brig/brigfrontend/brig-to-generic.cc

/* brig2tree.cc -- brig to gcc generic/gimple tree conversion
   Copyright (C) 2016-2018 Free Software Foundation, Inc.
   Contributed by Pekka Jaaskelainen <pekka.jaaskelainen@parmance.com>
   for General Processor Tech.

   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
   <http://www.gnu.org/licenses/>.  */

#include <cassert>
#include <iostream>
#include <iomanip>
#include <sstream>

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "target.h"
#include "function.h"
#include "brig-to-generic.h"
#include "stringpool.h"
#include "tree-iterator.h"
#include "toplev.h"
#include "gimplify.h"
#include "gimple-expr.h"
#include "print-tree.h"
#include "hsa-brig-format.h"
#include "stor-layout.h"
#include "diagnostic-core.h"
#include "brig-code-entry-handler.h"
#include "brig-machine.h"
#include "brig-util.h"
#include "phsa.h"
#include "tree-pretty-print.h"
#include "dumpfile.h"
#include "profile-count.h"
#include "tree-cfg.h"
#include "errors.h"
#include "fold-const.h"
#include "cgraph.h"
#include "dumpfile.h"
#include "tree-pretty-print.h"

extern int gccbrig_verbose;

tree brig_to_generic::s_fp16_type;
tree brig_to_generic::s_fp32_type;
tree brig_to_generic::s_fp64_type;

brig_to_generic::brig_to_generic ()
  : m_cf (NULL), m_analyzing (true), m_total_group_segment_usage (0),
    m_brig (NULL), m_next_private_offset (0)
{
  m_globals = NULL_TREE;

  /* Initialize the basic REAL types.
     This doesn't work straight away because most of the targets
     do not support fp16 natively.  Let's by default convert
     to fp32 and back before and after each instruction (handle it as
     a storage format only), and later add an optimization pass
     that removes the extra converts (in case of multiple fp16 ops
     in a row).  */
  s_fp16_type = make_node (REAL_TYPE);
  TYPE_PRECISION (s_fp16_type) = 16;
  TYPE_SIZE (s_fp16_type) = bitsize_int (16);
  TYPE_SIZE_UNIT (s_fp16_type) = size_int (2);
  SET_TYPE_ALIGN (s_fp16_type, 16);
  layout_type (s_fp16_type);

  s_fp32_type = gccbrig_tree_type_for_hsa_type (BRIG_TYPE_F32);
  s_fp64_type = gccbrig_tree_type_for_hsa_type (BRIG_TYPE_F64);

  /* TODO: (machine)query the preferred rounding mode that is set by
     the machine by default.  This can be redefined by each BRIG module
     header.  */
  m_default_float_rounding_mode = BRIG_ROUND_FLOAT_ZERO;

  m_dump_file = dump_begin (TDI_original, &m_dump_flags);
}

class unimplemented_entry_handler : public brig_code_entry_handler
{
public:
  unimplemented_entry_handler (brig_to_generic &parent)
    : brig_code_entry_handler (parent)
  {
  }

  size_t
  operator () (const BrigBase *base)
  {
    gcc_unreachable ();
    return base->byteCount;
  }
};

/* Handler for entries that can be (and are) safely skipped for the purposes
   of GENERIC generation.  */

class skipped_entry_handler : public brig_code_entry_handler
{
public:
  skipped_entry_handler (brig_to_generic &parent)
    : brig_code_entry_handler (parent)
  {
  }

  size_t
  operator () (const BrigBase *base)
  {
    return base->byteCount;
  }
};

class brig_reg_use_analyzer : public brig_code_entry_handler
{
public:
  brig_reg_use_analyzer (brig_to_generic &parent)
    : brig_code_entry_handler (parent)
  {
  }

  size_t
  operator () (const BrigBase *base)
  {
    const BrigInstBase *brig_inst = (const BrigInstBase *) base;
    analyze_operands (*brig_inst);
    return base->byteCount;
  }

};

/* Helper struct for pairing a BrigKind and a BrigCodeEntryHandler that
   should handle its data.  */

struct code_entry_handler_info
{
  BrigKind kind;
  brig_code_entry_handler *handler;
};


/* Finds the BRIG file sections in the currently processed file.  */

void
brig_to_generic::find_brig_sections ()
{
  m_data = m_code = m_operand = NULL;
  const BrigModuleHeader *mheader = (const BrigModuleHeader *) m_brig;

  /* Find the positions of the different sections.  */
  for (uint32_t sec = 0; sec < mheader->sectionCount; ++sec)
    {
      uint64_t offset
	= ((const uint64_t *) (m_brig + mheader->sectionIndex))[sec];

      const BrigSectionHeader *section_header
	= (const BrigSectionHeader *) (m_brig + offset);

      std::string name ((const char *) (&section_header->name),
			section_header->nameLength);

      if (sec == BRIG_SECTION_INDEX_DATA && name == "hsa_data")
	{
	  m_data = (const char *) section_header;
	  m_data_size = section_header->byteCount;
	}
      else if (sec == BRIG_SECTION_INDEX_CODE && name == "hsa_code")
	{
	  m_code = (const char *) section_header;
	  m_code_size = section_header->byteCount;
	}
      else if (sec == BRIG_SECTION_INDEX_OPERAND && name == "hsa_operand")
	{
	  m_operand = (const char *) section_header;
	  m_operand_size = section_header->byteCount;
	}
      else
	{
	  gcc_unreachable ();
	}
    }

  if (m_code == NULL)
    gcc_unreachable ();
  if (m_data == NULL)
    gcc_unreachable ();
  if (m_operand == NULL)
    gcc_unreachable ();

}

/* Does a first pass over the given BRIG to collect data needed for the
   actual parsing.  Currently this includes only collecting the
   group segment variable usage to support the experimental HSA PRM feature
   where group variables can be declared also in module and function scope
   (in addition to kernel scope).
*/

void
brig_to_generic::analyze (const char *brig_blob)
{
  const BrigModuleHeader *mheader = (const BrigModuleHeader *) brig_blob;

  if (strncmp (mheader->identification, "HSA BRIG", 8) != 0)
    fatal_error (UNKNOWN_LOCATION, PHSA_ERROR_PREFIX_INCOMPATIBLE_MODULE
		 "Unrecognized file format.");
  if (mheader->brigMajor != 1 || mheader->brigMinor != 0)
    fatal_error (UNKNOWN_LOCATION, PHSA_ERROR_PREFIX_INCOMPATIBLE_MODULE
		 "BRIG version not supported. BRIG 1.0 required.");

  m_brig = brig_blob;

  find_brig_sections ();

  brig_directive_variable_handler var_handler (*this);
  brig_directive_fbarrier_handler fbar_handler (*this);
  brig_directive_function_handler func_handler (*this);
  brig_reg_use_analyzer reg_use_analyzer (*this);

  /* Need this for grabbing the module names for mangling the
     group variable names.  */
  brig_directive_module_handler module_handler (*this);
  skipped_entry_handler skipped_handler (*this);

  const BrigSectionHeader *csection_header = (const BrigSectionHeader *) m_code;

  code_entry_handler_info handlers[]
    = {{BRIG_KIND_INST_BASIC, &reg_use_analyzer},
       {BRIG_KIND_INST_MOD, &reg_use_analyzer},
       {BRIG_KIND_INST_CMP, &reg_use_analyzer},
       {BRIG_KIND_INST_MEM, &reg_use_analyzer},
       {BRIG_KIND_INST_CVT, &reg_use_analyzer},
       {BRIG_KIND_INST_SEG_CVT, &reg_use_analyzer},
       {BRIG_KIND_INST_SEG, &reg_use_analyzer},
       {BRIG_KIND_INST_ADDR, &reg_use_analyzer},
       {BRIG_KIND_INST_SOURCE_TYPE, &reg_use_analyzer},
       {BRIG_KIND_INST_ATOMIC, &reg_use_analyzer},
       {BRIG_KIND_INST_SIGNAL, &reg_use_analyzer},
       {BRIG_KIND_INST_BR, &reg_use_analyzer},
       {BRIG_KIND_INST_LANE, &reg_use_analyzer},
       {BRIG_KIND_INST_QUEUE, &reg_use_analyzer},
       {BRIG_KIND_DIRECTIVE_VARIABLE, &var_handler},
       {BRIG_KIND_DIRECTIVE_FBARRIER, &fbar_handler},
       {BRIG_KIND_DIRECTIVE_KERNEL, &func_handler},
       {BRIG_KIND_DIRECTIVE_MODULE, &module_handler},
       {BRIG_KIND_DIRECTIVE_FUNCTION, &func_handler}};

  m_analyzing = true;
  for (size_t b = csection_header->headerByteCount; b < m_code_size;)
    {
      const BrigBase *entry = (const BrigBase *) (m_code + b);

      brig_code_entry_handler *handler = &skipped_handler;

      if (m_cf != NULL && b >= m_cf->m_brig_def->nextModuleEntry)
	{
	  /* The function definition ended.  We can just discard the place
	     holder function. */
	  m_total_group_segment_usage += m_cf->m_local_group_variables.size ();
	  delete m_cf;
	  m_cf = NULL;
	}

      /* Find a handler.  */
      for (size_t i = 0;
	   i < sizeof (handlers) / sizeof (code_entry_handler_info); ++i)
	{
	  if (handlers[i].kind == entry->kind)
	    handler = handlers[i].handler;
	}

      int bytes_processed = (*handler) (entry);
      if (bytes_processed == 0)
	fatal_error (UNKNOWN_LOCATION, PHSA_ERROR_PREFIX_CORRUPTED_MODULE
		     "Element with 0 bytes.");
      b += bytes_processed;
    }

  if (m_cf != NULL)
    {
      m_total_group_segment_usage += m_cf->m_local_group_variables.size ();
      delete m_cf;
      m_cf = NULL;
    }

  m_total_group_segment_usage += m_module_group_variables.size ();
  m_analyzing = false;
}

/* Parses the given BRIG blob.  */

void
brig_to_generic::parse (const char *brig_blob)
{
  m_brig = brig_blob;
  find_brig_sections ();

  brig_basic_inst_handler inst_handler (*this);
  brig_branch_inst_handler branch_inst_handler (*this);
  brig_cvt_inst_handler cvt_inst_handler (*this);
  brig_seg_inst_handler seg_inst_handler (*this);
  brig_copy_move_inst_handler copy_move_inst_handler (*this);
  brig_signal_inst_handler signal_inst_handler (*this);
  brig_atomic_inst_handler atomic_inst_handler (*this);
  brig_cmp_inst_handler cmp_inst_handler (*this);
  brig_mem_inst_handler mem_inst_handler (*this);
  brig_inst_mod_handler inst_mod_handler (*this);
  brig_directive_label_handler label_handler (*this);
  brig_directive_variable_handler var_handler (*this);
  brig_directive_fbarrier_handler fbar_handler (*this);
  brig_directive_comment_handler comment_handler (*this);
  brig_directive_function_handler func_handler (*this);
  brig_directive_control_handler control_handler (*this);
  brig_directive_arg_block_handler arg_block_handler (*this);
  brig_directive_module_handler module_handler (*this);
  brig_lane_inst_handler lane_inst_handler (*this);
  brig_queue_inst_handler queue_inst_handler (*this);
  skipped_entry_handler skipped_handler (*this);
  unimplemented_entry_handler unimplemented_handler (*this);

  struct code_entry_handler_info
  {
    BrigKind kind;
    brig_code_entry_handler *handler;
  };

  /* TODO: Convert to a hash table / map.  For now, put the more common
     entries to the top to keep the scan fast on average.  */
  code_entry_handler_info handlers[]
    = {{BRIG_KIND_INST_BASIC, &inst_handler},
       {BRIG_KIND_INST_CMP, &cmp_inst_handler},
       {BRIG_KIND_INST_MEM, &mem_inst_handler},
       {BRIG_KIND_INST_MOD, &inst_mod_handler},
       {BRIG_KIND_INST_CVT, &cvt_inst_handler},
       {BRIG_KIND_INST_SEG_CVT, &seg_inst_handler},
       {BRIG_KIND_INST_SEG, &seg_inst_handler},
       {BRIG_KIND_INST_ADDR, &copy_move_inst_handler},
       {BRIG_KIND_INST_SOURCE_TYPE, &copy_move_inst_handler},
       {BRIG_KIND_INST_ATOMIC, &atomic_inst_handler},
       {BRIG_KIND_INST_SIGNAL, &signal_inst_handler},
       {BRIG_KIND_INST_BR, &branch_inst_handler},
       {BRIG_KIND_INST_LANE, &lane_inst_handler},
       {BRIG_KIND_INST_QUEUE, &queue_inst_handler},
       /* Assuming fences are not needed.  FIXME: call builtins
	  when porting to a platform where they are.  */
       {BRIG_KIND_INST_MEM_FENCE, &skipped_handler},
       {BRIG_KIND_DIRECTIVE_LABEL, &label_handler},
       {BRIG_KIND_DIRECTIVE_VARIABLE, &var_handler},
       {BRIG_KIND_DIRECTIVE_ARG_BLOCK_START, &arg_block_handler},
       {BRIG_KIND_DIRECTIVE_ARG_BLOCK_END, &arg_block_handler},
       {BRIG_KIND_DIRECTIVE_FBARRIER, &fbar_handler},
       {BRIG_KIND_DIRECTIVE_COMMENT, &comment_handler},
       {BRIG_KIND_DIRECTIVE_KERNEL, &func_handler},
       {BRIG_KIND_DIRECTIVE_SIGNATURE, &func_handler},
       {BRIG_KIND_DIRECTIVE_FUNCTION, &func_handler},
       {BRIG_KIND_DIRECTIVE_INDIRECT_FUNCTION, &func_handler},
       {BRIG_KIND_DIRECTIVE_MODULE, &module_handler},
       /* Skipping debug locations for now as not needed for conformance.  */
       {BRIG_KIND_DIRECTIVE_LOC, &skipped_handler},
       /* There are no supported pragmas at this moment.  */
       {BRIG_KIND_DIRECTIVE_PRAGMA, &skipped_handler},
       {BRIG_KIND_DIRECTIVE_CONTROL, &control_handler},
       {BRIG_KIND_DIRECTIVE_EXTENSION, &skipped_handler},
       /* BRIG_KIND_NONE entries are valid anywhere.  They can be used
	  for patching BRIGs before finalization.  */
       {BRIG_KIND_NONE, &skipped_handler}};

  const BrigSectionHeader *csection_header = (const BrigSectionHeader *) m_code;

  for (size_t b = csection_header->headerByteCount; b < m_code_size;)
    {
      const BrigBase *entry = (const BrigBase *) (m_code + b);

      brig_code_entry_handler *handler = &unimplemented_handler;

      if (m_cf != NULL && b >= m_cf->m_brig_def->nextModuleEntry)
	finish_function (); /* The function definition ended.  */

      /* Find a handler.  */
      for (size_t i = 0;
	   i < sizeof (handlers) / sizeof (code_entry_handler_info); ++i)
	{
	  if (handlers[i].kind == entry->kind)
	    handler = handlers[i].handler;
	}
      b += (*handler) (entry);
    }

  finish_function ();
}

const BrigData *
brig_to_generic::get_brig_data_entry (size_t entry_offset) const
{
  return (const BrigData *) (m_data + entry_offset);
}

const BrigBase *
brig_to_generic::get_brig_operand_entry (size_t entry_offset) const
{
  return (const BrigBase *) (m_operand + entry_offset);
}

const BrigBase *
brig_to_generic::get_brig_code_entry (size_t entry_offset) const
{
  return (const BrigBase *) (m_code + entry_offset);
}

void
brig_to_generic::append_global (tree g)
{
  if (m_globals == NULL_TREE)
    {
      m_globals = g;
      return;
    }
  else
    {
      tree last = tree_last (m_globals);
      TREE_CHAIN (last) = g;
    }
}

tree
brig_to_generic::global_variable (const std::string &name) const
{
  label_index::const_iterator i = m_global_variables.find (name);
  if (i == m_global_variables.end ())
    return NULL_TREE;
  else
    return (*i).second;
}

/* Returns a function declaration with the given name.  Assumes it has been
   created previously via a DirectiveFunction or similar.  */

tree
brig_to_generic::function_decl (const std::string &name)
{
  label_index::const_iterator i = m_function_index.find (name);
  if (i == m_function_index.end ())
    return NULL_TREE;
  return (*i).second;
}

void
brig_to_generic::add_function_decl (const std::string &name, tree func_decl)
{
  m_function_index[name] = func_decl;
}

/* Adds a GENERIC global variable VAR_DECL with the given NAME to the
   current module.  If we have generated a host def var ptr (a place holder
   for variables that are defined by the HSA host code) for this global
   variable definition (because there was a declaration earlier which looked
   like it might have been a host defined variable), we now have
   to assign its address and make it private to allow the references to
   point to the defined variable instead.  */

void
brig_to_generic::add_global_variable (const std::string &name, tree var_decl)
{
  append_global (var_decl);
  m_global_variables[name] = var_decl;

  std::string host_def_var_name
    = std::string (PHSA_HOST_DEF_PTR_PREFIX) + name;
  tree host_def_var = global_variable (host_def_var_name);
  if (host_def_var == NULL_TREE)
    return;

  tree ptype = build_pointer_type (TREE_TYPE (var_decl));
  tree var_addr = build1 (ADDR_EXPR, ptype, var_decl);

  DECL_INITIAL (host_def_var) = var_addr;
  TREE_PUBLIC (host_def_var) = 0;
}

/* Adds an indirection pointer for a potential host-defined program scope
   variable declaration.  */

void
brig_to_generic::add_host_def_var_ptr (const std::string &name, tree var_decl)
{
  std::string var_name = std::string (PHSA_HOST_DEF_PTR_PREFIX) + name;

  tree name_identifier = get_identifier (var_name.c_str ());

  tree ptr_var = build_decl (UNKNOWN_LOCATION, VAR_DECL, name_identifier,
			     build_pointer_type (TREE_TYPE (var_decl)));
  DECL_EXTERNAL (ptr_var) = 0;
  DECL_ARTIFICIAL (ptr_var) = 0;

  TREE_PUBLIC (ptr_var) = 1;
  TREE_USED (ptr_var) = 1;
  TREE_ADDRESSABLE (ptr_var) = 1;
  TREE_STATIC (ptr_var) = 1;

  append_global (ptr_var);
  m_global_variables[var_name] = ptr_var;
}

/* Produce a "mangled name" for the given brig function or kernel.
   The mangling is used to make unique global symbol name in case of
   module scope functions.  Program scope functions are not mangled
   (except for dropping the leading &), which makes the functions
   directly visible for linking using the original function name.  */

std::string
brig_to_generic::get_mangled_name
(const BrigDirectiveExecutable *func) const
{
  /* Strip the leading &.  */
  std::string func_name = get_string (func->name).substr (1);
  if (func->linkage == BRIG_LINKAGE_MODULE)
    {
      /* Mangle the module scope function names with the module name and
	 make them public so they can be queried by the HSA runtime from
	 the produced binary.  Assume it's the currently processed function
	 we are always referring to.  */
      func_name = "gccbrig." + m_module_name + "." + func_name;
    }
  return func_name;
}

std::string
brig_to_generic::get_string (size_t entry_offset) const
{
  const BrigData *data_item = get_brig_data_entry (entry_offset);
  return std::string ((const char *) &data_item->bytes, data_item->byteCount);
}

/* Adapted from c-semantics.c.  */

tree
build_stmt (enum tree_code code, ...)
{
  tree ret;
  int length, i;
  va_list p;
  bool side_effects;

  /* This function cannot be used to construct variably-sized nodes.  */
  gcc_assert (TREE_CODE_CLASS (code) != tcc_vl_exp);

  va_start (p, code);

  ret = make_node (code);
  TREE_TYPE (ret) = void_type_node;
  length = TREE_CODE_LENGTH (code);

  /* TREE_SIDE_EFFECTS will already be set for statements with
     implicit side effects.  Here we make sure it is set for other
     expressions by checking whether the parameters have side
     effects.  */

  side_effects = false;
  for (i = 0; i < length; i++)
    {
      tree t = va_arg (p, tree);
      if (t && !TYPE_P (t))
	side_effects |= TREE_SIDE_EFFECTS (t);
      TREE_OPERAND (ret, i) = t;
    }

  TREE_SIDE_EFFECTS (ret) |= side_effects;

  va_end (p);
  return ret;
}

/* BRIG regs are untyped, but GENERIC is not.  We need to add implicit casts
   in case treating the operand with an instruction with a type different
   than the created reg var type in order to select correct instruction type
   later on.  This function creates the necessary reinterpret type cast from
   a source variable to the destination type.  In case no cast is needed to
   the same type, SOURCE is returned directly.

   In case of mismatched type sizes, casting:
   - to narrower type the upper bits are clipped and
   - to wider type the source value is zero extended.  */

tree
build_resize_convert_view (tree destination_type, tree source)
{

  gcc_assert (source && destination_type && TREE_TYPE (source) != NULL_TREE
	      && destination_type != NULL_TREE);

  tree source_type = TREE_TYPE (source);
  if (TREE_CODE (source) == CALL_EXPR)
    {
      tree func_decl = TREE_OPERAND (TREE_OPERAND (source, 1), 0);
      source_type = TREE_TYPE (TREE_TYPE (func_decl));
    }

  if (destination_type == source_type)
    return source;

  size_t src_size = int_size_in_bytes (source_type);
  size_t dst_size = int_size_in_bytes (destination_type);
  if (src_size == dst_size)
    return build1 (VIEW_CONVERT_EXPR, destination_type, source);
  else /* src_size != dst_size  */
    {
      /* The src_size can be smaller at least with f16 scalars which are
	 stored to 32b register variables.  First convert to an equivalent
	 size unsigned type, then extend to an unsigned type of the
	 target width, after which VIEW_CONVERT_EXPR can be used to
	 force to the target type.  */
      tree resized = convert (get_scalar_unsigned_int_type (destination_type),
			      build_reinterpret_to_uint (source));
      gcc_assert ((size_t)int_size_in_bytes (TREE_TYPE (resized)) == dst_size);
      return build_resize_convert_view (destination_type, resized);
    }
}

/* Reinterprets SOURCE as a scalar unsigned int with the size
   corresponding to the orignal.  */

tree build_reinterpret_to_uint (tree source)
{
  tree src_type = TREE_TYPE (source);
  if (INTEGRAL_TYPE_P (src_type) && TYPE_UNSIGNED (src_type))
    return source;
  tree dest_type = get_scalar_unsigned_int_type (src_type);
  return build1 (VIEW_CONVERT_EXPR, dest_type, source);
}

/* Returns the finished brig_function for the given generic FUNC_DECL,
   or NULL, if not found.  */

brig_function *
brig_to_generic::get_finished_function (tree func_decl)
{
  std::string func_name
    = identifier_to_locale (IDENTIFIER_POINTER (DECL_NAME (func_decl)));
  std::map<std::string, brig_function *>::iterator i
    = m_finished_functions.find (func_name);
  if (i != m_finished_functions.end ())
    return (*i).second;
  else
    return NULL;
}

/* Adds a group variable to a correct book keeping structure depending
   on its segment.  */

void
brig_to_generic::add_group_variable (const std::string &name, size_t size,
				     size_t alignment, bool function_scope)
{
  /* Module and function scope group region variables are an experimental
     feature.  We implement module scope group variables with a separate
     book keeping inside brig_to_generic which is populated in the 'analyze()'
     prepass.  This is to ensure we know the group segment offsets when
     processing the functions that might refer to them.  */
  if (!function_scope)
    {
      if (!m_module_group_variables.has_variable (name))
	m_module_group_variables.add (name, size, alignment);
      return;
    }

  if (!m_cf->m_local_group_variables.has_variable (name))
    m_cf->m_local_group_variables.add (name, size, alignment);
}

/* Finalizes the currently handled function.  Should be called before
   setting a new function.  */

void
brig_to_generic::finish_function ()
{
  if (m_cf == NULL || m_cf->m_func_decl == NULL_TREE)
    {
      /* It can be a finished func declaration fingerprint, in that case we
	 don't have m_func_decl.  */
      m_cf = NULL;
      return;
    }

  if (!m_cf->m_is_kernel)
    {
      tree bind_expr = m_cf->m_current_bind_expr;
      tree stmts = BIND_EXPR_BODY (bind_expr);
      m_cf->finish ();
      m_cf->emit_metadata (stmts);
      dump_function (m_dump_file, m_cf);
      gimplify_function_tree (m_cf->m_func_decl);
      cgraph_node::finalize_function (m_cf->m_func_decl, true);
    }
  else
    /* Emit the kernel only at the very end so we can analyze the total
       group and private memory usage.  */
    m_kernels.push_back (m_cf);

  pop_cfun ();

  m_finished_functions[m_cf->m_name] = m_cf;
  m_cf = NULL;
}

/* Initializes a new currently handled function.  */

void
brig_to_generic::start_function (tree f)
{
  if (DECL_STRUCT_FUNCTION (f) == NULL)
    push_struct_function (f);
  else
    push_cfun (DECL_STRUCT_FUNCTION (f));

  m_cf->m_func_decl = f;
}

/* Appends a new variable to the current kernel's private segment.  */

void
brig_to_generic::append_private_variable (const std::string &name,
					  size_t size, size_t alignment)
{
  /* We need to take care of two cases of alignment with private
     variables because of the layout where the same variable for
     each work-item is laid out in successive addresses.

     1) Ensure the first work-item's variable is in an aligned
     offset:  */
  size_t align_padding = m_next_private_offset % alignment == 0 ?
    0 : (alignment - m_next_private_offset % alignment);

  /* 2) Each successive per-work-item copy should be aligned.
     If the variable has wider alignment than size then we need
     to add extra padding to ensure it.  The padding must be
     included in the size to allow per-work-item offset computation
     to find their own aligned copy.  */

  size_t per_var_padding = size % alignment == 0 ?
    0 : (alignment - size % alignment);
  m_private_data_sizes[name] = size + per_var_padding;

  m_next_private_offset += align_padding;
  m_private_offsets[name] = m_next_private_offset;
  m_next_private_offset += size + per_var_padding;
}

size_t
brig_to_generic::private_variable_segment_offset
  (const std::string &name) const
{
  var_offset_table::const_iterator i = m_private_offsets.find (name);
  gcc_assert (i != m_private_offsets.end ());
  return (*i).second;
}

bool
brig_to_generic::has_private_variable (const std::string &name) const
{
  std::map<std::string, size_t>::const_iterator i
    = m_private_data_sizes.find (name);
  return i != m_private_data_sizes.end ();
}

size_t
brig_to_generic::private_variable_size (const std::string &name) const
{
  std::map<std::string, size_t>::const_iterator i
    = m_private_data_sizes.find (name);
  gcc_assert (i != m_private_data_sizes.end ());
  return (*i).second;
}


/* The size of private segment required by a single work-item executing
   the currently processed kernel.  */

size_t
brig_to_generic::private_segment_size () const
{
  return m_next_private_offset;
}

/* Cached builtins indexed by name.  */

typedef std::map<std::string, tree> builtin_index;
builtin_index builtin_cache_;

/* Build a call to a builtin function.  PDECL is the builtin function to
   call.  NARGS is the number of input arguments, RETTYPE the built-in
   functions return value type, and ... is the list of arguments passed to
   the call with type first, then the value.  */

tree
call_builtin (tree pdecl, int nargs, tree rettype, ...)
{
  if (rettype == error_mark_node)
    return error_mark_node;

  tree *types = new tree[nargs];
  tree *args = new tree[nargs];

  va_list ap;
  va_start (ap, rettype);
  for (int i = 0; i < nargs; ++i)
    {
      types[i] = va_arg (ap, tree);
      tree arg = va_arg (ap, tree);
      args[i] = build_resize_convert_view (types[i], arg);
      if (types[i] == error_mark_node || args[i] == error_mark_node)
	{
	  delete[] types;
	  delete[] args;
	  va_end (ap);
	  return error_mark_node;
	}
    }
  va_end (ap);

  tree fnptr = build_fold_addr_expr (pdecl);

  tree ret = build_call_array (rettype, fnptr, nargs, args);

  delete[] types;
  delete[] args;

  return ret;
}

/* Generate all global declarations.  Should be called after the last
   BRIG has been fed in.  */

void
brig_to_generic::write_globals ()
{
  /* Now that the whole BRIG module has been processed, build a launcher
     and a metadata section for each built kernel.  */
  for (size_t i = 0; i < m_kernels.size (); ++i)
    {
      brig_function *f = m_kernels[i];

      /* Finish kernels now that we know the call graphs and their barrier
	 usage.  */
      f->finish_kernel ();

      dump_function (m_dump_file, f);
      gimplify_function_tree (f->m_func_decl);
      cgraph_node::finalize_function (f->m_func_decl, true);

      f->m_descriptor.is_kernel = 1;
      /* TODO: analyze the kernel's actual private and group segment usage
	 using call graph.  Now the mem size is overly
	 pessimistic in case of multiple kernels in the same module.
      */
      f->m_descriptor.group_segment_size = m_total_group_segment_usage;
      f->m_descriptor.private_segment_size = private_segment_size ();

      /* The kernarg size is rounded up to a multiple of 16 according to
	 the PRM specs.  */
      f->m_descriptor.kernarg_segment_size = f->m_next_kernarg_offset;
      if (f->m_descriptor.kernarg_segment_size % 16 > 0)
	f->m_descriptor.kernarg_segment_size
	  += 16 - f->m_next_kernarg_offset % 16;
      f->m_descriptor.kernarg_max_align = f->m_kernarg_max_align;

      tree launcher = f->emit_launcher_and_metadata ();

      append_global (launcher);

      gimplify_function_tree (launcher);
      cgraph_node::finalize_function (launcher, true);
      pop_cfun ();
    }

  int no_globals = list_length (m_globals);
  tree *vec = new tree[no_globals];

  int i = 0;
  tree global = m_globals;
  while (global)
    {
      vec[i] = global;
      ++i;
      global = TREE_CHAIN (global);
    }

  wrapup_global_declarations (vec, no_globals);

  delete[] vec;

}

/* Returns an type with unsigned int elements corresponding to the
   size and element count of ORIGINAL_TYPE.  */

tree
get_unsigned_int_type (tree original_type)
{
  if (VECTOR_TYPE_P (original_type))
    {
      size_t esize
	= int_size_in_bytes (TREE_TYPE (original_type)) * BITS_PER_UNIT;
      poly_uint64 ecount = TYPE_VECTOR_SUBPARTS (original_type);
      return build_vector_type (build_nonstandard_integer_type (esize, true),
				ecount);
    }
  else
    return build_nonstandard_integer_type (int_size_in_bytes (original_type)
					   * BITS_PER_UNIT,
					   true);
}

/* Returns a type with unsigned int corresponding to the size
   ORIGINAL_TYPE.  */

tree
get_scalar_unsigned_int_type (tree original_type)
{
  return build_nonstandard_integer_type (int_size_in_bytes (original_type)
					 * BITS_PER_UNIT, true);
}

void
dump_function (FILE *dump_file, brig_function *f)
{
  /* Dump the BRIG-specific tree IR.  */
  if (dump_file)
    {
      fprintf (dump_file, "\n;; Function %s", f->m_name.c_str ());
      fprintf (dump_file, "\n;; enabled by -%s\n\n",
	       dump_flag_name (TDI_original));
      print_generic_decl (dump_file, f->m_func_decl, 0);
      print_generic_expr (dump_file, f->m_current_bind_expr, 0);
      fprintf (dump_file, "\n");
    }
}

/* Records use of the BRIG_REG as a TYPE in the current function.  */

void
brig_to_generic::add_reg_used_as_type (const BrigOperandRegister &brig_reg,
				       tree type)
{
  gcc_assert (m_cf);
  reg_use_info &info
    = m_fn_regs_use_index[m_cf->m_name][gccbrig_hsa_reg_id (brig_reg)];

  if (info.m_type_refs_lookup.count (type))
    info.m_type_refs[info.m_type_refs_lookup[type]].second++;
  else
    {
      info.m_type_refs.push_back (std::make_pair (type, 1));
      info.m_type_refs_lookup[type] = info.m_type_refs.size () - 1;
    }
}