/* Definitions of target machine for GNU compiler.
- Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007
- Free Software Foundation, Inc.
+ Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008,
+ 2009 Free Software Foundation, Inc.
Contributed by James E. Wilson <wilson@cygnus.com> and
David Mosberger <davidm@hpl.hp.com>.
#include "hashtab.h"
#include "langhooks.h"
#include "cfglayout.h"
-#include "tree-gimple.h"
+#include "gimple.h"
#include "intl.h"
#include "df.h"
#include "debug.h"
#include "params.h"
#include "dbgcnt.h"
#include "tm-constrs.h"
+#include "sel-sched.h"
/* This is used for communication between ASM_OUTPUT_LABEL and
ASM_OUTPUT_LABELREF. */
static void ia64_init_dfa_pre_cycle_insn (void);
static rtx ia64_dfa_pre_cycle_insn (void);
static int ia64_first_cycle_multipass_dfa_lookahead_guard (rtx);
-static bool ia64_first_cycle_multipass_dfa_lookahead_guard_spec (rtx);
+static bool ia64_first_cycle_multipass_dfa_lookahead_guard_spec (const_rtx);
static int ia64_dfa_new_cycle (FILE *, int, rtx, int, int, int *);
static void ia64_h_i_d_extended (void);
+static void * ia64_alloc_sched_context (void);
+static void ia64_init_sched_context (void *, bool);
+static void ia64_set_sched_context (void *);
+static void ia64_clear_sched_context (void *);
+static void ia64_free_sched_context (void *);
static int ia64_mode_to_int (enum machine_mode);
static void ia64_set_sched_flags (spec_info_t);
+static ds_t ia64_get_insn_spec_ds (rtx);
+static ds_t ia64_get_insn_checked_ds (rtx);
+static bool ia64_skip_rtx_p (const_rtx);
static int ia64_speculate_insn (rtx, ds_t, rtx *);
-static rtx ia64_gen_spec_insn (rtx, ds_t, int, bool, bool);
-static bool ia64_needs_block_p (rtx);
-static rtx ia64_gen_check (rtx, rtx, bool);
+static bool ia64_needs_block_p (int);
+static rtx ia64_gen_spec_check (rtx, rtx, ds_t);
static int ia64_spec_check_p (rtx);
static int ia64_spec_check_src_p (rtx);
static rtx gen_tls_get_addr (void);
static rtx gen_fr_spill_x (rtx, rtx, rtx);
static rtx gen_fr_restore_x (rtx, rtx, rtx);
-static enum machine_mode hfa_element_mode (tree, bool);
+static enum machine_mode hfa_element_mode (const_tree, bool);
static void ia64_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode,
tree, int *, int);
static int ia64_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
tree, bool);
static bool ia64_function_ok_for_sibcall (tree, tree);
-static bool ia64_return_in_memory (tree, tree);
-static bool ia64_rtx_costs (rtx, int, int, int *);
+static bool ia64_return_in_memory (const_tree, const_tree);
+static bool ia64_rtx_costs (rtx, int, int, int *, bool);
+static int ia64_unspec_may_trap_p (const_rtx, unsigned);
static void fix_range (const char *);
static bool ia64_handle_option (size_t, const char *, int);
static struct machine_function * ia64_init_machine_status (void);
static void final_emit_insn_group_barriers (FILE *);
static void emit_predicate_relation_info (void);
static void ia64_reorg (void);
-static bool ia64_in_small_data_p (tree);
+static bool ia64_in_small_data_p (const_tree);
static void process_epilogue (FILE *, rtx, bool, bool);
static int process_set (FILE *, rtx, rtx, bool, bool);
static void ia64_output_function_end_prologue (FILE *);
static int ia64_issue_rate (void);
-static int ia64_adjust_cost (rtx, rtx, rtx, int);
+static int ia64_adjust_cost_2 (rtx, int, rtx, int, dw_t);
static void ia64_sched_init (FILE *, int, int);
static void ia64_sched_init_global (FILE *, int, int);
static void ia64_sched_finish_global (FILE *, int);
static int get_template (state_t, int);
static rtx get_next_important_insn (rtx, rtx);
+static bool important_for_bundling_p (rtx);
static void bundling (FILE *, int, rtx, rtx);
static void ia64_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
ATTRIBUTE_UNUSED;
static void ia64_vms_init_libfuncs (void)
ATTRIBUTE_UNUSED;
+static void ia64_soft_fp_init_libfuncs (void)
+ ATTRIBUTE_UNUSED;
static tree ia64_handle_model_attribute (tree *, tree, tree, int, bool *);
static tree ia64_handle_version_id_attribute (tree *, tree, tree, int, bool *);
static void ia64_encode_section_info (tree, rtx, int);
static rtx ia64_struct_value_rtx (tree, int);
-static tree ia64_gimplify_va_arg (tree, tree, tree *, tree *);
+static tree ia64_gimplify_va_arg (tree, tree, gimple_seq *, gimple_seq *);
static bool ia64_scalar_mode_supported_p (enum machine_mode mode);
static bool ia64_vector_mode_supported_p (enum machine_mode mode);
static bool ia64_cannot_force_const_mem (rtx);
-static const char *ia64_mangle_type (tree);
-static const char *ia64_invalid_conversion (tree, tree);
-static const char *ia64_invalid_unary_op (int, tree);
-static const char *ia64_invalid_binary_op (int, tree, tree);
+static const char *ia64_mangle_type (const_tree);
+static const char *ia64_invalid_conversion (const_tree, const_tree);
+static const char *ia64_invalid_unary_op (int, const_tree);
+static const char *ia64_invalid_binary_op (int, const_tree, const_tree);
+static enum machine_mode ia64_c_mode_for_suffix (char);
\f
/* Table of valid machine attributes. */
static const struct attribute_spec ia64_attribute_table[] =
#undef TARGET_IN_SMALL_DATA_P
#define TARGET_IN_SMALL_DATA_P ia64_in_small_data_p
-#undef TARGET_SCHED_ADJUST_COST
-#define TARGET_SCHED_ADJUST_COST ia64_adjust_cost
+#undef TARGET_SCHED_ADJUST_COST_2
+#define TARGET_SCHED_ADJUST_COST_2 ia64_adjust_cost_2
#undef TARGET_SCHED_ISSUE_RATE
#define TARGET_SCHED_ISSUE_RATE ia64_issue_rate
#undef TARGET_SCHED_VARIABLE_ISSUE
#undef TARGET_SCHED_H_I_D_EXTENDED
#define TARGET_SCHED_H_I_D_EXTENDED ia64_h_i_d_extended
+#undef TARGET_SCHED_ALLOC_SCHED_CONTEXT
+#define TARGET_SCHED_ALLOC_SCHED_CONTEXT ia64_alloc_sched_context
+
+#undef TARGET_SCHED_INIT_SCHED_CONTEXT
+#define TARGET_SCHED_INIT_SCHED_CONTEXT ia64_init_sched_context
+
+#undef TARGET_SCHED_SET_SCHED_CONTEXT
+#define TARGET_SCHED_SET_SCHED_CONTEXT ia64_set_sched_context
+
+#undef TARGET_SCHED_CLEAR_SCHED_CONTEXT
+#define TARGET_SCHED_CLEAR_SCHED_CONTEXT ia64_clear_sched_context
+
+#undef TARGET_SCHED_FREE_SCHED_CONTEXT
+#define TARGET_SCHED_FREE_SCHED_CONTEXT ia64_free_sched_context
+
#undef TARGET_SCHED_SET_SCHED_FLAGS
#define TARGET_SCHED_SET_SCHED_FLAGS ia64_set_sched_flags
+#undef TARGET_SCHED_GET_INSN_SPEC_DS
+#define TARGET_SCHED_GET_INSN_SPEC_DS ia64_get_insn_spec_ds
+
+#undef TARGET_SCHED_GET_INSN_CHECKED_DS
+#define TARGET_SCHED_GET_INSN_CHECKED_DS ia64_get_insn_checked_ds
+
#undef TARGET_SCHED_SPECULATE_INSN
#define TARGET_SCHED_SPECULATE_INSN ia64_speculate_insn
#undef TARGET_SCHED_NEEDS_BLOCK_P
#define TARGET_SCHED_NEEDS_BLOCK_P ia64_needs_block_p
-#undef TARGET_SCHED_GEN_CHECK
-#define TARGET_SCHED_GEN_CHECK ia64_gen_check
+#undef TARGET_SCHED_GEN_SPEC_CHECK
+#define TARGET_SCHED_GEN_SPEC_CHECK ia64_gen_spec_check
#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD_SPEC
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD_SPEC\
ia64_first_cycle_multipass_dfa_lookahead_guard_spec
+#undef TARGET_SCHED_SKIP_RTX_P
+#define TARGET_SCHED_SKIP_RTX_P ia64_skip_rtx_p
+
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL ia64_function_ok_for_sibcall
#undef TARGET_ARG_PARTIAL_BYTES
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK ia64_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
-#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
+#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
#undef TARGET_ASM_FILE_START
#define TARGET_ASM_FILE_START ia64_file_start
#undef TARGET_RTX_COSTS
#define TARGET_RTX_COSTS ia64_rtx_costs
#undef TARGET_ADDRESS_COST
-#define TARGET_ADDRESS_COST hook_int_rtx_0
+#define TARGET_ADDRESS_COST hook_int_rtx_bool_0
+
+#undef TARGET_UNSPEC_MAY_TRAP_P
+#define TARGET_UNSPEC_MAY_TRAP_P ia64_unspec_may_trap_p
#undef TARGET_MACHINE_DEPENDENT_REORG
#define TARGET_MACHINE_DEPENDENT_REORG ia64_reorg
#undef TARGET_INVALID_BINARY_OP
#define TARGET_INVALID_BINARY_OP ia64_invalid_binary_op
+#undef TARGET_C_MODE_FOR_SUFFIX
+#define TARGET_C_MODE_FOR_SUFFIX ia64_c_mode_for_suffix
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
typedef enum
static bool
ia64_cannot_force_const_mem (rtx x)
{
+ if (GET_MODE (x) == RFmode)
+ return true;
return tls_symbolic_operand_type (x) != 0;
}
computation below are also more natural to compute as 64-bit quantities.
If we've been given an SImode destination register, change it. */
if (GET_MODE (dest) != Pmode)
- dest = gen_rtx_REG_offset (dest, Pmode, REGNO (dest), 0);
+ dest = gen_rtx_REG_offset (dest, Pmode, REGNO (dest),
+ byte_lowpart_offset (Pmode, GET_MODE (dest)));
if (TARGET_NO_PIC)
return false;
}
else
{
- p[0] = (((unsigned HOST_WIDE_INT) l[3]) << 32) + l[2];
- p[1] = (((unsigned HOST_WIDE_INT) l[1]) << 32) + l[0];
+ p[0] = (((unsigned HOST_WIDE_INT) l[1]) << 32) + l[0];
+ p[1] = (((unsigned HOST_WIDE_INT) l[3]) << 32) + l[2];
}
out[0] = GEN_INT (p[0]);
out[1] = GEN_INT (p[1]);
/* HPUX TFmode compare requires a library call to _U_Qfcmp, which takes a
magic number as its third argument, that indicates what to do.
The return value is an integer to be compared against zero. */
- else if (GET_MODE (op0) == TFmode)
+ else if (TARGET_HPUX && GET_MODE (op0) == TFmode)
{
enum qfcmp_magic {
QCMP_INV = 1, /* Raise FP_INVALID on SNaN as a side effect. */
new_reg = cmp_reg;
if (code == NOT)
{
- new_reg = expand_simple_unop (DImode, NOT, new_reg, NULL_RTX, true);
- code = AND;
+ new_reg = expand_simple_binop (DImode, AND, new_reg, val, NULL_RTX,
+ true, OPTAB_DIRECT);
+ new_reg = expand_simple_unop (DImode, code, new_reg, NULL_RTX, true);
}
- new_reg = expand_simple_binop (DImode, code, new_reg, val, NULL_RTX,
- true, OPTAB_DIRECT);
+ else
+ new_reg = expand_simple_binop (DImode, code, new_reg, val, NULL_RTX,
+ true, OPTAB_DIRECT);
if (mode != DImode)
new_reg = gen_lowpart (mode, new_reg);
if (emitted_frame_related_regs[r] != 0)
{
regno = emitted_frame_related_regs[r];
- if (regno >= LOC_REG (0) && regno < LOC_REG (80 - frame_pointer_needed))
+ if (regno >= LOC_REG (0) && regno < LOC_REG (80 - frame_pointer_needed)
+ && current_frame_info.n_local_regs < regno - LOC_REG (0) + 1)
current_frame_info.n_local_regs = regno - LOC_REG (0) + 1;
else if (current_function_is_leaf
&& regno >= GR_REG (1) && regno <= GR_REG (31))
/* If there is a frame pointer, then we can't use loc79, because
that is HARD_FRAME_POINTER_REGNUM. In particular, see the
reg_name switching code in ia64_expand_prologue. */
- if (regno < (80 - frame_pointer_needed))
- {
- current_frame_info.n_local_regs = regno + 1;
- return LOC_REG (0) + regno;
- }
+ while (regno < (80 - frame_pointer_needed))
+ if (! is_emitted (LOC_REG (regno++)))
+ {
+ current_frame_info.n_local_regs = regno;
+ return LOC_REG (regno - 1);
+ }
}
/* Failed to find a general register to spill to. Must use stack. */
int spilled_gr_p = 0;
int spilled_fr_p = 0;
unsigned int regno;
+ int min_regno;
+ int max_regno;
int i;
if (current_frame_info.initialized)
Likewise for -a profiling for the bb_init_func argument. For -ax
profiling, we need two output registers for the two bb_init_trace_func
arguments. */
- if (current_function_profile)
+ if (crtl->profile)
i = MAX (i, 1);
#endif
current_frame_info.n_output_regs = i;
/* Similarly for gp. Note that if we're calling setjmp, the stacked
registers are clobbered, so we fall back to the stack. */
current_frame_info.r[reg_save_gp]
- = (current_function_calls_setjmp ? 0 : find_gr_spill (reg_save_gp, 1));
+ = (cfun->calls_setjmp ? 0 : find_gr_spill (reg_save_gp, 1));
if (current_frame_info.r[reg_save_gp] == 0)
{
SET_HARD_REG_BIT (mask, GR_REG (1));
If we have already emitted code for any of those registers,
then it's already too late to change. */
- if (current_frame_info.r[reg_fp] != 0
- && current_frame_info.r[reg_save_b0] == current_frame_info.r[reg_fp] + 1
- && current_frame_info.r[reg_save_ar_pfs] == current_frame_info.r[reg_fp] + 2
- && emitted_frame_related_regs[reg_save_b0] == 0
- && emitted_frame_related_regs[reg_save_ar_pfs] == 0
- && emitted_frame_related_regs[reg_fp] == 0)
- {
- current_frame_info.r[reg_save_b0] = current_frame_info.r[reg_fp];
- current_frame_info.r[reg_save_ar_pfs] = current_frame_info.r[reg_fp] + 1;
- current_frame_info.r[reg_fp] = current_frame_info.r[reg_fp] + 2;
+ min_regno = MIN (current_frame_info.r[reg_fp],
+ MIN (current_frame_info.r[reg_save_b0],
+ current_frame_info.r[reg_save_ar_pfs]));
+ max_regno = MAX (current_frame_info.r[reg_fp],
+ MAX (current_frame_info.r[reg_save_b0],
+ current_frame_info.r[reg_save_ar_pfs]));
+ if (min_regno > 0
+ && min_regno + 2 == max_regno
+ && (current_frame_info.r[reg_fp] == min_regno + 1
+ || current_frame_info.r[reg_save_b0] == min_regno + 1
+ || current_frame_info.r[reg_save_ar_pfs] == min_regno + 1)
+ && (emitted_frame_related_regs[reg_save_b0] == 0
+ || emitted_frame_related_regs[reg_save_b0] == min_regno)
+ && (emitted_frame_related_regs[reg_save_ar_pfs] == 0
+ || emitted_frame_related_regs[reg_save_ar_pfs] == min_regno + 1)
+ && (emitted_frame_related_regs[reg_fp] == 0
+ || emitted_frame_related_regs[reg_fp] == min_regno + 2))
+ {
+ current_frame_info.r[reg_save_b0] = min_regno;
+ current_frame_info.r[reg_save_ar_pfs] = min_regno + 1;
+ current_frame_info.r[reg_fp] = min_regno + 2;
}
/* See if we need to store the predicate register block. */
the stack, then the FR save area will be unaligned. We round the
size of this area up to keep things 16 byte aligned. */
if (spilled_fr_p)
- pretend_args_size = IA64_STACK_ALIGN (current_function_pretend_args_size);
+ pretend_args_size = IA64_STACK_ALIGN (crtl->args.pretend_args_size);
else
- pretend_args_size = current_function_pretend_args_size;
+ pretend_args_size = crtl->args.pretend_args_size;
total_size = (spill_size + extra_spill_size + size + pretend_args_size
- + current_function_outgoing_args_size);
+ + crtl->outgoing_args_size);
total_size = IA64_STACK_ALIGN (total_size);
/* We always use the 16-byte scratch area provided by the caller, but
offset = -current_frame_info.total_size;
else
offset = -(current_frame_info.total_size
- - current_function_outgoing_args_size - 16);
+ - crtl->outgoing_args_size - 16);
break;
case STACK_POINTER_REGNUM:
if (current_function_is_leaf)
offset = 0;
else
- offset = 16 + current_function_outgoing_args_size;
+ offset = 16 + crtl->outgoing_args_size;
break;
default:
switch (to)
{
case HARD_FRAME_POINTER_REGNUM:
- offset = 16 - current_function_pretend_args_size;
+ offset = 16 - crtl->args.pretend_args_size;
break;
case STACK_POINTER_REGNUM:
offset = (current_frame_info.total_size
- + 16 - current_function_pretend_args_size);
+ + 16 - crtl->args.pretend_args_size);
break;
default:
/* We don't need an alloc instruction if we've used no outputs or locals. */
if (current_frame_info.n_local_regs == 0
&& current_frame_info.n_output_regs == 0
- && current_frame_info.n_input_regs <= current_function_args_info.int_regs
+ && current_frame_info.n_input_regs <= crtl->args.info.int_regs
&& !TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
{
/* If there is no alloc, but there are input registers used, then we
finish_spill_pointers ();
- if (current_frame_info.total_size || cfun->machine->ia64_eh_epilogue_sp)
+ if (current_frame_info.total_size
+ || cfun->machine->ia64_eh_epilogue_sp
+ || frame_pointer_needed)
{
/* ??? At this point we must generate a magic insn that appears to
modify the spill iterators, the stack pointer, and the frame
aggregates are excluded because our parallels crash the middle-end. */
static enum machine_mode
-hfa_element_mode (tree type, bool nested)
+hfa_element_mode (const_tree type, bool nested)
{
enum machine_mode element_mode = VOIDmode;
enum machine_mode mode;
return false;
/* We must always return with our current GP. This means we can
- only sibcall to functions defined in the current module. */
- return decl && (*targetm.binds_local_p) (decl);
+ only sibcall to functions defined in the current module unless
+ TARGET_CONST_GP is set to true. */
+ return (decl && (*targetm.binds_local_p) (decl)) || TARGET_CONST_GP;
}
\f
/* Implement va_arg. */
static tree
-ia64_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
+ia64_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
+ gimple_seq *post_p)
{
/* Variable sized types are passed by reference. */
if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t,
size_int (-2 * UNITS_PER_WORD));
t = fold_convert (TREE_TYPE (valist), t);
- t = build2 (GIMPLE_MODIFY_STMT, TREE_TYPE (valist), valist, t);
- gimplify_and_add (t, pre_p);
+ gimplify_assign (unshare_expr (valist), t, pre_p);
}
return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
in a register. */
static bool
-ia64_return_in_memory (tree valtype, tree fntype ATTRIBUTE_UNUSED)
+ia64_return_in_memory (const_tree valtype, const_tree fntype ATTRIBUTE_UNUSED)
{
enum machine_mode mode;
enum machine_mode hfa_mode;
/* Return rtx for register that holds the function return value. */
rtx
-ia64_function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
+ia64_function_value (const_tree valtype, const_tree func ATTRIBUTE_UNUSED)
{
enum machine_mode mode;
enum machine_mode hfa_mode;
/* ??? This is incomplete. */
static bool
-ia64_rtx_costs (rtx x, int code, int outer_code, int *total)
+ia64_rtx_costs (rtx x, int code, int outer_code, int *total,
+ bool speed ATTRIBUTE_UNUSED)
{
switch (code)
{
return 2;
}
-/* Implement PREFERRED_RELOAD_CLASS. Place additional restrictions on CLASS
+/* Implement PREFERRED_RELOAD_CLASS. Place additional restrictions on RCLASS
to use when copying X into that class. */
enum reg_class
-ia64_preferred_reload_class (rtx x, enum reg_class class)
+ia64_preferred_reload_class (rtx x, enum reg_class rclass)
{
- switch (class)
+ switch (rclass)
{
case FR_REGS:
case FP_REGS:
break;
}
- return class;
+ return rclass;
}
/* This function returns the register class required for a secondary
- register when copying between one of the registers in CLASS, and X,
+ register when copying between one of the registers in RCLASS, and X,
using MODE. A return value of NO_REGS means that no secondary register
is required. */
enum reg_class
-ia64_secondary_reload_class (enum reg_class class,
+ia64_secondary_reload_class (enum reg_class rclass,
enum machine_mode mode ATTRIBUTE_UNUSED, rtx x)
{
int regno = -1;
if (GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
regno = true_regnum (x);
- switch (class)
+ switch (rclass)
{
case BR_REGS:
case AR_M_REGS:
/* ??? This happens if we cse/gcse a BImode value across a call,
and the function has a nonlocal goto. This is because global
does not allocate call crossing pseudos to hard registers when
- current_function_has_nonlocal_goto is true. This is relatively
+ crtl->has_nonlocal_goto is true. This is relatively
common for C++ programs that use exceptions. To reproduce,
return NO_REGS and compile libstdc++. */
if (GET_CODE (x) == MEM)
}
\f
+/* Implement targetm.unspec_may_trap_p hook. */
+static int
+ia64_unspec_may_trap_p (const_rtx x, unsigned flags)
+{
+ if (GET_CODE (x) == UNSPEC)
+ {
+ switch (XINT (x, 1))
+ {
+ case UNSPEC_LDA:
+ case UNSPEC_LDS:
+ case UNSPEC_LDSA:
+ case UNSPEC_LDCCLR:
+ case UNSPEC_CHKACLR:
+ case UNSPEC_CHKS:
+ /* These unspecs are just wrappers. */
+ return may_trap_p_1 (XVECEXP (x, 0, 0), flags);
+ }
+ }
+
+ return default_unspec_may_trap_p (x, flags);
+}
+
+\f
/* Parse the -mfixed-range= option string. */
static void
static bool
ia64_handle_option (size_t code, const char *arg, int value)
{
+ static bool warned_itanium1_deprecated;
+
switch (code)
{
case OPT_mfixed_range_:
if (!strcmp (arg, processor_alias_table[i].name))
{
ia64_tune = processor_alias_table[i].processor;
+ if (ia64_tune == PROCESSOR_ITANIUM
+ && ! warned_itanium1_deprecated)
+ {
+ inform (0,
+ "value %<%s%> for -mtune= switch is deprecated",
+ arg);
+ inform (0, "GCC 4.4 is the last release with "
+ "Itanium1 tuning support");
+ warned_itanium1_deprecated = true;
+ }
break;
}
if (i == pta_size)
ia64_flag_schedule_insns2 = flag_schedule_insns_after_reload;
flag_schedule_insns_after_reload = 0;
+ if (optimize >= 3
+ && ! sel_sched_switch_set)
+ {
+ flag_selective_scheduling2 = 1;
+ flag_sel_sched_pipelining = 1;
+ }
+ if (mflag_sched_control_spec == 2)
+ {
+ /* Control speculation is on by default for the selective scheduler,
+ but not for the Haifa scheduler. */
+ mflag_sched_control_spec = flag_selective_scheduling2 ? 1 : 0;
+ }
+ if (flag_sel_sched_pipelining && flag_auto_inc_dec)
+ {
+ /* FIXME: remove this when we'd implement breaking autoinsns as
+ a transformation. */
+ flag_auto_inc_dec = 0;
+ }
+
ia64_section_threshold = g_switch_set ? g_switch_value : IA64_DEFAULT_GVALUE;
init_machine_status = ia64_init_machine_status;
+
+ if (align_functions <= 0)
+ align_functions = 64;
+ if (align_loops <= 0)
+ align_loops = 32;
}
/* Initialize the record of emitted frame related registers. */
static struct machine_function *
ia64_init_machine_status (void)
{
- return ggc_alloc_cleared (sizeof (struct machine_function));
+ return GGC_CNEW (struct machine_function);
}
\f
static enum attr_itanium_class ia64_safe_itanium_class (rtx);
If a predicate register is written by an AND.ORCM we set WRITTEN_BY_AND
to true; if it was written by an OR.ANDCM we set WRITTEN_BY_OR to true. */
+#if GCC_VERSION >= 4000
+#define RWS_FIELD_TYPE __extension__ unsigned short
+#else
+#define RWS_FIELD_TYPE unsigned int
+#endif
struct reg_write_state
{
- unsigned int write_count : 2;
- unsigned int first_pred : 16;
- unsigned int written_by_fp : 1;
- unsigned int written_by_and : 1;
- unsigned int written_by_or : 1;
+ RWS_FIELD_TYPE write_count : 2;
+ RWS_FIELD_TYPE first_pred : 10;
+ RWS_FIELD_TYPE written_by_fp : 1;
+ RWS_FIELD_TYPE written_by_and : 1;
+ RWS_FIELD_TYPE written_by_or : 1;
};
/* Cumulative info for the current instruction group. */
struct reg_write_state rws_sum[NUM_REGS];
-/* Info for the current instruction. This gets copied to rws_sum after a
- stop bit is emitted. */
-struct reg_write_state rws_insn[NUM_REGS];
+#ifdef ENABLE_CHECKING
+/* Bitmap whether a register has been written in the current insn. */
+HARD_REG_ELT_TYPE rws_insn[(NUM_REGS + HOST_BITS_PER_WIDEST_FAST_INT - 1)
+ / HOST_BITS_PER_WIDEST_FAST_INT];
+
+static inline void
+rws_insn_set (int regno)
+{
+ gcc_assert (!TEST_HARD_REG_BIT (rws_insn, regno));
+ SET_HARD_REG_BIT (rws_insn, regno);
+}
+
+static inline int
+rws_insn_test (int regno)
+{
+ return TEST_HARD_REG_BIT (rws_insn, regno);
+}
+#else
+/* When not checking, track just REG_AR_CFM and REG_VOLATILE. */
+unsigned char rws_insn[2];
+
+static inline void
+rws_insn_set (int regno)
+{
+ if (regno == REG_AR_CFM)
+ rws_insn[0] = 1;
+ else if (regno == REG_VOLATILE)
+ rws_insn[1] = 1;
+}
+
+static inline int
+rws_insn_test (int regno)
+{
+ if (regno == REG_AR_CFM)
+ return rws_insn[0];
+ if (regno == REG_VOLATILE)
+ return rws_insn[1];
+ return 0;
+}
+#endif
/* Indicates whether this is the first instruction after a stop bit,
in which case we don't need another stop bit. Without this,
unsigned int is_sibcall : 1; /* Is this a sibling or normal call? */
};
-static void rws_update (struct reg_write_state *, int, struct reg_flags, int);
+static void rws_update (int, struct reg_flags, int);
static int rws_access_regno (int, struct reg_flags, int);
static int rws_access_reg (rtx, struct reg_flags, int);
static void update_set_flags (rtx, struct reg_flags *);
static void init_insn_group_barriers (void);
static int group_barrier_needed (rtx);
static int safe_group_barrier_needed (rtx);
+static int in_safe_group_barrier;
/* Update *RWS for REGNO, which is being written by the current instruction,
with predicate PRED, and associated register flags in FLAGS. */
static void
-rws_update (struct reg_write_state *rws, int regno, struct reg_flags flags, int pred)
+rws_update (int regno, struct reg_flags flags, int pred)
{
if (pred)
- rws[regno].write_count++;
+ rws_sum[regno].write_count++;
else
- rws[regno].write_count = 2;
- rws[regno].written_by_fp |= flags.is_fp;
+ rws_sum[regno].write_count = 2;
+ rws_sum[regno].written_by_fp |= flags.is_fp;
/* ??? Not tracking and/or across differing predicates. */
- rws[regno].written_by_and = flags.is_and;
- rws[regno].written_by_or = flags.is_or;
- rws[regno].first_pred = pred;
+ rws_sum[regno].written_by_and = flags.is_and;
+ rws_sum[regno].written_by_or = flags.is_or;
+ rws_sum[regno].first_pred = pred;
}
/* Handle an access to register REGNO of type FLAGS using predicate register
- PRED. Update rws_insn and rws_sum arrays. Return 1 if this access creates
+ PRED. Update rws_sum array. Return 1 if this access creates
a dependency with an earlier instruction in the same group. */
static int
{
int write_count;
- /* One insn writes same reg multiple times? */
- gcc_assert (!rws_insn[regno].write_count);
-
- /* Update info for current instruction. */
- rws_update (rws_insn, regno, flags, pred);
+ rws_insn_set (regno);
write_count = rws_sum[regno].write_count;
switch (write_count)
{
case 0:
/* The register has not been written yet. */
- rws_update (rws_sum, regno, flags, pred);
+ if (!in_safe_group_barrier)
+ rws_update (regno, flags, pred);
break;
case 1:
;
else if ((rws_sum[regno].first_pred ^ 1) != pred)
need_barrier = 1;
- rws_update (rws_sum, regno, flags, pred);
+ if (!in_safe_group_barrier)
+ rws_update (regno, flags, pred);
break;
case 2:
;
else
need_barrier = 1;
- rws_sum[regno].written_by_and = flags.is_and;
- rws_sum[regno].written_by_or = flags.is_or;
+ if (!in_safe_group_barrier)
+ {
+ rws_sum[regno].written_by_and = flags.is_and;
+ rws_sum[regno].written_by_or = flags.is_or;
+ }
break;
default:
/* Avoid multiple register writes, in case this is a pattern with
multiple CALL rtx. This avoids a failure in rws_access_reg. */
- if (! flags.is_sibcall && ! rws_insn[REG_AR_CFM].write_count)
+ if (! flags.is_sibcall && ! rws_insn_test (REG_AR_CFM))
{
new_flags.is_write = 1;
need_barrier |= rws_access_regno (REG_RP, new_flags, pred);
{
/* Avoid writing the register multiple times if we have multiple
asm outputs. This avoids a failure in rws_access_reg. */
- if (! rws_insn[REG_VOLATILE].write_count)
+ if (! rws_insn_test (REG_VOLATILE))
{
new_flags.is_write = 1;
rws_access_regno (REG_VOLATILE, new_flags, pred);
case UNSPEC_SETF_EXP:
case UNSPEC_ADDP4:
case UNSPEC_FR_SQRT_RECIP_APPROX:
+ case UNSPEC_FR_SQRT_RECIP_APPROX_RES:
case UNSPEC_LDA:
case UNSPEC_LDS:
+ case UNSPEC_LDS_A:
case UNSPEC_LDSA:
case UNSPEC_CHKACLR:
case UNSPEC_CHKS:
case UNSPEC_FR_RECIP_APPROX:
case UNSPEC_SHRP:
case UNSPEC_COPYSIGN:
+ case UNSPEC_FR_RECIP_APPROX_RES:
need_barrier = rtx_needs_barrier (XVECEXP (x, 0, 0), flags, pred);
need_barrier |= rtx_needs_barrier (XVECEXP (x, 0, 1), flags, pred);
break;
asm. */
if (! need_barrier)
need_barrier = rws_access_regno (REG_VOLATILE, flags, 0);
+
break;
default:
static int
safe_group_barrier_needed (rtx insn)
{
- struct reg_write_state rws_saved[NUM_REGS];
int saved_first_instruction;
int t;
- memcpy (rws_saved, rws_sum, NUM_REGS * sizeof *rws_saved);
saved_first_instruction = first_instruction;
+ in_safe_group_barrier = 1;
t = group_barrier_needed (insn);
- memcpy (rws_sum, rws_saved, NUM_REGS * sizeof *rws_saved);
first_instruction = saved_first_instruction;
+ in_safe_group_barrier = 0;
return t;
}
static rtx last_scheduled_insn;
-/* The following variable value is size of the DFA state. */
-
-static size_t dfa_state_size;
-
/* The following variable value is pointer to a DFA state used as
temporary variable. */
static char *stops_p = NULL;
-/* The following array element values are ZERO for non-speculative
- instructions and hold corresponding speculation check number for
- speculative instructions. */
-static int *spec_check_no = NULL;
-
-/* Size of spec_check_no array. */
-static int max_uid = 0;
-
/* The following variable is used to set up the mentioned above array. */
static int stop_before_p = 0;
/* The following variable value is number of data speculations in progress. */
static int pending_data_specs = 0;
+/* Number of memory references on current and three future processor cycles. */
+static char mem_ops_in_group[4];
+
+/* Number of current processor cycle (from scheduler's point of view). */
+static int current_cycle;
+
static rtx ia64_single_set (rtx);
static void ia64_emit_insn_before (rtx, rtx);
return ret;
}
-/* Adjust the cost of a scheduling dependency. Return the new cost of
- a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
-
+/* Adjust the cost of a scheduling dependency.
+ Return the new cost of a dependency of type DEP_TYPE or INSN on DEP_INSN.
+ COST is the current cost, DW is dependency weakness. */
static int
-ia64_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
+ia64_adjust_cost_2 (rtx insn, int dep_type1, rtx dep_insn, int cost, dw_t dw)
{
+ enum reg_note dep_type = (enum reg_note) dep_type1;
enum attr_itanium_class dep_class;
enum attr_itanium_class insn_class;
- if (REG_NOTE_KIND (link) != REG_DEP_OUTPUT)
- return cost;
-
insn_class = ia64_safe_itanium_class (insn);
dep_class = ia64_safe_itanium_class (dep_insn);
+
+ /* Treat true memory dependencies separately. Ignore apparent true
+ dependence between store and call (call has a MEM inside a SYMBOL_REF). */
+ if (dep_type == REG_DEP_TRUE
+ && (dep_class == ITANIUM_CLASS_ST || dep_class == ITANIUM_CLASS_STF)
+ && (insn_class == ITANIUM_CLASS_BR || insn_class == ITANIUM_CLASS_SCALL))
+ return 0;
+
+ if (dw == MIN_DEP_WEAK)
+ /* Store and load are likely to alias, use higher cost to avoid stall. */
+ return PARAM_VALUE (PARAM_SCHED_MEM_TRUE_DEP_COST);
+ else if (dw > MIN_DEP_WEAK)
+ {
+ /* Store and load are less likely to alias. */
+ if (mflag_sched_fp_mem_deps_zero_cost && dep_class == ITANIUM_CLASS_STF)
+ /* Assume there will be no cache conflict for floating-point data.
+ For integer data, L1 conflict penalty is huge (17 cycles), so we
+ never assume it will not cause a conflict. */
+ return 0;
+ else
+ return cost;
+ }
+
+ if (dep_type != REG_DEP_OUTPUT)
+ return cost;
+
if (dep_class == ITANIUM_CLASS_ST || dep_class == ITANIUM_CLASS_STF
|| insn_class == ITANIUM_CLASS_ST || insn_class == ITANIUM_CLASS_STF)
return 0;
if (INSN_P (insn)
&& ia64_safe_itanium_class (insn) == ITANIUM_CLASS_IALU)
{
- dep_link_t link;
+ sd_iterator_def sd_it;
+ dep_t dep;
+ bool has_mem_op_consumer_p = false;
- FOR_EACH_DEP_LINK (link, INSN_FORW_DEPS (insn))
+ FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
{
enum attr_itanium_class c;
- if (DEP_LINK_KIND (link) != REG_DEP_TRUE)
+ if (DEP_TYPE (dep) != REG_DEP_TRUE)
continue;
- next = DEP_LINK_CON (link);
+ next = DEP_CON (dep);
c = ia64_safe_itanium_class (next);
if ((c == ITANIUM_CLASS_ST
|| c == ITANIUM_CLASS_STF)
&& ia64_st_address_bypass_p (insn, next))
- break;
+ {
+ has_mem_op_consumer_p = true;
+ break;
+ }
else if ((c == ITANIUM_CLASS_LD
|| c == ITANIUM_CLASS_FLD
|| c == ITANIUM_CLASS_FLDP)
&& ia64_ld_address_bypass_p (insn, next))
- break;
+ {
+ has_mem_op_consumer_p = true;
+ break;
+ }
}
- insn->call = link != 0;
+
+ insn->call = has_mem_op_consumer_p;
}
}
#ifdef ENABLE_CHECKING
rtx insn;
- if (reload_completed)
+ if (!sel_sched_p () && reload_completed)
for (insn = NEXT_INSN (current_sched_info->prev_head);
insn != current_sched_info->next_tail;
insn = NEXT_INSN (insn))
#endif
last_scheduled_insn = NULL_RTX;
init_insn_group_barriers ();
+
+ current_cycle = 0;
+ memset (mem_ops_in_group, 0, sizeof (mem_ops_in_group));
}
/* We're beginning a scheduling pass. Check assertion. */
int sched_verbose ATTRIBUTE_UNUSED,
int max_ready ATTRIBUTE_UNUSED)
{
- gcc_assert (!pending_data_specs);
+ gcc_assert (pending_data_specs == 0);
}
/* Scheduling pass is now finished. Free/reset static variable. */
ia64_sched_finish_global (FILE *dump ATTRIBUTE_UNUSED,
int sched_verbose ATTRIBUTE_UNUSED)
{
- free (spec_check_no);
- spec_check_no = 0;
- max_uid = 0;
+ gcc_assert (pending_data_specs == 0);
+}
+
+/* Return TRUE if INSN is a load (either normal or speculative, but not a
+ speculation check), FALSE otherwise. */
+static bool
+is_load_p (rtx insn)
+{
+ enum attr_itanium_class insn_class = ia64_safe_itanium_class (insn);
+
+ return
+ ((insn_class == ITANIUM_CLASS_LD || insn_class == ITANIUM_CLASS_FLD)
+ && get_attr_check_load (insn) == CHECK_LOAD_NO);
+}
+
+/* If INSN is a memory reference, memoize it in MEM_OPS_IN_GROUP global array
+ (taking account for 3-cycle cache reference postponing for stores: Intel
+ Itanium 2 Reference Manual for Software Development and Optimization,
+ 6.7.3.1). */
+static void
+record_memory_reference (rtx insn)
+{
+ enum attr_itanium_class insn_class = ia64_safe_itanium_class (insn);
+
+ switch (insn_class) {
+ case ITANIUM_CLASS_FLD:
+ case ITANIUM_CLASS_LD:
+ mem_ops_in_group[current_cycle % 4]++;
+ break;
+ case ITANIUM_CLASS_STF:
+ case ITANIUM_CLASS_ST:
+ mem_ops_in_group[(current_cycle + 3) % 4]++;
+ break;
+ default:;
+ }
}
/* We are about to being issuing insns for this clock cycle.
static int
ia64_dfa_sched_reorder (FILE *dump, int sched_verbose, rtx *ready,
- int *pn_ready, int clock_var ATTRIBUTE_UNUSED,
+ int *pn_ready, int clock_var,
int reorder_type)
{
int n_asms;
ready += deleted;
}
+ current_cycle = clock_var;
+ if (reload_completed && mem_ops_in_group[clock_var % 4] >= ia64_max_memory_insns)
+ {
+ int moved = 0;
+
+ insnp = e_ready;
+ /* Move down loads/stores, preserving relative order. */
+ while (insnp-- > ready + moved)
+ while (insnp >= ready + moved)
+ {
+ rtx insn = *insnp;
+ if (! is_load_p (insn))
+ break;
+ memmove (ready + 1, ready, (insnp - ready) * sizeof (rtx));
+ *ready = insn;
+ moved++;
+ }
+ n_ready -= moved;
+ ready += moved;
+ }
+
return 1;
}
rtx insn ATTRIBUTE_UNUSED,
int can_issue_more ATTRIBUTE_UNUSED)
{
- if (current_sched_info->flags & DO_SPECULATION)
+ if (sched_deps_info->generate_spec_deps && !sel_sched_p ())
/* Modulo scheduling does not extend h_i_d when emitting
- new instructions. Deal with it. */
+ new instructions. Don't use h_i_d, if we don't have to. */
{
if (DONE_SPEC (insn) & BEGIN_DATA)
pending_data_specs++;
init_insn_group_barriers ();
stops_p [INSN_UID (insn)] = stop_before_p;
stop_before_p = 0;
+
+ record_memory_reference (insn);
}
return 1;
}
static int
ia64_first_cycle_multipass_dfa_lookahead_guard (rtx insn)
{
- gcc_assert (insn && INSN_P (insn));
+ gcc_assert (insn && INSN_P (insn));
return ((!reload_completed
|| !safe_group_barrier_needed (insn))
- && ia64_first_cycle_multipass_dfa_lookahead_guard_spec (insn));
+ && ia64_first_cycle_multipass_dfa_lookahead_guard_spec (insn)
+ && (!mflag_sched_mem_insns_hard_limit
+ || !is_load_p (insn)
+ || mem_ops_in_group[current_cycle % 4] < ia64_max_memory_insns));
}
/* We are choosing insn from the ready queue. Return nonzero if INSN
can be chosen. */
static bool
-ia64_first_cycle_multipass_dfa_lookahead_guard_spec (rtx insn)
+ia64_first_cycle_multipass_dfa_lookahead_guard_spec (const_rtx insn)
{
gcc_assert (insn && INSN_P (insn));
/* Size of ALAT is 32. As far as we perform conservative data speculation,
static rtx dfa_pre_cycle_insn;
+/* Returns 1 when a meaningful insn was scheduled between the last group
+ barrier and LAST. */
+static int
+scheduled_good_insn (rtx last)
+{
+ if (last && recog_memoized (last) >= 0)
+ return 1;
+
+ for ( ;
+ last != NULL && !NOTE_INSN_BASIC_BLOCK_P (last)
+ && !stops_p[INSN_UID (last)];
+ last = PREV_INSN (last))
+ /* We could hit a NOTE_INSN_DELETED here which is actually outside
+ the ebb we're scheduling. */
+ if (INSN_P (last) && recog_memoized (last) >= 0)
+ return 1;
+
+ return 0;
+}
+
/* We are about to being issuing INSN. Return nonzero if we cannot
issue it on given cycle CLOCK and return zero if we should not sort
the ready queue on the next clock start. */
int setup_clocks_p = FALSE;
gcc_assert (insn && INSN_P (insn));
- if ((reload_completed && safe_group_barrier_needed (insn))
+ /* When a group barrier is needed for insn, last_scheduled_insn
+ should be set. */
+ gcc_assert (!(reload_completed && safe_group_barrier_needed (insn))
+ || last_scheduled_insn);
+
+ if ((reload_completed
+ && (safe_group_barrier_needed (insn)
+ || (mflag_sched_stop_bits_after_every_cycle
+ && last_clock != clock
+ && last_scheduled_insn
+ && scheduled_good_insn (last_scheduled_insn))))
|| (last_scheduled_insn
&& (GET_CODE (last_scheduled_insn) == CALL_INSN
|| GET_CODE (PATTERN (last_scheduled_insn)) == ASM_INPUT
|| asm_noperands (PATTERN (last_scheduled_insn)) >= 0)))
{
init_insn_group_barriers ();
+
if (verbose && dump)
fprintf (dump, "// Stop should be before %d%s\n", INSN_UID (insn),
last_clock == clock ? " + cycle advance" : "");
+
stop_before_p = 1;
+ current_cycle = clock;
+ mem_ops_in_group[current_cycle % 4] = 0;
+
if (last_clock == clock)
{
state_transition (curr_state, dfa_stop_insn);
}
else if (reload_completed)
setup_clocks_p = TRUE;
- if (GET_CODE (PATTERN (last_scheduled_insn)) == ASM_INPUT
- || asm_noperands (PATTERN (last_scheduled_insn)) >= 0)
- state_reset (curr_state);
- else
+
+ if (last_scheduled_insn)
{
- memcpy (curr_state, prev_cycle_state, dfa_state_size);
- state_transition (curr_state, dfa_stop_insn);
- state_transition (curr_state, dfa_pre_cycle_insn);
- state_transition (curr_state, NULL);
+ if (GET_CODE (PATTERN (last_scheduled_insn)) == ASM_INPUT
+ || asm_noperands (PATTERN (last_scheduled_insn)) >= 0)
+ state_reset (curr_state);
+ else
+ {
+ memcpy (curr_state, prev_cycle_state, dfa_state_size);
+ state_transition (curr_state, dfa_stop_insn);
+ state_transition (curr_state, dfa_pre_cycle_insn);
+ state_transition (curr_state, NULL);
+ }
}
}
else if (reload_completed)
setup_clocks_p = TRUE;
+
if (setup_clocks_p && ia64_tune == PROCESSOR_ITANIUM
&& GET_CODE (PATTERN (insn)) != ASM_INPUT
&& asm_noperands (PATTERN (insn)) < 0)
if (c != ITANIUM_CLASS_MMMUL && c != ITANIUM_CLASS_MMSHF)
{
- dep_link_t link;
+ sd_iterator_def sd_it;
+ dep_t dep;
int d = -1;
- FOR_EACH_DEP_LINK (link, INSN_BACK_DEPS (insn))
- if (DEP_LINK_KIND (link) == REG_DEP_TRUE)
+ FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
+ if (DEP_TYPE (dep) == REG_DEP_TRUE)
{
enum attr_itanium_class dep_class;
- rtx dep_insn = DEP_LINK_PRO (link);
+ rtx dep_insn = DEP_PRO (dep);
dep_class = ia64_safe_itanium_class (dep_insn);
if ((dep_class == ITANIUM_CLASS_MMMUL
add_cycles [INSN_UID (insn)] = 3 - d;
}
}
+
return 0;
}
static void
ia64_h_i_d_extended (void)
{
- if (current_sched_info->flags & DO_SPECULATION)
- {
- int new_max_uid = get_max_uid () + 1;
-
- spec_check_no = xrecalloc (spec_check_no, new_max_uid,
- max_uid, sizeof (*spec_check_no));
- max_uid = new_max_uid;
- }
-
if (stops_p != NULL)
{
- int new_clocks_length = get_max_uid () + 1;
+ int new_clocks_length = get_max_uid () * 3 / 2;
- stops_p = xrecalloc (stops_p, new_clocks_length, clocks_length, 1);
+ stops_p = (char *) xrecalloc (stops_p, new_clocks_length, clocks_length, 1);
if (ia64_tune == PROCESSOR_ITANIUM)
{
- clocks = xrecalloc (clocks, new_clocks_length, clocks_length,
- sizeof (int));
- add_cycles = xrecalloc (add_cycles, new_clocks_length, clocks_length,
- sizeof (int));
+ clocks = (int *) xrecalloc (clocks, new_clocks_length, clocks_length,
+ sizeof (int));
+ add_cycles = (int *) xrecalloc (add_cycles, new_clocks_length,
+ clocks_length, sizeof (int));
}
clocks_length = new_clocks_length;
}
}
+\f
+
+/* This structure describes the data used by the backend to guide scheduling.
+ When the current scheduling point is switched, this data should be saved
+ and restored later, if the scheduler returns to this point. */
+struct _ia64_sched_context
+{
+ state_t prev_cycle_state;
+ rtx last_scheduled_insn;
+ struct reg_write_state rws_sum[NUM_REGS];
+ struct reg_write_state rws_insn[NUM_REGS];
+ int first_instruction;
+ int pending_data_specs;
+ int current_cycle;
+ char mem_ops_in_group[4];
+};
+typedef struct _ia64_sched_context *ia64_sched_context_t;
+
+/* Allocates a scheduling context. */
+static void *
+ia64_alloc_sched_context (void)
+{
+ return xmalloc (sizeof (struct _ia64_sched_context));
+}
+
+/* Initializes the _SC context with clean data, if CLEAN_P, and from
+ the global context otherwise. */
+static void
+ia64_init_sched_context (void *_sc, bool clean_p)
+{
+ ia64_sched_context_t sc = (ia64_sched_context_t) _sc;
+
+ sc->prev_cycle_state = xmalloc (dfa_state_size);
+ if (clean_p)
+ {
+ state_reset (sc->prev_cycle_state);
+ sc->last_scheduled_insn = NULL_RTX;
+ memset (sc->rws_sum, 0, sizeof (rws_sum));
+ memset (sc->rws_insn, 0, sizeof (rws_insn));
+ sc->first_instruction = 1;
+ sc->pending_data_specs = 0;
+ sc->current_cycle = 0;
+ memset (sc->mem_ops_in_group, 0, sizeof (mem_ops_in_group));
+ }
+ else
+ {
+ memcpy (sc->prev_cycle_state, prev_cycle_state, dfa_state_size);
+ sc->last_scheduled_insn = last_scheduled_insn;
+ memcpy (sc->rws_sum, rws_sum, sizeof (rws_sum));
+ memcpy (sc->rws_insn, rws_insn, sizeof (rws_insn));
+ sc->first_instruction = first_instruction;
+ sc->pending_data_specs = pending_data_specs;
+ sc->current_cycle = current_cycle;
+ memcpy (sc->mem_ops_in_group, mem_ops_in_group, sizeof (mem_ops_in_group));
+ }
+}
+
+/* Sets the global scheduling context to the one pointed to by _SC. */
+static void
+ia64_set_sched_context (void *_sc)
+{
+ ia64_sched_context_t sc = (ia64_sched_context_t) _sc;
+
+ gcc_assert (sc != NULL);
+
+ memcpy (prev_cycle_state, sc->prev_cycle_state, dfa_state_size);
+ last_scheduled_insn = sc->last_scheduled_insn;
+ memcpy (rws_sum, sc->rws_sum, sizeof (rws_sum));
+ memcpy (rws_insn, sc->rws_insn, sizeof (rws_insn));
+ first_instruction = sc->first_instruction;
+ pending_data_specs = sc->pending_data_specs;
+ current_cycle = sc->current_cycle;
+ memcpy (mem_ops_in_group, sc->mem_ops_in_group, sizeof (mem_ops_in_group));
+}
+
+/* Clears the data in the _SC scheduling context. */
+static void
+ia64_clear_sched_context (void *_sc)
+{
+ ia64_sched_context_t sc = (ia64_sched_context_t) _sc;
+
+ free (sc->prev_cycle_state);
+ sc->prev_cycle_state = NULL;
+}
+
+/* Frees the _SC scheduling context. */
+static void
+ia64_free_sched_context (void *_sc)
+{
+ gcc_assert (_sc != NULL);
+
+ free (_sc);
+}
+
+typedef rtx (* gen_func_t) (rtx, rtx);
+
+/* Return a function that will generate a load of mode MODE_NO
+ with speculation types TS. */
+static gen_func_t
+get_spec_load_gen_function (ds_t ts, int mode_no)
+{
+ static gen_func_t gen_ld_[] = {
+ gen_movbi,
+ gen_movqi_internal,
+ gen_movhi_internal,
+ gen_movsi_internal,
+ gen_movdi_internal,
+ gen_movsf_internal,
+ gen_movdf_internal,
+ gen_movxf_internal,
+ gen_movti_internal,
+ gen_zero_extendqidi2,
+ gen_zero_extendhidi2,
+ gen_zero_extendsidi2,
+ };
+
+ static gen_func_t gen_ld_a[] = {
+ gen_movbi_advanced,
+ gen_movqi_advanced,
+ gen_movhi_advanced,
+ gen_movsi_advanced,
+ gen_movdi_advanced,
+ gen_movsf_advanced,
+ gen_movdf_advanced,
+ gen_movxf_advanced,
+ gen_movti_advanced,
+ gen_zero_extendqidi2_advanced,
+ gen_zero_extendhidi2_advanced,
+ gen_zero_extendsidi2_advanced,
+ };
+ static gen_func_t gen_ld_s[] = {
+ gen_movbi_speculative,
+ gen_movqi_speculative,
+ gen_movhi_speculative,
+ gen_movsi_speculative,
+ gen_movdi_speculative,
+ gen_movsf_speculative,
+ gen_movdf_speculative,
+ gen_movxf_speculative,
+ gen_movti_speculative,
+ gen_zero_extendqidi2_speculative,
+ gen_zero_extendhidi2_speculative,
+ gen_zero_extendsidi2_speculative,
+ };
+ static gen_func_t gen_ld_sa[] = {
+ gen_movbi_speculative_advanced,
+ gen_movqi_speculative_advanced,
+ gen_movhi_speculative_advanced,
+ gen_movsi_speculative_advanced,
+ gen_movdi_speculative_advanced,
+ gen_movsf_speculative_advanced,
+ gen_movdf_speculative_advanced,
+ gen_movxf_speculative_advanced,
+ gen_movti_speculative_advanced,
+ gen_zero_extendqidi2_speculative_advanced,
+ gen_zero_extendhidi2_speculative_advanced,
+ gen_zero_extendsidi2_speculative_advanced,
+ };
+ static gen_func_t gen_ld_s_a[] = {
+ gen_movbi_speculative_a,
+ gen_movqi_speculative_a,
+ gen_movhi_speculative_a,
+ gen_movsi_speculative_a,
+ gen_movdi_speculative_a,
+ gen_movsf_speculative_a,
+ gen_movdf_speculative_a,
+ gen_movxf_speculative_a,
+ gen_movti_speculative_a,
+ gen_zero_extendqidi2_speculative_a,
+ gen_zero_extendhidi2_speculative_a,
+ gen_zero_extendsidi2_speculative_a,
+ };
+
+ gen_func_t *gen_ld;
+
+ if (ts & BEGIN_DATA)
+ {
+ if (ts & BEGIN_CONTROL)
+ gen_ld = gen_ld_sa;
+ else
+ gen_ld = gen_ld_a;
+ }
+ else if (ts & BEGIN_CONTROL)
+ {
+ if ((spec_info->flags & SEL_SCHED_SPEC_DONT_CHECK_CONTROL)
+ || ia64_needs_block_p (ts))
+ gen_ld = gen_ld_s;
+ else
+ gen_ld = gen_ld_s_a;
+ }
+ else if (ts == 0)
+ gen_ld = gen_ld_;
+ else
+ gcc_unreachable ();
+
+ return gen_ld[mode_no];
+}
/* Constants that help mapping 'enum machine_mode' to int. */
enum SPEC_MODES
SPEC_MODE_LAST = 8
};
+enum
+ {
+ /* Offset to reach ZERO_EXTEND patterns. */
+ SPEC_GEN_EXTEND_OFFSET = SPEC_MODE_LAST - SPEC_MODE_FOR_EXTEND_FIRST + 1
+ };
+
/* Return index of the MODE. */
static int
ia64_mode_to_int (enum machine_mode mode)
unsigned int *flags = &(current_sched_info->flags);
if (*flags & SCHED_RGN
- || *flags & SCHED_EBB)
+ || *flags & SCHED_EBB
+ || *flags & SEL_SCHED)
{
int mask = 0;
if ((mflag_sched_br_data_spec && !reload_completed && optimize > 0)
- || (mflag_sched_ar_data_spec && reload_completed))
+ || (mflag_sched_ar_data_spec && reload_completed))
{
mask |= BEGIN_DATA;
-
- if ((mflag_sched_br_in_data_spec && !reload_completed)
- || (mflag_sched_ar_in_data_spec && reload_completed))
+
+ if (!sel_sched_p ()
+ && ((mflag_sched_br_in_data_spec && !reload_completed)
+ || (mflag_sched_ar_in_data_spec && reload_completed)))
mask |= BE_IN_DATA;
}
- if (mflag_sched_control_spec)
+ if (mflag_sched_control_spec
+ && (!sel_sched_p ()
+ || reload_completed))
{
mask |= BEGIN_CONTROL;
- if (mflag_sched_in_control_spec)
+ if (!sel_sched_p () && mflag_sched_in_control_spec)
mask |= BE_IN_CONTROL;
}
+ spec_info->mask = mask;
+
if (mask)
{
*flags |= USE_DEPS_LIST | DO_SPECULATION;
if (mask & BE_IN_SPEC)
*flags |= NEW_BBS;
- spec_info->mask = mask;
spec_info->flags = 0;
if ((mask & DATA_SPEC) && mflag_sched_prefer_non_data_spec_insns)
spec_info->flags |= PREFER_NON_DATA_SPEC;
- if ((mask & CONTROL_SPEC)
- && mflag_sched_prefer_non_control_spec_insns)
- spec_info->flags |= PREFER_NON_CONTROL_SPEC;
-
- if (mflag_sched_spec_verbose)
+ if (mask & CONTROL_SPEC)
{
- if (sched_verbose >= 1)
- spec_info->dump = sched_dump;
- else
- spec_info->dump = stderr;
+ if (mflag_sched_prefer_non_control_spec_insns)
+ spec_info->flags |= PREFER_NON_CONTROL_SPEC;
+
+ if (sel_sched_p () && mflag_sel_sched_dont_check_control_spec)
+ spec_info->flags |= SEL_SCHED_SPEC_DONT_CHECK_CONTROL;
}
+
+ if (sched_verbose >= 1)
+ spec_info->dump = sched_dump;
else
spec_info->dump = 0;
spec_info->flags |= COUNT_SPEC_IN_CRITICAL_PATH;
}
}
+ else
+ spec_info->mask = 0;
}
-/* Implement targetm.sched.speculate_insn hook.
- Check if the INSN can be TS speculative.
- If 'no' - return -1.
- If 'yes' - generate speculative pattern in the NEW_PAT and return 1.
- If current pattern of the INSN already provides TS speculation, return 0. */
+/* If INSN is an appropriate load return its mode.
+ Return -1 otherwise. */
static int
-ia64_speculate_insn (rtx insn, ds_t ts, rtx *new_pat)
-{
- rtx pat, reg, mem, mem_reg;
- int mode_no, gen_p = 1;
+get_mode_no_for_insn (rtx insn)
+{
+ rtx reg, mem, mode_rtx;
+ int mode_no;
bool extend_p;
-
- gcc_assert (!(ts & ~BEGIN_SPEC) && ts);
-
- pat = PATTERN (insn);
- if (GET_CODE (pat) == COND_EXEC)
- pat = COND_EXEC_CODE (pat);
+ extract_insn_cached (insn);
- /* This should be a SET ... */
- if (GET_CODE (pat) != SET)
- return -1;
+ /* We use WHICH_ALTERNATIVE only after reload. This will
+ guarantee that reload won't touch a speculative insn. */
- reg = SET_DEST (pat);
- /* ... to the general/fp register ... */
- if (!REG_P (reg) || !(GR_REGNO_P (REGNO (reg)) || FP_REGNO_P (REGNO (reg))))
+ if (recog_data.n_operands != 2)
return -1;
- /* ... from the mem ... */
- mem = SET_SRC (pat);
+ reg = recog_data.operand[0];
+ mem = recog_data.operand[1];
- /* ... that can, possibly, be a zero_extend ... */
- if (GET_CODE (mem) == ZERO_EXTEND)
+ /* We should use MEM's mode since REG's mode in presence of
+ ZERO_EXTEND will always be DImode. */
+ if (get_attr_speculable1 (insn) == SPECULABLE1_YES)
+ /* Process non-speculative ld. */
+ {
+ if (!reload_completed)
+ {
+ /* Do not speculate into regs like ar.lc. */
+ if (!REG_P (reg) || AR_REGNO_P (REGNO (reg)))
+ return -1;
+
+ if (!MEM_P (mem))
+ return -1;
+
+ {
+ rtx mem_reg = XEXP (mem, 0);
+
+ if (!REG_P (mem_reg))
+ return -1;
+ }
+
+ mode_rtx = mem;
+ }
+ else if (get_attr_speculable2 (insn) == SPECULABLE2_YES)
+ {
+ gcc_assert (REG_P (reg) && MEM_P (mem));
+ mode_rtx = mem;
+ }
+ else
+ return -1;
+ }
+ else if (get_attr_data_speculative (insn) == DATA_SPECULATIVE_YES
+ || get_attr_control_speculative (insn) == CONTROL_SPECULATIVE_YES
+ || get_attr_check_load (insn) == CHECK_LOAD_YES)
+ /* Process speculative ld or ld.c. */
{
- mem = XEXP (mem, 0);
- extend_p = true;
+ gcc_assert (REG_P (reg) && MEM_P (mem));
+ mode_rtx = mem;
}
else
- extend_p = false;
-
- /* ... or a speculative load. */
- if (GET_CODE (mem) == UNSPEC)
{
- int code;
-
- code = XINT (mem, 1);
- if (code != UNSPEC_LDA && code != UNSPEC_LDS && code != UNSPEC_LDSA)
- return -1;
-
- if ((code == UNSPEC_LDA && !(ts & BEGIN_CONTROL))
- || (code == UNSPEC_LDS && !(ts & BEGIN_DATA))
- || code == UNSPEC_LDSA)
- gen_p = 0;
+ enum attr_itanium_class attr_class = get_attr_itanium_class (insn);
- mem = XVECEXP (mem, 0, 0);
- gcc_assert (MEM_P (mem));
+ if (attr_class == ITANIUM_CLASS_CHK_A
+ || attr_class == ITANIUM_CLASS_CHK_S_I
+ || attr_class == ITANIUM_CLASS_CHK_S_F)
+ /* Process chk. */
+ mode_rtx = reg;
+ else
+ return -1;
}
- /* Source should be a mem ... */
- if (!MEM_P (mem))
- return -1;
+ mode_no = ia64_mode_to_int (GET_MODE (mode_rtx));
- /* ... addressed by a register. */
- mem_reg = XEXP (mem, 0);
- if (!REG_P (mem_reg))
- return -1;
-
- /* We should use MEM's mode since REG's mode in presence of ZERO_EXTEND
- will always be DImode. */
- mode_no = ia64_mode_to_int (GET_MODE (mem));
-
- if (mode_no == SPEC_MODE_INVALID
- || (extend_p
- && !(SPEC_MODE_FOR_EXTEND_FIRST <= mode_no
- && mode_no <= SPEC_MODE_FOR_EXTEND_LAST)))
+ if (mode_no == SPEC_MODE_INVALID)
return -1;
- extract_insn_cached (insn);
- gcc_assert (reg == recog_data.operand[0] && mem == recog_data.operand[1]);
+ extend_p = (GET_MODE (reg) != GET_MODE (mode_rtx));
+
+ if (extend_p)
+ {
+ if (!(SPEC_MODE_FOR_EXTEND_FIRST <= mode_no
+ && mode_no <= SPEC_MODE_FOR_EXTEND_LAST))
+ return -1;
- *new_pat = ia64_gen_spec_insn (insn, ts, mode_no, gen_p != 0, extend_p);
+ mode_no += SPEC_GEN_EXTEND_OFFSET;
+ }
- return gen_p;
+ return mode_no;
}
-enum
- {
- /* Offset to reach ZERO_EXTEND patterns. */
- SPEC_GEN_EXTEND_OFFSET = SPEC_MODE_LAST - SPEC_MODE_FOR_EXTEND_FIRST + 1,
- /* Number of patterns for each speculation mode. */
- SPEC_N = (SPEC_MODE_LAST
- + SPEC_MODE_FOR_EXTEND_LAST - SPEC_MODE_FOR_EXTEND_FIRST + 2)
- };
+/* If X is an unspec part of a speculative load, return its code.
+ Return -1 otherwise. */
+static int
+get_spec_unspec_code (const_rtx x)
+{
+ if (GET_CODE (x) != UNSPEC)
+ return -1;
-enum SPEC_GEN_LD_MAP
{
- /* Offset to ld.a patterns. */
- SPEC_GEN_A = 0 * SPEC_N,
- /* Offset to ld.s patterns. */
- SPEC_GEN_S = 1 * SPEC_N,
- /* Offset to ld.sa patterns. */
- SPEC_GEN_SA = 2 * SPEC_N,
- /* Offset to ld.sa patterns. For this patterns corresponding ld.c will
- mutate to chk.s. */
- SPEC_GEN_SA_FOR_S = 3 * SPEC_N
- };
+ int code;
-/* These offsets are used to get (4 * SPEC_N). */
-enum SPEC_GEN_CHECK_OFFSET
- {
- SPEC_GEN_CHKA_FOR_A_OFFSET = 4 * SPEC_N - SPEC_GEN_A,
- SPEC_GEN_CHKA_FOR_SA_OFFSET = 4 * SPEC_N - SPEC_GEN_SA
- };
+ code = XINT (x, 1);
-/* If GEN_P is true, calculate the index of needed speculation check and return
- speculative pattern for INSN with speculative mode TS, machine mode
- MODE_NO and with ZERO_EXTEND (if EXTEND_P is true).
- If GEN_P is false, just calculate the index of needed speculation check. */
-static rtx
-ia64_gen_spec_insn (rtx insn, ds_t ts, int mode_no, bool gen_p, bool extend_p)
+ switch (code)
+ {
+ case UNSPEC_LDA:
+ case UNSPEC_LDS:
+ case UNSPEC_LDS_A:
+ case UNSPEC_LDSA:
+ return code;
+
+ default:
+ return -1;
+ }
+ }
+}
+
+/* Implement skip_rtx_p hook. */
+static bool
+ia64_skip_rtx_p (const_rtx x)
{
- rtx pat, new_pat;
- int load_no;
- int shift = 0;
+ return get_spec_unspec_code (x) != -1;
+}
- static rtx (* const gen_load[]) (rtx, rtx) = {
- gen_movbi_advanced,
- gen_movqi_advanced,
- gen_movhi_advanced,
- gen_movsi_advanced,
- gen_movdi_advanced,
- gen_movsf_advanced,
- gen_movdf_advanced,
- gen_movxf_advanced,
- gen_movti_advanced,
- gen_zero_extendqidi2_advanced,
- gen_zero_extendhidi2_advanced,
- gen_zero_extendsidi2_advanced,
+/* If INSN is a speculative load, return its UNSPEC code.
+ Return -1 otherwise. */
+static int
+get_insn_spec_code (const_rtx insn)
+{
+ rtx pat, reg, mem;
- gen_movbi_speculative,
- gen_movqi_speculative,
- gen_movhi_speculative,
- gen_movsi_speculative,
- gen_movdi_speculative,
- gen_movsf_speculative,
- gen_movdf_speculative,
- gen_movxf_speculative,
- gen_movti_speculative,
- gen_zero_extendqidi2_speculative,
- gen_zero_extendhidi2_speculative,
- gen_zero_extendsidi2_speculative,
+ pat = PATTERN (insn);
- gen_movbi_speculative_advanced,
- gen_movqi_speculative_advanced,
- gen_movhi_speculative_advanced,
- gen_movsi_speculative_advanced,
- gen_movdi_speculative_advanced,
- gen_movsf_speculative_advanced,
- gen_movdf_speculative_advanced,
- gen_movxf_speculative_advanced,
- gen_movti_speculative_advanced,
- gen_zero_extendqidi2_speculative_advanced,
- gen_zero_extendhidi2_speculative_advanced,
- gen_zero_extendsidi2_speculative_advanced,
+ if (GET_CODE (pat) == COND_EXEC)
+ pat = COND_EXEC_CODE (pat);
- gen_movbi_speculative_advanced,
- gen_movqi_speculative_advanced,
- gen_movhi_speculative_advanced,
- gen_movsi_speculative_advanced,
- gen_movdi_speculative_advanced,
- gen_movsf_speculative_advanced,
- gen_movdf_speculative_advanced,
- gen_movxf_speculative_advanced,
- gen_movti_speculative_advanced,
- gen_zero_extendqidi2_speculative_advanced,
- gen_zero_extendhidi2_speculative_advanced,
- gen_zero_extendsidi2_speculative_advanced
- };
+ if (GET_CODE (pat) != SET)
+ return -1;
- load_no = extend_p ? mode_no + SPEC_GEN_EXTEND_OFFSET : mode_no;
+ reg = SET_DEST (pat);
+ if (!REG_P (reg))
+ return -1;
- if (ts & BEGIN_DATA)
+ mem = SET_SRC (pat);
+ if (GET_CODE (mem) == ZERO_EXTEND)
+ mem = XEXP (mem, 0);
+
+ return get_spec_unspec_code (mem);
+}
+
+/* If INSN is a speculative load, return a ds with the speculation types.
+ Otherwise [if INSN is a normal instruction] return 0. */
+static ds_t
+ia64_get_insn_spec_ds (rtx insn)
+{
+ int code = get_insn_spec_code (insn);
+
+ switch (code)
{
- /* We don't need recovery because even if this is ld.sa
- ALAT entry will be allocated only if NAT bit is set to zero.
- So it is enough to use ld.c here. */
+ case UNSPEC_LDA:
+ return BEGIN_DATA;
- if (ts & BEGIN_CONTROL)
- {
- load_no += SPEC_GEN_SA;
+ case UNSPEC_LDS:
+ case UNSPEC_LDS_A:
+ return BEGIN_CONTROL;
- if (!mflag_sched_ldc)
- shift = SPEC_GEN_CHKA_FOR_SA_OFFSET;
- }
- else
- {
- load_no += SPEC_GEN_A;
+ case UNSPEC_LDSA:
+ return BEGIN_DATA | BEGIN_CONTROL;
- if (!mflag_sched_ldc)
- shift = SPEC_GEN_CHKA_FOR_A_OFFSET;
- }
+ default:
+ return 0;
}
- else if (ts & BEGIN_CONTROL)
+}
+
+/* If INSN is a speculative load return a ds with the speculation types that
+ will be checked.
+ Otherwise [if INSN is a normal instruction] return 0. */
+static ds_t
+ia64_get_insn_checked_ds (rtx insn)
+{
+ int code = get_insn_spec_code (insn);
+
+ switch (code)
{
- /* ld.sa can be used instead of ld.s to avoid basic block splitting. */
- if (!mflag_control_ldc)
- load_no += SPEC_GEN_S;
- else
- {
- gcc_assert (mflag_sched_ldc);
- load_no += SPEC_GEN_SA_FOR_S;
- }
+ case UNSPEC_LDA:
+ return BEGIN_DATA | BEGIN_CONTROL;
+
+ case UNSPEC_LDS:
+ return BEGIN_CONTROL;
+
+ case UNSPEC_LDS_A:
+ case UNSPEC_LDSA:
+ return BEGIN_DATA | BEGIN_CONTROL;
+
+ default:
+ return 0;
}
- else
- gcc_unreachable ();
+}
- /* Set the desired check index. We add '1', because zero element in this
- array means, that instruction with such uid is non-speculative. */
- spec_check_no[INSN_UID (insn)] = load_no + shift + 1;
+/* If GEN_P is true, calculate the index of needed speculation check and return
+ speculative pattern for INSN with speculative mode TS, machine mode
+ MODE_NO and with ZERO_EXTEND (if EXTEND_P is true).
+ If GEN_P is false, just calculate the index of needed speculation check. */
+static rtx
+ia64_gen_spec_load (rtx insn, ds_t ts, int mode_no)
+{
+ rtx pat, new_pat;
+ gen_func_t gen_load;
- if (!gen_p)
- return 0;
+ gen_load = get_spec_load_gen_function (ts, mode_no);
- new_pat = gen_load[load_no] (copy_rtx (recog_data.operand[0]),
- copy_rtx (recog_data.operand[1]));
+ new_pat = gen_load (copy_rtx (recog_data.operand[0]),
+ copy_rtx (recog_data.operand[1]));
pat = PATTERN (insn);
if (GET_CODE (pat) == COND_EXEC)
- new_pat = gen_rtx_COND_EXEC (VOIDmode, copy_rtx
- (COND_EXEC_TEST (pat)), new_pat);
+ new_pat = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (COND_EXEC_TEST (pat)),
+ new_pat);
return new_pat;
}
-/* Offset to branchy checks. */
-enum { SPEC_GEN_CHECK_MUTATION_OFFSET = 5 * SPEC_N };
-
-/* Return nonzero, if INSN needs branchy recovery check. */
static bool
-ia64_needs_block_p (rtx insn)
+insn_can_be_in_speculative_p (rtx insn ATTRIBUTE_UNUSED,
+ ds_t ds ATTRIBUTE_UNUSED)
{
- int check_no;
+ return false;
+}
- check_no = spec_check_no[INSN_UID(insn)] - 1;
- gcc_assert (0 <= check_no && check_no < SPEC_GEN_CHECK_MUTATION_OFFSET);
+/* Implement targetm.sched.speculate_insn hook.
+ Check if the INSN can be TS speculative.
+ If 'no' - return -1.
+ If 'yes' - generate speculative pattern in the NEW_PAT and return 1.
+ If current pattern of the INSN already provides TS speculation,
+ return 0. */
+static int
+ia64_speculate_insn (rtx insn, ds_t ts, rtx *new_pat)
+{
+ int mode_no;
+ int res;
+
+ gcc_assert (!(ts & ~SPECULATIVE));
- return ((SPEC_GEN_S <= check_no && check_no < SPEC_GEN_S + SPEC_N)
- || (4 * SPEC_N <= check_no && check_no < 4 * SPEC_N + SPEC_N));
-}
+ if (ia64_spec_check_p (insn))
+ return -1;
-/* Generate (or regenerate, if (MUTATE_P)) recovery check for INSN.
- If (LABEL != 0 || MUTATE_P), generate branchy recovery check.
- Otherwise, generate a simple check. */
-static rtx
-ia64_gen_check (rtx insn, rtx label, bool mutate_p)
-{
- rtx op1, pat, check_pat;
+ if ((ts & BE_IN_SPEC)
+ && !insn_can_be_in_speculative_p (insn, ts))
+ return -1;
- static rtx (* const gen_check[]) (rtx, rtx) = {
- gen_movbi_clr,
- gen_movqi_clr,
- gen_movhi_clr,
- gen_movsi_clr,
- gen_movdi_clr,
- gen_movsf_clr,
- gen_movdf_clr,
- gen_movxf_clr,
- gen_movti_clr,
- gen_zero_extendqidi2_clr,
- gen_zero_extendhidi2_clr,
- gen_zero_extendsidi2_clr,
+ mode_no = get_mode_no_for_insn (insn);
- gen_speculation_check_bi,
- gen_speculation_check_qi,
- gen_speculation_check_hi,
- gen_speculation_check_si,
- gen_speculation_check_di,
- gen_speculation_check_sf,
- gen_speculation_check_df,
- gen_speculation_check_xf,
- gen_speculation_check_ti,
- gen_speculation_check_di,
- gen_speculation_check_di,
- gen_speculation_check_di,
+ if (mode_no != SPEC_MODE_INVALID)
+ {
+ if (ia64_get_insn_spec_ds (insn) == ds_get_speculation_types (ts))
+ res = 0;
+ else
+ {
+ res = 1;
+ *new_pat = ia64_gen_spec_load (insn, ts, mode_no);
+ }
+ }
+ else
+ res = -1;
- gen_movbi_clr,
- gen_movqi_clr,
- gen_movhi_clr,
- gen_movsi_clr,
- gen_movdi_clr,
- gen_movsf_clr,
- gen_movdf_clr,
- gen_movxf_clr,
- gen_movti_clr,
- gen_zero_extendqidi2_clr,
- gen_zero_extendhidi2_clr,
- gen_zero_extendsidi2_clr,
+ return res;
+}
+/* Return a function that will generate a check for speculation TS with mode
+ MODE_NO.
+ If simple check is needed, pass true for SIMPLE_CHECK_P.
+ If clearing check is needed, pass true for CLEARING_CHECK_P. */
+static gen_func_t
+get_spec_check_gen_function (ds_t ts, int mode_no,
+ bool simple_check_p, bool clearing_check_p)
+{
+ static gen_func_t gen_ld_c_clr[] = {
gen_movbi_clr,
gen_movqi_clr,
gen_movhi_clr,
gen_zero_extendqidi2_clr,
gen_zero_extendhidi2_clr,
gen_zero_extendsidi2_clr,
-
- gen_advanced_load_check_clr_bi,
- gen_advanced_load_check_clr_qi,
- gen_advanced_load_check_clr_hi,
- gen_advanced_load_check_clr_si,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_sf,
- gen_advanced_load_check_clr_df,
- gen_advanced_load_check_clr_xf,
- gen_advanced_load_check_clr_ti,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_di,
-
- /* Following checks are generated during mutation. */
- gen_advanced_load_check_clr_bi,
- gen_advanced_load_check_clr_qi,
- gen_advanced_load_check_clr_hi,
- gen_advanced_load_check_clr_si,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_sf,
- gen_advanced_load_check_clr_df,
- gen_advanced_load_check_clr_xf,
- gen_advanced_load_check_clr_ti,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_di,
- gen_advanced_load_check_clr_di,
-
- 0,0,0,0,0,0,0,0,0,0,0,0,
-
+ };
+ static gen_func_t gen_ld_c_nc[] = {
+ gen_movbi_nc,
+ gen_movqi_nc,
+ gen_movhi_nc,
+ gen_movsi_nc,
+ gen_movdi_nc,
+ gen_movsf_nc,
+ gen_movdf_nc,
+ gen_movxf_nc,
+ gen_movti_nc,
+ gen_zero_extendqidi2_nc,
+ gen_zero_extendhidi2_nc,
+ gen_zero_extendsidi2_nc,
+ };
+ static gen_func_t gen_chk_a_clr[] = {
gen_advanced_load_check_clr_bi,
gen_advanced_load_check_clr_qi,
gen_advanced_load_check_clr_hi,
gen_advanced_load_check_clr_di,
gen_advanced_load_check_clr_di,
gen_advanced_load_check_clr_di,
-
+ };
+ static gen_func_t gen_chk_a_nc[] = {
+ gen_advanced_load_check_nc_bi,
+ gen_advanced_load_check_nc_qi,
+ gen_advanced_load_check_nc_hi,
+ gen_advanced_load_check_nc_si,
+ gen_advanced_load_check_nc_di,
+ gen_advanced_load_check_nc_sf,
+ gen_advanced_load_check_nc_df,
+ gen_advanced_load_check_nc_xf,
+ gen_advanced_load_check_nc_ti,
+ gen_advanced_load_check_nc_di,
+ gen_advanced_load_check_nc_di,
+ gen_advanced_load_check_nc_di,
+ };
+ static gen_func_t gen_chk_s[] = {
gen_speculation_check_bi,
gen_speculation_check_qi,
gen_speculation_check_hi,
gen_speculation_check_ti,
gen_speculation_check_di,
gen_speculation_check_di,
- gen_speculation_check_di
+ gen_speculation_check_di,
};
- extract_insn_cached (insn);
+ gen_func_t *gen_check;
- if (label)
- {
- gcc_assert (mutate_p || ia64_needs_block_p (insn));
- op1 = label;
- }
- else
+ if (ts & BEGIN_DATA)
{
- gcc_assert (!mutate_p && !ia64_needs_block_p (insn));
- op1 = copy_rtx (recog_data.operand[1]);
+ /* We don't need recovery because even if this is ld.sa
+ ALAT entry will be allocated only if NAT bit is set to zero.
+ So it is enough to use ld.c here. */
+
+ if (simple_check_p)
+ {
+ gcc_assert (mflag_sched_spec_ldc);
+
+ if (clearing_check_p)
+ gen_check = gen_ld_c_clr;
+ else
+ gen_check = gen_ld_c_nc;
+ }
+ else
+ {
+ if (clearing_check_p)
+ gen_check = gen_chk_a_clr;
+ else
+ gen_check = gen_chk_a_nc;
+ }
}
-
- if (mutate_p)
- /* INSN is ld.c.
- Find the speculation check number by searching for original
- speculative load in the RESOLVED_DEPS list of INSN.
- As long as patterns are unique for each instruction, this can be
- accomplished by matching ORIG_PAT fields. */
+ else if (ts & BEGIN_CONTROL)
{
- dep_link_t link;
- int check_no = 0;
- rtx orig_pat = ORIG_PAT (insn);
-
- FOR_EACH_DEP_LINK (link, INSN_RESOLVED_BACK_DEPS (insn))
+ if (simple_check_p)
+ /* We might want to use ld.sa -> ld.c instead of
+ ld.s -> chk.s. */
{
- rtx x = DEP_LINK_PRO (link);
+ gcc_assert (!ia64_needs_block_p (ts));
- if (ORIG_PAT (x) == orig_pat)
- check_no = spec_check_no[INSN_UID (x)];
+ if (clearing_check_p)
+ gen_check = gen_ld_c_clr;
+ else
+ gen_check = gen_ld_c_nc;
}
- gcc_assert (check_no);
+ else
+ {
+ gen_check = gen_chk_s;
+ }
+ }
+ else
+ gcc_unreachable ();
+
+ gcc_assert (mode_no >= 0);
+ return gen_check[mode_no];
+}
+
+/* Return nonzero, if INSN needs branchy recovery check. */
+static bool
+ia64_needs_block_p (ds_t ts)
+{
+ if (ts & BEGIN_DATA)
+ return !mflag_sched_spec_ldc;
+
+ gcc_assert ((ts & BEGIN_CONTROL) != 0);
+
+ return !(mflag_sched_spec_control_ldc && mflag_sched_spec_ldc);
+}
+
+/* Generate (or regenerate, if (MUTATE_P)) recovery check for INSN.
+ If (LABEL != 0 || MUTATE_P), generate branchy recovery check.
+ Otherwise, generate a simple check. */
+static rtx
+ia64_gen_spec_check (rtx insn, rtx label, ds_t ds)
+{
+ rtx op1, pat, check_pat;
+ gen_func_t gen_check;
+ int mode_no;
- spec_check_no[INSN_UID (insn)] = (check_no
- + SPEC_GEN_CHECK_MUTATION_OFFSET);
+ mode_no = get_mode_no_for_insn (insn);
+ gcc_assert (mode_no >= 0);
+
+ if (label)
+ op1 = label;
+ else
+ {
+ gcc_assert (!ia64_needs_block_p (ds));
+ op1 = copy_rtx (recog_data.operand[1]);
}
+
+ gen_check = get_spec_check_gen_function (ds, mode_no, label == NULL_RTX,
+ true);
- check_pat = (gen_check[spec_check_no[INSN_UID (insn)] - 1]
- (copy_rtx (recog_data.operand[0]), op1));
+ check_pat = gen_check (copy_rtx (recog_data.operand[0]), op1);
pat = PATTERN (insn);
if (GET_CODE (pat) == COND_EXEC)
code = XINT (t, 1);
- if (code == UNSPEC_CHKACLR
- || code == UNSPEC_CHKS
- || code == UNSPEC_LDCCLR)
+ if (code == UNSPEC_LDCCLR
+ || code == UNSPEC_LDCNC
+ || code == UNSPEC_CHKACLR
+ || code == UNSPEC_CHKANC
+ || code == UNSPEC_CHKS)
{
gcc_assert (code != 0);
return code;
int accumulated_insns_num; /* number of all previous insns including
nops. L is considered as 2 insns */
int branch_deviation; /* deviation of previous branches from 3rd slots */
+ int middle_bundle_stops; /* number of stop bits in the middle of bundles */
struct bundle_state *next; /* next state with the same insn_num */
struct bundle_state *originator; /* originator (previous insn state) */
/* All bundle states are in the following chain. */
}
else
{
- result = xmalloc (sizeof (struct bundle_state));
+ result = XNEW (struct bundle_state);
result->dfa_state = xmalloc (dfa_state_size);
result->allocated_states_chain = allocated_bundle_states_chain;
allocated_bundle_states_chain = result;
static unsigned
bundle_state_hash (const void *bundle_state)
{
- const struct bundle_state *state = (struct bundle_state *) bundle_state;
+ const struct bundle_state *const state
+ = (const struct bundle_state *) bundle_state;
unsigned result, i;
for (result = i = 0; i < dfa_state_size; i++)
static int
bundle_state_eq_p (const void *bundle_state_1, const void *bundle_state_2)
{
- const struct bundle_state * state1 = (struct bundle_state *) bundle_state_1;
- const struct bundle_state * state2 = (struct bundle_state *) bundle_state_2;
+ const struct bundle_state *const state1
+ = (const struct bundle_state *) bundle_state_1;
+ const struct bundle_state *const state2
+ = (const struct bundle_state *) bundle_state_2;
return (state1->insn_num == state2->insn_num
&& memcmp (state1->dfa_state, state2->dfa_state,
|| (((struct bundle_state *)
*entry_ptr)->accumulated_insns_num
== bundle_state->accumulated_insns_num
- && ((struct bundle_state *)
- *entry_ptr)->branch_deviation
- > bundle_state->branch_deviation))))
+ && (((struct bundle_state *)
+ *entry_ptr)->branch_deviation
+ > bundle_state->branch_deviation
+ || (((struct bundle_state *)
+ *entry_ptr)->branch_deviation
+ == bundle_state->branch_deviation
+ && ((struct bundle_state *)
+ *entry_ptr)->middle_bundle_stops
+ > bundle_state->middle_bundle_stops))))))
{
struct bundle_state temp;
curr_state->accumulated_insns_num
= originator->accumulated_insns_num + before_nops_num;
curr_state->branch_deviation = originator->branch_deviation;
+ curr_state->middle_bundle_stops = originator->middle_bundle_stops;
gcc_assert (insn);
if (INSN_CODE (insn) == CODE_FOR_insn_group_barrier)
{
if (!try_issue_insn (curr_state, insn))
return;
memcpy (temp_dfa_state, curr_state->dfa_state, dfa_state_size);
+ if (curr_state->accumulated_insns_num % 3 != 0)
+ curr_state->middle_bundle_stops++;
if (state_transition (temp_dfa_state, dfa_pre_cycle_insn) >= 0
&& curr_state->accumulated_insns_num % 3 != 0)
{
}
}
+/* True when INSN is important for bundling. */
+static bool
+important_for_bundling_p (rtx insn)
+{
+ return (INSN_P (insn)
+ && ia64_safe_itanium_class (insn) != ITANIUM_CLASS_IGNORE
+ && GET_CODE (PATTERN (insn)) != USE
+ && GET_CODE (PATTERN (insn)) != CLOBBER);
+}
+
/* The following function returns an insn important for insn bundling
followed by INSN and before TAIL. */
get_next_important_insn (rtx insn, rtx tail)
{
for (; insn && insn != tail; insn = NEXT_INSN (insn))
- if (INSN_P (insn)
- && ia64_safe_itanium_class (insn) != ITANIUM_CLASS_IGNORE
- && GET_CODE (PATTERN (insn)) != USE
- && GET_CODE (PATTERN (insn)) != CLOBBER)
+ if (important_for_bundling_p (insn))
return insn;
return NULL_RTX;
}
bundling_p = 1;
dfa_clean_insn_cache ();
initiate_bundle_state_table ();
- index_to_bundle_states = xmalloc ((insn_num + 2)
- * sizeof (struct bundle_state *));
+ index_to_bundle_states = XNEWVEC (struct bundle_state *, insn_num + 2);
/* First (forward) pass -- generation of bundle states. */
curr_state = get_free_bundle_state ();
curr_state->insn = NULL;
curr_state->cost = 0;
curr_state->accumulated_insns_num = 0;
curr_state->branch_deviation = 0;
+ curr_state->middle_bundle_stops = 0;
curr_state->next = NULL;
curr_state->originator = NULL;
state_reset (curr_state->dfa_state);
if (INSN_P (next_insn)
&& ia64_safe_itanium_class (next_insn) != ITANIUM_CLASS_IGNORE
&& GET_CODE (PATTERN (next_insn)) != USE
- && GET_CODE (PATTERN (next_insn)) != CLOBBER)
+ && GET_CODE (PATTERN (next_insn)) != CLOBBER
+ && INSN_CODE (next_insn) != CODE_FOR_insn_group_barrier)
{
PUT_MODE (next_insn, TImode);
break;
fprintf
(dump,
- "// Bundle state %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, state %d) for %d\n",
+ "// Bundle state %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, mid.stops %d state %d) for %d\n",
curr_state->unique_num,
(curr_state->originator == NULL
? -1 : curr_state->originator->unique_num),
curr_state->cost,
curr_state->before_nops_num, curr_state->after_nops_num,
curr_state->accumulated_insns_num, curr_state->branch_deviation,
+ curr_state->middle_bundle_stops,
(ia64_tune == PROCESSOR_ITANIUM
? ((struct DFA_chip *) curr_state->dfa_state)->oneb_automaton_state
: ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state),
< best_state->accumulated_insns_num
|| (curr_state->accumulated_insns_num
== best_state->accumulated_insns_num
- && curr_state->branch_deviation
- < best_state->branch_deviation)))))
+ && (curr_state->branch_deviation
+ < best_state->branch_deviation
+ || (curr_state->branch_deviation
+ == best_state->branch_deviation
+ && curr_state->middle_bundle_stops
+ < best_state->middle_bundle_stops)))))))
best_state = curr_state;
/* Second (backward) pass: adding nops and templates. */
+ gcc_assert (best_state);
insn_num = best_state->before_nops_num;
template0 = template1 = -1;
for (curr_state = best_state;
fprintf
(dump,
- "// Best %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, state %d) for %d\n",
+ "// Best %d (orig %d, cost %d, nops %d/%d, insns %d, branch %d, mid.stops %d, state %d) for %d\n",
curr_state->unique_num,
(curr_state->originator == NULL
? -1 : curr_state->originator->unique_num),
curr_state->cost,
curr_state->before_nops_num, curr_state->after_nops_num,
curr_state->accumulated_insns_num, curr_state->branch_deviation,
+ curr_state->middle_bundle_stops,
(ia64_tune == PROCESSOR_ITANIUM
? ((struct DFA_chip *) curr_state->dfa_state)->oneb_automaton_state
: ((struct DFA_chip *) curr_state->dfa_state)->twob_automaton_state),
insn);
}
}
+
+#ifdef ENABLE_CHECKING
+ {
+ /* Assert right calculation of middle_bundle_stops. */
+ int num = best_state->middle_bundle_stops;
+ bool start_bundle = true, end_bundle = false;
+
+ for (insn = NEXT_INSN (prev_head_insn);
+ insn && insn != tail;
+ insn = NEXT_INSN (insn))
+ {
+ if (!INSN_P (insn))
+ continue;
+ if (recog_memoized (insn) == CODE_FOR_bundle_selector)
+ start_bundle = true;
+ else
+ {
+ rtx next_insn;
+
+ for (next_insn = NEXT_INSN (insn);
+ next_insn && next_insn != tail;
+ next_insn = NEXT_INSN (next_insn))
+ if (INSN_P (next_insn)
+ && (ia64_safe_itanium_class (next_insn)
+ != ITANIUM_CLASS_IGNORE
+ || recog_memoized (next_insn)
+ == CODE_FOR_bundle_selector)
+ && GET_CODE (PATTERN (next_insn)) != USE
+ && GET_CODE (PATTERN (next_insn)) != CLOBBER)
+ break;
+
+ end_bundle = next_insn == NULL_RTX
+ || next_insn == tail
+ || (INSN_P (next_insn)
+ && recog_memoized (next_insn)
+ == CODE_FOR_bundle_selector);
+ if (recog_memoized (insn) == CODE_FOR_insn_group_barrier
+ && !start_bundle && !end_bundle
+ && next_insn
+ && GET_CODE (PATTERN (next_insn)) != ASM_INPUT
+ && asm_noperands (PATTERN (next_insn)) < 0)
+ num--;
+
+ start_bundle = false;
+ }
+ }
+
+ gcc_assert (num == 0);
+ }
+#endif
+
free (index_to_bundle_states);
finish_bundle_state_table ();
bundling_p = 0;
{
rtx insn;
int need_barrier_p = 0;
+ int seen_good_insn = 0;
rtx prev_insn = NULL_RTX;
init_insn_group_barriers ();
emit_insn_after (gen_insn_group_barrier (GEN_INT (3)), last);
init_insn_group_barriers ();
+ seen_good_insn = 0;
need_barrier_p = 0;
prev_insn = NULL_RTX;
}
if (recog_memoized (insn) == CODE_FOR_insn_group_barrier)
{
init_insn_group_barriers ();
+ seen_good_insn = 0;
need_barrier_p = 0;
prev_insn = NULL_RTX;
}
- else if (need_barrier_p || group_barrier_needed (insn))
+ else if (need_barrier_p || group_barrier_needed (insn)
+ || (mflag_sched_stop_bits_after_every_cycle
+ && GET_MODE (insn) == TImode
+ && seen_good_insn))
{
if (TARGET_EARLY_STOP_BITS)
{
last != insn;
last = NEXT_INSN (last))
if (INSN_P (last))
- group_barrier_needed (last);
+ {
+ group_barrier_needed (last);
+ if (recog_memoized (last) >= 0
+ && important_for_bundling_p (last))
+ seen_good_insn = 1;
+ }
}
else
{
emit_insn_before (gen_insn_group_barrier (GEN_INT (3)),
insn);
init_insn_group_barriers ();
+ seen_good_insn = 0;
}
group_barrier_needed (insn);
+ if (recog_memoized (insn) >= 0
+ && important_for_bundling_p (insn))
+ seen_good_insn = 1;
prev_insn = NULL_RTX;
}
- else if (recog_memoized (insn) >= 0)
- prev_insn = insn;
+ else if (recog_memoized (insn) >= 0
+ && important_for_bundling_p (insn))
+ {
+ prev_insn = insn;
+ seen_good_insn = 1;
+ }
need_barrier_p = (GET_CODE (insn) == CALL_INSN
|| GET_CODE (PATTERN (insn)) == ASM_INPUT
|| asm_noperands (PATTERN (insn)) >= 0);
{
int c = XINT (mem, 1);
- gcc_assert (c == UNSPEC_LDA || c == UNSPEC_LDS || c == UNSPEC_LDSA);
+ gcc_assert (c == UNSPEC_LDA || c == UNSPEC_LDS || c == UNSPEC_LDS_A
+ || c == UNSPEC_LDSA);
mem = XVECEXP (mem, 0, 0);
}
if (optimize == 0)
split_all_insns ();
- if (optimize && ia64_flag_schedule_insns2 && dbg_cnt (ia64_sched2))
+ if (optimize && ia64_flag_schedule_insns2
+ && dbg_cnt (ia64_sched2))
{
timevar_push (TV_SCHED2);
ia64_final_schedule = 1;
PREV_INSN (ia64_nop) = NEXT_INSN (ia64_nop) = NULL_RTX;
recog_memoized (ia64_nop);
clocks_length = get_max_uid () + 1;
- stops_p = xcalloc (1, clocks_length);
+ stops_p = XCNEWVEC (char, clocks_length);
if (ia64_tune == PROCESSOR_ITANIUM)
{
- clocks = xcalloc (clocks_length, sizeof (int));
- add_cycles = xcalloc (clocks_length, sizeof (int));
+ clocks = XCNEWVEC (int, clocks_length);
+ add_cycles = XCNEWVEC (int, clocks_length);
}
if (ia64_tune == PROCESSOR_ITANIUM2)
{
_1mfb_ = get_cpu_unit_code ("1b_1mfb.");
_1mlx_ = get_cpu_unit_code ("1b_1mlx.");
}
- schedule_ebbs ();
+
+ if (flag_selective_scheduling2
+ && !maybe_skip_selective_scheduling ())
+ run_selective_scheduling ();
+ else
+ schedule_ebbs ();
+
+ /* Redo alignment computation, as it might gone wrong. */
+ compute_alignments ();
+
/* We cannot reuse this one because it has been corrupted by the
evil glat. */
finish_bundle_states ();
types which can't go in sdata/sbss. */
static bool
-ia64_in_small_data_p (tree exp)
+ia64_in_small_data_p (const_tree exp)
{
if (TARGET_NO_SDATA)
return false;
enum ia64_builtins
{
IA64_BUILTIN_BSP,
- IA64_BUILTIN_FLUSHRS
+ IA64_BUILTIN_COPYSIGNQ,
+ IA64_BUILTIN_FABSQ,
+ IA64_BUILTIN_FLUSHRS,
+ IA64_BUILTIN_INFQ,
+ IA64_BUILTIN_HUGE_VALQ
};
void
/* The __float128 type. */
if (!TARGET_HPUX)
{
+ tree ftype, decl;
tree float128_type = make_node (REAL_TYPE);
+
TYPE_PRECISION (float128_type) = 128;
layout_type (float128_type);
(*lang_hooks.types.register_builtin_type) (float128_type, "__float128");
+
+ /* TFmode support builtins. */
+ ftype = build_function_type (float128_type, void_list_node);
+ add_builtin_function ("__builtin_infq", ftype,
+ IA64_BUILTIN_INFQ, BUILT_IN_MD,
+ NULL, NULL_TREE);
+
+ add_builtin_function ("__builtin_huge_valq", ftype,
+ IA64_BUILTIN_HUGE_VALQ, BUILT_IN_MD,
+ NULL, NULL_TREE);
+
+ ftype = build_function_type_list (float128_type,
+ float128_type,
+ NULL_TREE);
+ decl = add_builtin_function ("__builtin_fabsq", ftype,
+ IA64_BUILTIN_FABSQ, BUILT_IN_MD,
+ "__fabstf2", NULL_TREE);
+ TREE_READONLY (decl) = 1;
+
+ ftype = build_function_type_list (float128_type,
+ float128_type,
+ float128_type,
+ NULL_TREE);
+ decl = add_builtin_function ("__builtin_copysignq", ftype,
+ IA64_BUILTIN_COPYSIGNQ, BUILT_IN_MD,
+ "__copysigntf3", NULL_TREE);
+ TREE_READONLY (decl) = 1;
}
else
/* Under HPUX, this is a synonym for "long double". */
emit_insn (gen_flushrs ());
return const0_rtx;
+ case IA64_BUILTIN_INFQ:
+ case IA64_BUILTIN_HUGE_VALQ:
+ {
+ REAL_VALUE_TYPE inf;
+ rtx tmp;
+
+ real_inf (&inf);
+ tmp = CONST_DOUBLE_FROM_REAL_VALUE (inf, mode);
+
+ tmp = validize_mem (force_const_mem (mode, tmp));
+
+ if (target == 0)
+ target = gen_reg_rtx (mode);
+
+ emit_move_insn (target, tmp);
+ return target;
+ }
+
+ case IA64_BUILTIN_FABSQ:
+ case IA64_BUILTIN_COPYSIGNQ:
+ return expand_call (exp, target, ignore);
+
default:
- break;
+ gcc_unreachable ();
}
return NULL_RTX;
most significant bits of the stack slot. */
enum direction
-ia64_hpux_function_arg_padding (enum machine_mode mode, tree type)
+ia64_hpux_function_arg_padding (enum machine_mode mode, const_tree type)
{
/* Exception to normal case for structures/unions/etc. */
/* We leave out _U_Qfmin, _U_Qfmax and _U_Qfabs since soft-fp in
glibc doesn't have them. */
}
+
+/* Use soft-fp. */
+
+static void
+ia64_soft_fp_init_libfuncs (void)
+{
+}
\f
/* For HPUX, it is illegal to have relocations in shared segments. */
HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
tree function)
{
- rtx this, insn, funexp;
+ rtx this_rtx, insn, funexp;
unsigned int this_parmno;
unsigned int this_regno;
rtx delta_rtx;
if (!TARGET_REG_NAMES)
reg_names[this_regno] = ia64_reg_numbers[this_parmno];
- this = gen_rtx_REG (Pmode, this_regno);
+ this_rtx = gen_rtx_REG (Pmode, this_regno);
/* Apply the constant offset, if required. */
delta_rtx = GEN_INT (delta);
REG_POINTER (tmp) = 1;
if (delta && satisfies_constraint_I (delta_rtx))
{
- emit_insn (gen_ptr_extend_plus_imm (this, tmp, delta_rtx));
+ emit_insn (gen_ptr_extend_plus_imm (this_rtx, tmp, delta_rtx));
delta = 0;
}
else
- emit_insn (gen_ptr_extend (this, tmp));
+ emit_insn (gen_ptr_extend (this_rtx, tmp));
}
if (delta)
{
emit_move_insn (tmp, delta_rtx);
delta_rtx = tmp;
}
- emit_insn (gen_adddi3 (this, this, delta_rtx));
+ emit_insn (gen_adddi3 (this_rtx, this_rtx, delta_rtx));
}
/* Apply the offset from the vtable, if required. */
{
rtx t = gen_rtx_REG (ptr_mode, 2);
REG_POINTER (t) = 1;
- emit_move_insn (t, gen_rtx_MEM (ptr_mode, this));
+ emit_move_insn (t, gen_rtx_MEM (ptr_mode, this_rtx));
if (satisfies_constraint_I (vcall_offset_rtx))
{
emit_insn (gen_ptr_extend_plus_imm (tmp, t, vcall_offset_rtx));
emit_insn (gen_ptr_extend (tmp, t));
}
else
- emit_move_insn (tmp, gen_rtx_MEM (Pmode, this));
+ emit_move_insn (tmp, gen_rtx_MEM (Pmode, this_rtx));
if (vcall_offset)
{
else
emit_move_insn (tmp, gen_rtx_MEM (Pmode, tmp));
- emit_insn (gen_adddi3 (this, this, tmp));
+ emit_insn (gen_adddi3 (this_rtx, this_rtx, tmp));
}
/* Generate a tail call to the target function. */
final_start_function (insn, file, 1);
final (insn, file, 1);
final_end_function ();
+ free_after_compilation (cfun);
reload_completed = 0;
epilogue_completed = 0;
return true;
case TFmode:
- return TARGET_HPUX;
+ return true;
default:
return false;
/* Return the mangling of TYPE if it is an extended fundamental type. */
static const char *
-ia64_mangle_type (tree type)
+ia64_mangle_type (const_tree type)
{
type = TYPE_MAIN_VARIANT (type);
/* Return the diagnostic message string if conversion from FROMTYPE to
TOTYPE is not allowed, NULL otherwise. */
static const char *
-ia64_invalid_conversion (tree fromtype, tree totype)
+ia64_invalid_conversion (const_tree fromtype, const_tree totype)
{
/* Reject nontrivial conversion to or from __fpreg. */
if (TYPE_MODE (fromtype) == RFmode
/* Return the diagnostic message string if the unary operation OP is
not permitted on TYPE, NULL otherwise. */
static const char *
-ia64_invalid_unary_op (int op, tree type)
+ia64_invalid_unary_op (int op, const_tree type)
{
/* Reject operations on __fpreg other than unary + or &. */
if (TYPE_MODE (type) == RFmode
/* Return the diagnostic message string if the binary operation OP is
not permitted on TYPE1 and TYPE2, NULL otherwise. */
static const char *
-ia64_invalid_binary_op (int op ATTRIBUTE_UNUSED, tree type1, tree type2)
+ia64_invalid_binary_op (int op ATTRIBUTE_UNUSED, const_tree type1, const_tree type2)
{
/* Reject operations on __fpreg. */
if (TYPE_MODE (type1) == RFmode || TYPE_MODE (type2) == RFmode)
set_param_value ("simultaneous-prefetches", 6);
set_param_value ("l1-cache-line-size", 32);
+ set_param_value("sched-mem-true-dep-cost", 4);
}
/* HP-UX version_id attribute.
return NULL_TREE;
}
+/* Target hook for c_mode_for_suffix. */
+
+static enum machine_mode
+ia64_c_mode_for_suffix (char suffix)
+{
+ if (suffix == 'q')
+ return TFmode;
+ if (suffix == 'w')
+ return XFmode;
+
+ return VOIDmode;
+}
+
#include "gt-ia64.h"