perm_const +#undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE +#define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ + gcn_vector_alignment_reachable +#undef TARGET_VECTOR_MODE_SUPPORTED_P +#define TARGET_VECTOR_MODE_SUPPORTED_P gcn_vector_mode_supported_p + +struct gcc_target targetm = TARGET_INITIALIZER; + +#include "gt-gcn.h" +/* }}} */ diff --git a/gcc/config/gcn/gcn.h b/gcc/config/gcn/gcn.h new file mode 100644 index 00000000000..b3b2d1ad3f9 --- /dev/null +++ b/gcc/config/gcn/gcn.h @@ -0,0 +1,662 @@ +/* Copyright (C) 2016-2019 Free Software Foundation, Inc. + + This file is free software; you can redistribute it and/or modify it under + the terms of the GNU General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your option) + any later version. + + This file is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + for more details. + + You should have received a copy of the GNU General Public License + along with GCC; see the file COPYING3. If not see + . */ + +#include "config/gcn/gcn-opts.h" + +#define TARGET_CPU_CPP_BUILTINS() \ + do \ + { \ + builtin_define ("__AMDGCN__"); \ + if (TARGET_GCN3) \ + builtin_define ("__GCN3__"); \ + else if (TARGET_GCN5) \ + builtin_define ("__GCN5__"); \ + } \ + while(0) + +/* Support for a compile-time default architecture and tuning. + The rules are: + --with-arch is ignored if -march is specified. + --with-tune is ignored if -mtune is specified. */ +#define OPTION_DEFAULT_SPECS \ + {"arch", "%{!march=*:-march=%(VALUE)}" }, \ + {"tune", "%{!mtune=*:-mtune=%(VALUE)}" } + +/* Default target_flags if no switches specified. */ +#ifndef TARGET_DEFAULT +#define TARGET_DEFAULT 0 +#endif + + +/* Storage Layout */ +#define BITS_BIG_ENDIAN 0 +#define BYTES_BIG_ENDIAN 0 +#define WORDS_BIG_ENDIAN 0 + +#define BITS_PER_WORD 32 +#define UNITS_PER_WORD (BITS_PER_WORD/BITS_PER_UNIT) +#define LIBGCC2_UNITS_PER_WORD 4 + +#define POINTER_SIZE 64 +#define PARM_BOUNDARY 64 +#define STACK_BOUNDARY 64 +#define FUNCTION_BOUNDARY 32 +#define BIGGEST_ALIGNMENT 64 +#define EMPTY_FIELD_BOUNDARY 32 +#define MAX_FIXED_MODE_SIZE 64 +#define MAX_REGS_PER_ADDRESS 2 +#define STACK_SIZE_MODE DImode +#define Pmode DImode +#define CASE_VECTOR_MODE DImode +#define FUNCTION_MODE QImode + +#define DATA_ALIGNMENT(TYPE,ALIGN) ((ALIGN) > 128 ? (ALIGN) : 128) +#define LOCAL_ALIGNMENT(TYPE,ALIGN) ((ALIGN) > 64 ? (ALIGN) : 64) +#define STACK_SLOT_ALIGNMENT(TYPE,MODE,ALIGN) ((ALIGN) > 64 ? (ALIGN) : 64) +#define STRICT_ALIGNMENT 1 + +/* Type Layout: match what x86_64 does. */ +#define INT_TYPE_SIZE 32 +#define LONG_TYPE_SIZE 64 +#define LONG_LONG_TYPE_SIZE 64 +#define FLOAT_TYPE_SIZE 32 +#define DOUBLE_TYPE_SIZE 64 +#define LONG_DOUBLE_TYPE_SIZE 64 +#define DEFAULT_SIGNED_CHAR 1 +#define PCC_BITFIELD_TYPE_MATTERS 1 + +/* Frame Layout */ +#define FRAME_GROWS_DOWNWARD 0 +#define ARGS_GROW_DOWNWARD 1 +#define STACK_POINTER_OFFSET 0 +#define FIRST_PARM_OFFSET(FNDECL) 0 +#define DYNAMIC_CHAIN_ADDRESS(FP) plus_constant (Pmode, (FP), -16) +#define INCOMING_RETURN_ADDR_RTX gen_rtx_REG (Pmode, LINK_REGNUM) +#define STACK_DYNAMIC_OFFSET(FNDECL) (-crtl->outgoing_args_size) +#define ACCUMULATE_OUTGOING_ARGS 1 +#define RETURN_ADDR_RTX(COUNT,FRAMEADDR) \ + ((COUNT) == 0 ? get_hard_reg_initial_val (Pmode, LINK_REGNUM) : NULL_RTX) + +/* Register Basics */ +#define FIRST_SGPR_REG 0 +#define SGPR_REGNO(N) ((N)+FIRST_SGPR_REG) +#define LAST_SGPR_REG 101 + +#define FLAT_SCRATCH_REG 102 +#define FLAT_SCRATCH_LO_REG 102 +#define FLAT_SCRATCH_HI_REG 103 +#define XNACK_MASK_REG 104 +#define XNACK_MASK_LO_REG 104 +#define XNACK_MASK_HI_REG 105 +#define VCC_LO_REG 106 +#define VCC_HI_REG 107 +#define VCCZ_REG 108 +#define TBA_REG 109 +#define TBA_LO_REG 109 +#define TBA_HI_REG 110 +#define TMA_REG 111 +#define TMA_LO_REG 111 +#define TMA_HI_REG 112 +#define TTMP0_REG 113 +#define TTMP11_REG 124 +#define M0_REG 125 +#define EXEC_REG 126 +#define EXEC_LO_REG 126 +#define EXEC_HI_REG 127 +#define EXECZ_REG 128 +#define SCC_REG 129 +/* 132-159 are reserved to simplify masks. */ +#define FIRST_VGPR_REG 160 +#define VGPR_REGNO(N) ((N)+FIRST_VGPR_REG) +#define LAST_VGPR_REG 415 + +/* Frame Registers, and other registers */ + +#define HARD_FRAME_POINTER_REGNUM 14 +#define STACK_POINTER_REGNUM 16 +#define LINK_REGNUM 18 +#define EXEC_SAVE_REG 20 +#define CC_SAVE_REG 22 +#define RETURN_VALUE_REG 24 /* Must be divisible by 4. */ +#define STATIC_CHAIN_REGNUM 30 +#define WORK_ITEM_ID_Z_REG 162 +#define SOFT_ARG_REG 416 +#define FRAME_POINTER_REGNUM 418 +#define FIRST_PSEUDO_REGISTER 420 + +#define FIRST_PARM_REG 24 +#define NUM_PARM_REGS 6 + +/* There is no arg pointer. Just choose random fixed register that does + not intefere with anything. */ +#define ARG_POINTER_REGNUM SOFT_ARG_REG + +#define HARD_FRAME_POINTER_IS_ARG_POINTER 0 +#define HARD_FRAME_POINTER_IS_FRAME_POINTER 0 + +#define SGPR_OR_VGPR_REGNO_P(N) ((N)>=FIRST_VGPR_REG && (N) <= LAST_SGPR_REG) +#define SGPR_REGNO_P(N) ((N) <= LAST_SGPR_REG) +#define VGPR_REGNO_P(N) ((N)>=FIRST_VGPR_REG && (N) <= LAST_VGPR_REG) +#define SSRC_REGNO_P(N) ((N) <= SCC_REG && (N) != VCCZ_REG) +#define SDST_REGNO_P(N) ((N) <= EXEC_HI_REG && (N) != VCCZ_REG) +#define CC_REG_P(X) (REG_P (X) && CC_REGNO_P (REGNO (X))) +#define CC_REGNO_P(X) ((X) == SCC_REG || (X) == VCC_REG) +#define FUNCTION_ARG_REGNO_P(N) \ + ((N) >= FIRST_PARM_REG && (N) < (FIRST_PARM_REG + NUM_PARM_REGS)) + + +#define FIXED_REGISTERS { \ + /* Scalars. */ \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ +/* fp sp lr. */ \ + 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, \ +/* exec_save, cc_save */ \ + 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, \ + /* Special regs and padding. */ \ +/* flat xnack vcc tba tma ttmp */ \ + 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ +/* m0 exec scc */ \ + 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + /* VGRPs */ \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + /* Other registers. */ \ + 1, 1, 1, 1 \ +} + +#define CALL_USED_REGISTERS { \ + /* Scalars. */ \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, \ + /* Special regs and padding. */ \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + /* VGRPs */ \ + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + /* Other registers. */ \ + 1, 1, 1, 1 \ +} + + +#define HARD_REGNO_RENAME_OK(FROM, TO) \ + gcn_hard_regno_rename_ok (FROM, TO) + +#define HARD_REGNO_CALLER_SAVE_MODE(HARDREG, NREGS, MODE) \ + gcn_hard_regno_caller_save_mode ((HARDREG), (NREGS), (MODE)) + +/* Register Classes */ + +enum reg_class +{ + NO_REGS, + + /* SCC */ + SCC_CONDITIONAL_REG, + + /* VCCZ */ + VCCZ_CONDITIONAL_REG, + + /* VCC */ + VCC_CONDITIONAL_REG, + + /* EXECZ */ + EXECZ_CONDITIONAL_REG, + + /* SCC VCCZ EXECZ */ + ALL_CONDITIONAL_REGS, + + /* EXEC */ + EXEC_MASK_REG, + + /* SGPR0-101 */ + SGPR_REGS, + + /* SGPR0-101 EXEC_LO/EXEC_HI */ + SGPR_EXEC_REGS, + + /* SGPR0-101, FLAT_SCRATCH_LO/HI, VCC LO/HI, TBA LO/HI, TMA LO/HI, TTMP0-11, + M0, VCCZ, SCC + (EXEC_LO/HI, EXECZ excluded to prevent compiler misuse.) */ + SGPR_VOP_SRC_REGS, + + /* SGPR0-101, FLAT_SCRATCH_LO/HI, XNACK_MASK_LO/HI, VCC LO/HI, TBA LO/HI + TMA LO/HI, TTMP0-11 */ + SGPR_MEM_SRC_REGS, + + /* SGPR0-101, FLAT_SCRATCH_LO/HI, XNACK_MASK_LO/HI, VCC LO/HI, TBA LO/HI + TMA LO/HI, TTMP0-11, M0, EXEC LO/HI */ + SGPR_DST_REGS, + + /* SGPR0-101, FLAT_SCRATCH_LO/HI, XNACK_MASK_LO/HI, VCC LO/HI, TBA LO/HI + TMA LO/HI, TTMP0-11 */ + SGPR_SRC_REGS, + GENERAL_REGS, + VGPR_REGS, + ALL_GPR_REGS, + SRCDST_REGS, + AFP_REGS, + ALL_REGS, + LIM_REG_CLASSES +}; + +#define N_REG_CLASSES (int) LIM_REG_CLASSES + +#define REG_CLASS_NAMES \ +{ "NO_REGS", \ + "SCC_CONDITIONAL_REG", \ + "VCCZ_CONDITIONAL_REG", \ + "VCC_CONDITIONAL_REG", \ + "EXECZ_CONDITIONAL_REG", \ + "ALL_CONDITIONAL_REGS", \ + "EXEC_MASK_REG", \ + "SGPR_REGS", \ + "SGPR_EXEC_REGS", \ + "SGPR_VOP3A_SRC_REGS", \ + "SGPR_MEM_SRC_REGS", \ + "SGPR_DST_REGS", \ + "SGPR_SRC_REGS", \ + "GENERAL_REGS", \ + "VGPR_REGS", \ + "ALL_GPR_REGS", \ + "SRCDST_REGS", \ + "AFP_REGS", \ + "ALL_REGS" \ +} + +#define NAMED_REG_MASK(N) (1<<((N)-3*32)) +#define NAMED_REG_MASK2(N) (1<<((N)-4*32)) + +#define REG_CLASS_CONTENTS { \ + /* NO_REGS. */ \ + {0, 0, 0, 0, \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SCC_CONDITIONAL_REG. */ \ + {0, 0, 0, 0, \ + NAMED_REG_MASK2 (SCC_REG), 0, 0, 0, \ + 0, 0, 0, 0, 0}, \ + /* VCCZ_CONDITIONAL_REG. */ \ + {0, 0, 0, NAMED_REG_MASK (VCCZ_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* VCC_CONDITIONAL_REG. */ \ + {0, 0, 0, NAMED_REG_MASK (VCC_LO_REG)|NAMED_REG_MASK (VCC_HI_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* EXECZ_CONDITIONAL_REG. */ \ + {0, 0, 0, 0, \ + NAMED_REG_MASK2 (EXECZ_REG), 0, 0, 0, \ + 0, 0, 0, 0, 0}, \ + /* ALL_CONDITIONAL_REGS. */ \ + {0, 0, 0, NAMED_REG_MASK (VCCZ_REG), \ + NAMED_REG_MASK2 (EXECZ_REG) | NAMED_REG_MASK2 (SCC_REG), 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* EXEC_MASK_REG. */ \ + {0, 0, 0, NAMED_REG_MASK (EXEC_LO_REG) | NAMED_REG_MASK (EXEC_HI_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, 0xf1, \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_EXEC_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, \ + 0xf1 | NAMED_REG_MASK (EXEC_LO_REG) | NAMED_REG_MASK (EXEC_HI_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_VOP_SRC_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff \ + -NAMED_REG_MASK (EXEC_LO_REG) \ + -NAMED_REG_MASK (EXEC_HI_REG), \ + NAMED_REG_MASK2 (SCC_REG), 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_MEM_SRC_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff-NAMED_REG_MASK (VCCZ_REG)-NAMED_REG_MASK (M0_REG) \ + -NAMED_REG_MASK (EXEC_LO_REG)-NAMED_REG_MASK (EXEC_HI_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_DST_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff-NAMED_REG_MASK (VCCZ_REG), \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* SGPR_SRC_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, \ + NAMED_REG_MASK2 (EXECZ_REG) | NAMED_REG_MASK2 (SCC_REG), 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* GENERAL_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, 0xf1, \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0}, \ + /* VGPR_REGS. */ \ + {0, 0, 0, 0, \ + 0, 0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0}, \ + /* ALL_GPR_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, 0xf1, \ + 0, 0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0}, \ + /* SRCDST_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff-NAMED_REG_MASK (VCCZ_REG), \ + 0, 0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0}, \ + /* AFP_REGS. */ \ + {0, 0, 0, 0, \ + 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0xf}, \ + /* ALL_REGS. */ \ + {0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, \ + 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0 }} + +#define REGNO_REG_CLASS(REGNO) gcn_regno_reg_class (REGNO) +#define MODE_CODE_BASE_REG_CLASS(MODE, AS, OUTER, INDEX) \ + gcn_mode_code_base_reg_class (MODE, AS, OUTER, INDEX) +#define REGNO_MODE_CODE_OK_FOR_BASE_P(NUM, MODE, AS, OUTER, INDEX) \ + gcn_regno_mode_code_ok_for_base_p (NUM, MODE, AS, OUTER, INDEX) +#define INDEX_REG_CLASS VGPR_REGS +#define REGNO_OK_FOR_INDEX_P(regno) regno_ok_for_index_p (regno) + + +/* Address spaces. */ +enum gcn_address_spaces +{ + ADDR_SPACE_DEFAULT = 0, + ADDR_SPACE_FLAT, + ADDR_SPACE_SCALAR_FLAT, + ADDR_SPACE_FLAT_SCRATCH, + ADDR_SPACE_LDS, + ADDR_SPACE_GDS, + ADDR_SPACE_SCRATCH, + ADDR_SPACE_GLOBAL +}; +#define REGISTER_TARGET_PRAGMAS() do { \ + c_register_addr_space ("__flat", ADDR_SPACE_FLAT); \ + c_register_addr_space ("__flat_scratch", ADDR_SPACE_FLAT_SCRATCH); \ + c_register_addr_space ("__scalar_flat", ADDR_SPACE_SCALAR_FLAT); \ + c_register_addr_space ("__lds", ADDR_SPACE_LDS); \ + c_register_addr_space ("__gds", ADDR_SPACE_GDS); \ + c_register_addr_space ("__global", ADDR_SPACE_GLOBAL); \ +} while (0); + +#define STACK_ADDR_SPACE \ + (TARGET_GCN5_PLUS ? ADDR_SPACE_GLOBAL : ADDR_SPACE_FLAT) +#define DEFAULT_ADDR_SPACE \ + ((cfun && cfun->machine && !cfun->machine->use_flat_addressing) \ + ? ADDR_SPACE_GLOBAL : ADDR_SPACE_FLAT) +#define AS_SCALAR_FLAT_P(AS) ((AS) == ADDR_SPACE_SCALAR_FLAT) +#define AS_FLAT_SCRATCH_P(AS) ((AS) == ADDR_SPACE_FLAT_SCRATCH) +#define AS_FLAT_P(AS) ((AS) == ADDR_SPACE_FLAT \ + || ((AS) == ADDR_SPACE_DEFAULT \ + && DEFAULT_ADDR_SPACE == ADDR_SPACE_FLAT)) +#define AS_LDS_P(AS) ((AS) == ADDR_SPACE_LDS) +#define AS_GDS_P(AS) ((AS) == ADDR_SPACE_GDS) +#define AS_SCRATCH_P(AS) ((AS) == ADDR_SPACE_SCRATCH) +#define AS_GLOBAL_P(AS) ((AS) == ADDR_SPACE_GLOBAL \ + || ((AS) == ADDR_SPACE_DEFAULT \ + && DEFAULT_ADDR_SPACE == ADDR_SPACE_GLOBAL)) +#define AS_ANY_FLAT_P(AS) (AS_FLAT_SCRATCH_P (AS) || AS_FLAT_P (AS)) +#define AS_ANY_DS_P(AS) (AS_LDS_P (AS) || AS_GDS_P (AS)) + + +/* Instruction Output */ +#define REGISTER_NAMES \ + {"s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7", "s8", "s9", "s10", \ + "s11", "s12", "s13", "s14", "s15", "s16", "s17", "s18", "s19", "s20", \ + "s21", "s22", "s23", "s24", "s25", "s26", "s27", "s28", "s29", "s30", \ + "s31", "s32", "s33", "s34", "s35", "s36", "s37", "s38", "s39", "s40", \ + "s41", "s42", "s43", "s44", "s45", "s46", "s47", "s48", "s49", "s50", \ + "s51", "s52", "s53", "s54", "s55", "s56", "s57", "s58", "s59", "s60", \ + "s61", "s62", "s63", "s64", "s65", "s66", "s67", "s68", "s69", "s70", \ + "s71", "s72", "s73", "s74", "s75", "s76", "s77", "s78", "s79", "s80", \ + "s81", "s82", "s83", "s84", "s85", "s86", "s87", "s88", "s89", "s90", \ + "s91", "s92", "s93", "s94", "s95", "s96", "s97", "s98", "s99", \ + "s100", "s101", \ + "flat_scratch_lo", "flat_scratch_hi", "xnack_mask_lo", "xnack_mask_hi", \ + "vcc_lo", "vcc_hi", "vccz", "tba_lo", "tba_hi", "tma_lo", "tma_hi", \ + "ttmp0", "ttmp1", "ttmp2", "ttmp3", "ttmp4", "ttmp5", "ttmp6", "ttmp7", \ + "ttmp8", "ttmp9", "ttmp10", "ttmp11", "m0", "exec_lo", "exec_hi", \ + "execz", "scc", \ + "res130", "res131", "res132", "res133", "res134", "res135", "res136", \ + "res137", "res138", "res139", "res140", "res141", "res142", "res143", \ + "res144", "res145", "res146", "res147", "res148", "res149", "res150", \ + "res151", "res152", "res153", "res154", "res155", "res156", "res157", \ + "res158", "res159", \ + "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", \ + "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", \ + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", \ + "v31", "v32", "v33", "v34", "v35", "v36", "v37", "v38", "v39", "v40", \ + "v41", "v42", "v43", "v44", "v45", "v46", "v47", "v48", "v49", "v50", \ + "v51", "v52", "v53", "v54", "v55", "v56", "v57", "v58", "v59", "v60", \ + "v61", "v62", "v63", "v64", "v65", "v66", "v67", "v68", "v69", "v70", \ + "v71", "v72", "v73", "v74", "v75", "v76", "v77", "v78", "v79", "v80", \ + "v81", "v82", "v83", "v84", "v85", "v86", "v87", "v88", "v89", "v90", \ + "v91", "v92", "v93", "v94", "v95", "v96", "v97", "v98", "v99", "v100", \ + "v101", "v102", "v103", "v104", "v105", "v106", "v107", "v108", "v109", \ + "v110", "v111", "v112", "v113", "v114", "v115", "v116", "v117", "v118", \ + "v119", "v120", "v121", "v122", "v123", "v124", "v125", "v126", "v127", \ + "v128", "v129", "v130", "v131", "v132", "v133", "v134", "v135", "v136", \ + "v137", "v138", "v139", "v140", "v141", "v142", "v143", "v144", "v145", \ + "v146", "v147", "v148", "v149", "v150", "v151", "v152", "v153", "v154", \ + "v155", "v156", "v157", "v158", "v159", "v160", "v161", "v162", "v163", \ + "v164", "v165", "v166", "v167", "v168", "v169", "v170", "v171", "v172", \ + "v173", "v174", "v175", "v176", "v177", "v178", "v179", "v180", "v181", \ + "v182", "v183", "v184", "v185", "v186", "v187", "v188", "v189", "v190", \ + "v191", "v192", "v193", "v194", "v195", "v196", "v197", "v198", "v199", \ + "v200", "v201", "v202", "v203", "v204", "v205", "v206", "v207", "v208", \ + "v209", "v210", "v211", "v212", "v213", "v214", "v215", "v216", "v217", \ + "v218", "v219", "v220", "v221", "v222", "v223", "v224", "v225", "v226", \ + "v227", "v228", "v229", "v230", "v231", "v232", "v233", "v234", "v235", \ + "v236", "v237", "v238", "v239", "v240", "v241", "v242", "v243", "v244", \ + "v245", "v246", "v247", "v248", "v249", "v250", "v251", "v252", "v253", \ + "v254", "v255", \ + "?ap0", "?ap1", "?fp0", "?fp1" } + +#define PRINT_OPERAND(FILE, X, CODE) print_operand(FILE, X, CODE) +#define PRINT_OPERAND_ADDRESS(FILE, ADDR) print_operand_address (FILE, ADDR) +#define PRINT_OPERAND_PUNCT_VALID_P(CODE) (CODE == '^') + + +/* Register Arguments */ + +#ifndef USED_FOR_TARGET + +#define GCN_KERNEL_ARG_TYPES 19 +struct GTY(()) gcn_kernel_args +{ + long requested; + int reg[GCN_KERNEL_ARG_TYPES]; + int order[GCN_KERNEL_ARG_TYPES]; + int nargs, nsgprs; +}; + +typedef struct gcn_args +{ + /* True if this isn't a kernel (HSA runtime entrypoint). */ + bool normal_function; + tree fntype; + struct gcn_kernel_args args; + int num; + int offset; + int alignment; +} CUMULATIVE_ARGS; +#endif + +#define INIT_CUMULATIVE_ARGS(CUM,FNTYPE,LIBNAME,FNDECL,N_NAMED_ARGS) \ + gcn_init_cumulative_args (&(CUM), (FNTYPE), (LIBNAME), (FNDECL), \ + (N_NAMED_ARGS) != -1) + + +#ifndef USED_FOR_TARGET + +#include "hash-table.h" +#include "hash-map.h" +#include "vec.h" + +struct GTY(()) machine_function +{ + struct gcn_kernel_args args; + int kernarg_segment_alignment; + int kernarg_segment_byte_size; + /* Frame layout info for normal functions. */ + bool normal_function; + bool need_frame_pointer; + bool lr_needs_saving; + HOST_WIDE_INT outgoing_args_size; + HOST_WIDE_INT pretend_size; + HOST_WIDE_INT local_vars; + HOST_WIDE_INT callee_saves; + + unsigned lds_allocated; + hash_map *lds_allocs; + + vec *reduc_decls; + + bool use_flat_addressing; +}; +#endif + + +/* Codes for all the GCN builtins. */ + +enum gcn_builtin_codes +{ +#define DEF_BUILTIN(fcode, icode, name, type, params, expander) \ + GCN_BUILTIN_ ## fcode, +#define DEF_BUILTIN_BINOP_INT_FP(fcode, ic, name) \ + GCN_BUILTIN_ ## fcode ## _V64SI, \ + GCN_BUILTIN_ ## fcode ## _V64SI_unspec, +#include "gcn-builtins.def" +#undef DEF_BUILTIN +#undef DEF_BUILTIN_BINOP_INT_FP + GCN_BUILTIN_MAX +}; + + +/* Misc */ + +/* We can load/store 128-bit quantities, but having this larger than + MAX_FIXED_MODE_SIZE (which we want to be 64 bits) causes problems. */ +#define MOVE_MAX 8 + +#define AVOID_CCMODE_COPIES 1 +#define SLOW_BYTE_ACCESS 0 +#define WORD_REGISTER_OPERATIONS 1 + +/* Definitions for register eliminations. + + This is an array of structures. Each structure initializes one pair + of eliminable registers. The "from" register number is given first, + followed by "to". Eliminations of the same "from" register are listed + in order of preference. */ + +#define ELIMINABLE_REGS \ +{{ ARG_POINTER_REGNUM, STACK_POINTER_REGNUM }, \ + { ARG_POINTER_REGNUM, HARD_FRAME_POINTER_REGNUM }, \ + { FRAME_POINTER_REGNUM, STACK_POINTER_REGNUM }, \ + { FRAME_POINTER_REGNUM, HARD_FRAME_POINTER_REGNUM }} + +/* Define the offset between two registers, one to be eliminated, and the + other its replacement, at the start of a routine. */ + +#define INITIAL_ELIMINATION_OFFSET(FROM, TO, OFFSET) \ + ((OFFSET) = gcn_initial_elimination_offset ((FROM), (TO))) + + +/* Define this macro if it is advisable to hold scalars in registers + in a wider mode than that declared by the program. In such cases, + the value is constrained to be within the bounds of the declared + type, but kept valid in the wider mode. The signedness of the + extension may differ from that of the type. */ + +#define PROMOTE_MODE(MODE,UNSIGNEDP,TYPE) \ + if (GET_MODE_CLASS (MODE) == MODE_INT \ + && (TYPE == NULL || TREE_CODE (TYPE) != VECTOR_TYPE) \ + && GET_MODE_SIZE (MODE) < UNITS_PER_WORD) \ + { \ + (MODE) = SImode; \ + } + +/* This needs to match gcn_function_value. */ +#define LIBCALL_VALUE(MODE) gen_rtx_REG (MODE, SGPR_REGNO (RETURN_VALUE_REG)) + + +/* Costs. */ + +/* Branches are to be dicouraged when theres an alternative. + FIXME: This number is plucked from the air. */ +#define BRANCH_COST(SPEED_P, PREDICABLE_P) 10 + + +/* Profiling */ +#define FUNCTION_PROFILER(FILE, LABELNO) +#define NO_PROFILE_COUNTERS 1 +#define PROFILE_BEFORE_PROLOGUE 0 + +/* Trampolines */ +#define TRAMPOLINE_SIZE 36 +#define TRAMPOLINE_ALIGNMENT 64 diff --git a/gcc/config/gcn/gcn.opt b/gcc/config/gcn/gcn.opt new file mode 100644 index 00000000000..2fd3996edba --- /dev/null +++ b/gcc/config/gcn/gcn.opt @@ -0,0 +1,78 @@ +; Options for the GCN port of the compiler. + +; Copyright (C) 2016-2019 Free Software Foundation, Inc. +; +; This file is part of GCC. +; +; GCC is free software; you can redistribute it and/or modify it under +; the terms of the GNU General Public License as published by the Free +; Software Foundation; either version 3, or (at your option) any later +; version. +; +; GCC is distributed in the hope that it will be useful, but WITHOUT ANY +; WARRANTY; without even the implied warranty of MERCHANTABILITY or +; FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +; for more details. +; +; You should have received a copy of the GNU General Public License +; along with GCC; see the file COPYING3. If not see +; . + +HeaderInclude +config/gcn/gcn-opts.h + +Enum +Name(gpu_type) Type(enum processor_type) +GCN GPU type to use: + +EnumValue +Enum(gpu_type) String(carrizo) Value(PROCESSOR_CARRIZO) + +EnumValue +Enum(gpu_type) String(fiji) Value(PROCESSOR_FIJI) + +EnumValue +Enum(gpu_type) String(gfx900) Value(PROCESSOR_VEGA) + +march= +Target RejectNegative Joined ToLower Enum(gpu_type) Var(gcn_arch) Init(PROCESSOR_CARRIZO) +Specify the name of the target GPU. + +mtune= +Target RejectNegative Joined ToLower Enum(gpu_type) Var(gcn_tune) Init(PROCESSOR_CARRIZO) +Specify the name of the target GPU. + +m32 +Target Report RejectNegative InverseMask(ABI64) +Generate code for a 32-bit ABI. + +m64 +Target Report RejectNegative Mask(ABI64) +Generate code for a 64-bit ABI. + +mgomp +Target Report RejectNegative +Enable OpenMP GPU offloading. + +bool flag_bypass_init_error = false + +mbypass-init-error +Target Report RejectNegative Var(flag_bypass_init_error) + +bool flag_worker_partitioning = false + +macc-experimental-workers +Target Report Var(flag_worker_partitioning) Init(1) + +int stack_size_opt = -1 + +mstack-size= +Target Report RejectNegative Joined UInteger Var(stack_size_opt) Init(-1) +-mstack-size= Set the private segment size per wave-front, in bytes. + +mlocal-symbol-id= +Target RejectNegative Report JoinedOrMissing Var(local_symbol_id) Init(0) + +Wopenacc-dims +Target Var(warn_openacc_dims) Warning +Warn about invalid OpenACC dimensions. diff --git a/gcc/config/gcn/t-gcn-hsa b/gcc/config/gcn/t-gcn-hsa new file mode 100644 index 00000000000..085ba429c9d --- /dev/null +++ b/gcc/config/gcn/t-gcn-hsa @@ -0,0 +1,52 @@ +# Copyright (C) 2016-2019 Free Software Foundation, Inc. +# +# This file is free software; you can redistribute it and/or modify it under +# the terms of the GNU General Public License as published by the Free +# Software Foundation; either version 3 of the License, or (at your option) +# any later version. +# +# This file is distributed in the hope that it will be useful, but WITHOUT +# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +# FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +# for more details. +# +# You should have received a copy of the GNU General Public License +# along with GCC; see the file COPYING3. If not see +# . + +GTM_H += $(HASH_TABLE_H) + +driver-gcn.o: $(srcdir)/config/gcn/driver-gcn.c + $(COMPILE) $< + $(POSTCOMPILE) + +CFLAGS-mkoffload.o += $(DRIVER_DEFINES) \ + -DGCC_INSTALL_NAME=\"$(GCC_INSTALL_NAME)\" +mkoffload.o: $(srcdir)/config/gcn/mkoffload.c + $(COMPILE) $< + $(POSTCOMPILE) +ALL_HOST_OBJS += mkoffload.o + +mkoffload$(exeext): mkoffload.o collect-utils.o libcommon-target.a \ + $(LIBIBERTY) $(LIBDEPS) + +$(LINKER) $(ALL_LINKERFLAGS) $(LDFLAGS) -o $@ \ + mkoffload.o collect-utils.o libcommon-target.a $(LIBIBERTY) $(LIBS) + +CFLAGS-gcn-run.o += -DVERSION_STRING=$(PKGVERSION_s) +COMPILE-gcn-run.o = $(filter-out -fno-rtti,$(COMPILE)) +gcn-run.o: $(srcdir)/config/gcn/gcn-run.c + $(COMPILE-gcn-run.o) -x c -std=gnu11 -Wno-error=pedantic $< + $(POSTCOMPILE) +ALL_HOST_OBJS += gcn-run.o + +gcn-run$(exeext): gcn-run.o + +$(LINKER) $(ALL_LINKERFLAGS) $(LDFLAGS) -o $@ $< -ldl + +MULTILIB_OPTIONS = march=gfx900 +MULTILIB_DIRNAMES = gcn5 + +PASSES_EXTRA += $(srcdir)/config/gcn/gcn-passes.def +gcn-tree.o: $(srcdir)/config/gcn/gcn-tree.c + $(COMPILE) $< + $(POSTCOMPILE) +ALL_HOST_OBJS += gcn-tree.o -- 2.30.2