-/* Output routines for GCC for Hitachi / SuperH SH.
- Copyright (C) 1993, 1994, 1995, 1997, 1997, 1998, 1999, 2000, 2001, 2002, 2003
- Free Software Foundation, Inc.
+/* Output routines for GCC for Renesas / SuperH SH.
+ Copyright (C) 1993, 1994, 1995, 1997, 1997, 1998, 1999, 2000, 2001, 2002,
+ 2003 Free Software Foundation, Inc.
Contributed by Steve Chamberlain (sac@cygnus.com).
Improved by Jim Wilson (wilson@cygnus.com).
#include "real.h"
#include "langhooks.h"
#include "basic-block.h"
+#include "ra.h"
+#include "cfglayout.h"
int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch;
/* These are some macros to abstract register modes. */
#define CONST_OK_FOR_ADD(size) \
- (TARGET_SHMEDIA ? CONST_OK_FOR_P (size) : CONST_OK_FOR_I (size))
+ (TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size))
#define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi))
#define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3))
#define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3))
/* Provides the class number of the smallest class containing
reg number. */
-int regno_reg_class[FIRST_PSEUDO_REGISTER] =
+enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] =
{
R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
int assembler_dialect;
+static bool shmedia_space_reserved_for_target_registers;
+
static void split_branches PARAMS ((rtx));
static int branch_dest PARAMS ((rtx));
static void force_into PARAMS ((rtx, rtx));
static rtx find_barrier PARAMS ((int, rtx, rtx));
static int noncall_uses_reg PARAMS ((rtx, rtx, rtx *));
static rtx gen_block_redirect PARAMS ((rtx, int, int));
-static void output_stack_adjust PARAMS ((int, rtx, int, rtx (*) (rtx)));
+static void sh_reorg PARAMS ((void));
+static void output_stack_adjust (int, rtx, int, HARD_REG_SET *);
static rtx frame_insn PARAMS ((rtx));
static rtx push PARAMS ((int));
static void pop PARAMS ((int));
-static void push_regs PARAMS ((HOST_WIDE_INT *));
-static void calc_live_regs PARAMS ((int *, HOST_WIDE_INT *));
+static void push_regs PARAMS ((HARD_REG_SET *, int));
+static int calc_live_regs PARAMS ((HARD_REG_SET *));
static void mark_use PARAMS ((rtx, rtx *));
static HOST_WIDE_INT rounded_frame_size PARAMS ((int));
static rtx mark_constant_pool_use PARAMS ((rtx));
static bool sh_function_ok_for_sibcall PARAMS ((tree, tree));
static bool sh_cannot_modify_jumps_p PARAMS ((void));
+static int sh_target_reg_class (void);
+static bool sh_optimize_target_register_callee_saved (bool);
static bool sh_ms_bitfield_layout_p PARAMS ((tree));
-static void sh_encode_section_info PARAMS ((tree, int));
-static const char *sh_strip_name_encoding PARAMS ((const char *));
static void sh_init_builtins PARAMS ((void));
static void sh_media_init_builtins PARAMS ((void));
static rtx sh_expand_builtin PARAMS ((tree, rtx, rtx, enum machine_mode, int));
static void sh_output_mi_thunk PARAMS ((FILE *, tree, HOST_WIDE_INT,
HOST_WIDE_INT, tree));
+static void sh_file_start PARAMS ((void));
static int flow_dependent_p PARAMS ((rtx, rtx));
static void flow_dependent_p_1 PARAMS ((rtx, rtx, void *));
static int shiftcosts PARAMS ((rtx));
static bool sh_cannot_copy_insn_p PARAMS ((rtx));
static bool sh_rtx_costs PARAMS ((rtx, int, int, int *));
static int sh_address_cost PARAMS ((rtx));
+static int shmedia_target_regs_stack_space (HARD_REG_SET *);
+static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *);
+static int shmedia_target_regs_stack_adjust (HARD_REG_SET *);
+static int scavenge_reg (HARD_REG_SET *s);
+struct save_schedule_s;
+static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *,
+ struct save_schedule_s *, int);
\f
/* Initialize the GCC target structure. */
#undef TARGET_ATTRIBUTE_TABLE
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
+#undef TARGET_ASM_FILE_START
+#define TARGET_ASM_FILE_START sh_file_start
+#undef TARGET_ASM_FILE_START_FILE_DIRECTIVE
+#define TARGET_ASM_FILE_START_FILE_DIRECTIVE true
+
#undef TARGET_INSERT_ATTRIBUTES
#define TARGET_INSERT_ATTRIBUTES sh_insert_attributes
#undef TARGET_CANNOT_MODIFY_JUMPS_P
#define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p
+#undef TARGET_BRANCH_TARGET_REGISTER_CLASS
+#define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class
+#undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED
+#define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \
+ sh_optimize_target_register_callee_saved
#undef TARGET_MS_BITFIELD_LAYOUT_P
#define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p
-#undef TARGET_ENCODE_SECTION_INFO
-#define TARGET_ENCODE_SECTION_INFO sh_encode_section_info
-#undef TARGET_STRIP_NAME_ENCODING
-#define TARGET_STRIP_NAME_ENCODING sh_strip_name_encoding
-
#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS sh_init_builtins
#undef TARGET_EXPAND_BUILTIN
#undef TARGET_ADDRESS_COST
#define TARGET_ADDRESS_COST sh_address_cost
+#undef TARGET_MACHINE_DEPENDENT_REORG
+#define TARGET_MACHINE_DEPENDENT_REORG sh_reorg
+
#ifdef HAVE_AS_TLS
#undef TARGET_HAVE_TLS
#define TARGET_HAVE_TLS true
',' print LOCAL_LABEL_PREFIX
'@' print trap, rte or rts depending upon pragma interruptness
'#' output a nop if there is nothing to put in the delay slot
- ''' print likelyhood suffix (/u for unlikely).
+ ''' print likelihood suffix (/u for unlikely).
'O' print a constant without the #
'R' print the LSW of a dp value - changes if in little endian
'S' print the MSW of a dp value - changes if in little endian
entry_name = get_identifier ("__movstrSI12_i4");
- sym = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name));
+ sym = function_symbol (IDENTIFIER_POINTER (entry_name));
func_addr_rtx = copy_to_mode_reg (Pmode, sym);
force_into (XEXP (operands[0], 0), r4);
force_into (XEXP (operands[1], 0), r5);
entry_name = get_identifier (bytes & 4
? "__movstr_i4_odd"
: "__movstr_i4_even");
- sym = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name));
+ sym = function_symbol (IDENTIFIER_POINTER (entry_name));
func_addr_rtx = copy_to_mode_reg (Pmode, sym);
force_into (XEXP (operands[0], 0), r4);
force_into (XEXP (operands[1], 0), r5);
sprintf (entry, "__movstrSI%d", bytes);
entry_name = get_identifier (entry);
- sym = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name));
+ sym = function_symbol (IDENTIFIER_POINTER (entry_name));
func_addr_rtx = copy_to_mode_reg (Pmode, sym);
force_into (XEXP (operands[0], 0), r4);
force_into (XEXP (operands[1], 0), r5);
rtx r6 = gen_rtx_REG (SImode, 6);
entry_name = get_identifier ("__movstr");
- sym = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name));
+ sym = function_symbol (IDENTIFIER_POINTER (entry_name));
func_addr_rtx = copy_to_mode_reg (Pmode, sym);
force_into (XEXP (operands[0], 0), r4);
force_into (XEXP (operands[1], 0), r5);
&& ! sh_register_operand (operands[1], mode))
operands[1] = copy_to_mode_reg (mode, operands[1]);
+ if (GET_CODE (operands[0]) == MEM && ! memory_operand (operands[0], mode))
+ {
+ /* This is like change_address_1 (operands[0], mode, 0, 1) ,
+ except that we can't use that function because it is static. */
+ rtx new = change_address (operands[0], mode, 0);
+ MEM_COPY_ATTRIBUTES (new, operands[0]);
+ operands[0] = new;
+ }
+
/* This case can happen while generating code to move the result
of a library call to the target. Reject `st r0,@(rX,rY)' because
reload will fail to find a spill register for rX, since r0 is already
being used for the source. */
- else if (GET_CODE (operands[1]) == REG && REGNO (operands[1]) == 0
+ else if (refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0)
&& GET_CODE (operands[0]) == MEM
&& GET_CODE (XEXP (operands[0], 0)) == PLUS
&& GET_CODE (XEXP (XEXP (operands[0], 0), 1)) == REG)
return output_branchy_insn (NE, "bt\t%l9\\;fcmp/eq\t%1,%0", insn, operands);
}
\f
-/* Output to FILE the start of the assembler file. */
-
-void
-output_file_start (file)
- FILE *file;
-{
- output_file_directive (file, main_input_filename);
+/* Output the start of the assembler file. */
- /* Switch to the data section so that the coffsem symbol
- isn't in the text section. */
- data_section ();
+static void
+sh_file_start ()
+{
+ default_file_start ();
+
+ if (TARGET_ELF)
+ /* We need to show the text section with the proper
+ attributes as in TEXT_SECTION_ASM_OP, before dwarf2out
+ emits it without attributes in TEXT_SECTION, else GAS
+ will complain. We can teach GAS specifically about the
+ default attributes for our choice of text section, but
+ then we would have to change GAS again if/when we change
+ the text section name. */
+ fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP);
+ else
+ /* Switch to the data section so that the coffsem symbol
+ isn't in the text section. */
+ data_section ();
if (TARGET_LITTLE_ENDIAN)
- fprintf (file, "\t.little\n");
+ fputs ("\t.little\n", asm_out_file);
- if (TARGET_SHCOMPACT)
- fprintf (file, "\t.mode\tSHcompact\n");
- else if (TARGET_SHMEDIA)
- fprintf (file, "\t.mode\tSHmedia\n\t.abi\t%i\n",
- TARGET_SHMEDIA64 ? 64 : 32);
+ if (!TARGET_ELF)
+ {
+ if (TARGET_SHCOMPACT)
+ fputs ("\t.mode\tSHcompact\n", asm_out_file);
+ else if (TARGET_SHMEDIA)
+ fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n",
+ TARGET_SHMEDIA64 ? 64 : 32);
+ }
}
\f
/* Check if PAT includes UNSPEC_CALLER unspec pattern. */
if (TARGET_SHMEDIA)
{
if ((GET_CODE (XEXP (x, 1)) == CONST_INT
- && CONST_OK_FOR_J (INTVAL (XEXP (x, 1))))
- || EXTRA_CONSTRAINT_S (XEXP (x, 1)))
+ && CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1))))
+ || EXTRA_CONSTRAINT_C16 (XEXP (x, 1)))
return 1;
else
return 2;
/* These constants are single cycle extu.[bw] instructions. */
if (i == 0xff || i == 0xffff)
return 1;
- /* Constants that can be used in an and immediate instruction is a single
+ /* Constants that can be used in an and immediate instruction in a single
cycle, but this requires r0, so make it a little more expensive. */
- if (CONST_OK_FOR_L (i))
+ if (CONST_OK_FOR_K08 (i))
return 2;
/* Constants that can be loaded with a mov immediate and an and.
This case is probably unnecessary. */
- if (CONST_OK_FOR_I (i))
+ if (CONST_OK_FOR_I08 (i))
return 2;
/* Any other constants requires a 2 cycle pc-relative load plus an and.
This case is probably unnecessary. */
return TARGET_SHMEDIA64 ? 5 : 3;
case CONST_INT:
- if (CONST_OK_FOR_J (INTVAL (XEXP (x, 1))))
+ if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1))))
return 2;
- else if (CONST_OK_FOR_J (INTVAL (XEXP (x, 1)) >> 16))
+ else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16))
return 3;
- else if (CONST_OK_FOR_J ((INTVAL (XEXP (x, 1)) >> 16) >> 16))
+ else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16))
return 4;
/* Fall through. */
*total = 0;
else if ((outer_code == IOR || outer_code == XOR
|| outer_code == PLUS)
- && CONST_OK_FOR_P (INTVAL (x)))
+ && CONST_OK_FOR_I10 (INTVAL (x)))
*total = 0;
- else if (CONST_OK_FOR_J (INTVAL (x)))
+ else if (CONST_OK_FOR_I16 (INTVAL (x)))
*total = COSTS_N_INSNS (outer_code != SET);
- else if (CONST_OK_FOR_J (INTVAL (x) >> 16))
+ else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16))
*total = COSTS_N_INSNS (2);
- else if (CONST_OK_FOR_J ((INTVAL (x) >> 16) >> 16))
+ else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16))
*total = COSTS_N_INSNS (3);
else
*total = COSTS_N_INSNS (4);
return true;
}
- if (CONST_OK_FOR_I (INTVAL (x)))
+ if (CONST_OK_FOR_I08 (INTVAL (x)))
*total = 0;
else if ((outer_code == AND || outer_code == IOR || outer_code == XOR)
- && CONST_OK_FOR_L (INTVAL (x)))
+ && CONST_OK_FOR_K08 (INTVAL (x)))
*total = 1;
else
*total = 8;
emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]);
sprintf (func, "__ashiftrt_r4_%d", value);
func_name = get_identifier (func);
- sym = gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (func_name));
+ sym = function_symbol (IDENTIFIER_POINTER (func_name));
emit_move_insn (wrk, sym);
emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk));
emit_move_insn (operands[0], gen_rtx_REG (SImode, 4));
{
if (i > right)
break;
- if (! CONST_OK_FOR_L (mask >> i))
+ if (! CONST_OK_FOR_K08 (mask >> i))
continue;
cost = (i != 0) + 2 + ext_shift_insns[left + i];
if (cost < best_cost)
{
if (i > right)
break;
- cost = (i != 0) + (CONST_OK_FOR_I (mask >> i) ? 2 : 3)
+ cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3)
+ (can_ext ? ext_shift_insns : shift_insns)[left + i];
if (cost < best_cost)
{
best = 4 - can_ext;
best_cost = cost;
best_right = i;
- best_len = cost - 1 - ! CONST_OK_FOR_I (mask >> i);
+ best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i);
}
}
if (GET_CODE (sym) != SYMBOL_REF)
abort ();
- XSTR (sym, 0) = concat (SH_DATALABEL_ENCODING, XSTR (sym, 0), NULL);
-
return sym;
}
/* Nonzero if the insn is a move instruction which needs to be fixed. */
/* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the
- CONST_DOUBLE input value is CONST_OK_FOR_I. For a SFmode move, we don't
- need to fix it if the input value is CONST_OK_FOR_I. */
+ CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't
+ need to fix it if the input value is CONST_OK_FOR_I08. */
static int
broken_move (insn)
&& GET_CODE (SET_DEST (pat)) == REG
&& FP_REGISTER_P (REGNO (SET_DEST (pat))))
&& (GET_CODE (SET_SRC (pat)) != CONST_INT
- || ! CONST_OK_FOR_I (INTVAL (SET_SRC (pat)))))
+ || ! CONST_OK_FOR_I08 (INTVAL (SET_SRC (pat)))))
return 1;
}
rtx next = next_active_insn (next_active_insn (dest));
if (next && GET_CODE (next) == JUMP_INSN
&& GET_CODE (PATTERN (next)) == SET
- && recog_memoized (next) == CODE_FOR_jump)
+ && recog_memoized (next) == CODE_FOR_jump_compact)
{
dest = JUMP_LABEL (next);
if (dest
NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and
the jump. */
- INSN_SCOPE (insn) = INSN_SCOPE (jump);
+ INSN_LOCATOR (insn) = INSN_LOCATOR (jump);
INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch;
return insn;
}
jump = emit_jump_insn_after (gen_return (), insn);
/* Emit a barrier so that reorg knows that any following instructions
are not reachable via a fall-through path.
- But don't do this when not optimizing, since we wouldn't supress the
+ But don't do this when not optimizing, since we wouldn't suppress the
alignment for the barrier then, and could end up with out-of-range
pc-relative loads. */
if (optimize)
|| (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))),
(INSN_P (x)
&& (INSN_CODE (x) == CODE_FOR_block_branch_redirect
- || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch))))
+ || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch
+ || INSN_CODE (x) == CODE_FOR_stuff_delay_slot))))
{
rtx pat = PATTERN (prev);
if (GET_CODE (pat) == PARALLEL)
return align_loops_log;
}
-/* Exported to toplev.c.
-
- Do a final pass over the function, just before delayed branch
+/* Do a final pass over the function, just before delayed branch
scheduling. */
-void
-machine_dependent_reorg (first)
- rtx first;
+static void
+sh_reorg ()
{
- rtx insn, mova = NULL_RTX;
+ rtx first, insn, mova = NULL_RTX;
int num_mova;
rtx r0_rtx = gen_rtx_REG (Pmode, 0);
rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx);
+ first = get_insns ();
+
/* We must split call insns before introducing `mova's. If we're
optimizing, they'll have already been split. Otherwise, make
sure we don't split them too late. */
|| ((beyond = next_active_insn (beyond))
&& GET_CODE (beyond) == JUMP_INSN))
&& GET_CODE (PATTERN (beyond)) == SET
- && recog_memoized (beyond) == CODE_FOR_jump
+ && recog_memoized (beyond) == CODE_FOR_jump_compact
&& ((INSN_ADDRESSES
(INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0)))
- INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252)
if ((GET_CODE (next) == JUMP_INSN
|| GET_CODE (next = next_active_insn (next)) == JUMP_INSN)
&& GET_CODE (PATTERN (next)) == SET
- && recog_memoized (next) == CODE_FOR_jump
+ && recog_memoized (next) == CODE_FOR_jump_compact
&& ((INSN_ADDRESSES
(INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0)))
- INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252)
static int extra_push;
-/* Adjust the stack by SIZE bytes. REG holds the rtl of the register
- to be adjusted, and TEMP, if nonnegative, holds the register number
- of a general register that we may clobber. */
+/* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be
+ adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's
+ for an epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET
+ of all the registers that are about to be restored, and hence dead. */
static void
-output_stack_adjust (size, reg, temp, emit_fn)
- int size;
- rtx reg;
- int temp;
- rtx (*emit_fn) PARAMS ((rtx));
+output_stack_adjust (int size, rtx reg, int epilogue_p,
+ HARD_REG_SET *live_regs_mask)
{
+ rtx (*emit_fn) (rtx) = epilogue_p ? &emit_insn : &frame_insn;
if (size)
{
HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT;
{
rtx const_reg;
rtx insn;
+ int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1);
+ int i;
/* If TEMP is invalid, we could temporarily save a general
register to MACL. However, there is currently no need
to handle this case, so just abort when we see it. */
+ if (current_function_interrupt
+ || ! call_used_regs[temp] || fixed_regs[temp])
+ temp = -1;
+ if (temp < 0 && ! current_function_interrupt)
+ {
+ HARD_REG_SET temps;
+ COPY_HARD_REG_SET (temps, call_used_reg_set);
+ AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set);
+ if (epilogue_p)
+ {
+ for (i = 0; i < HARD_REGNO_NREGS (FIRST_RET_REG, DImode); i++)
+ CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i);
+ if (current_function_calls_eh_return)
+ {
+ CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO);
+ for (i = 0; i <= 3; i++)
+ CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i));
+ }
+ }
+ else
+ {
+ for (i = FIRST_PARM_REG;
+ i < FIRST_PARM_REG + NPARM_REGS (SImode); i++)
+ CLEAR_HARD_REG_BIT (temps, i);
+ if (current_function_needs_context)
+ CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM);
+ }
+ temp = scavenge_reg (&temps);
+ }
+ if (temp < 0 && live_regs_mask)
+ temp = scavenge_reg (live_regs_mask);
if (temp < 0)
abort ();
const_reg = gen_rtx_REG (GET_MODE (reg), temp);
emit_insn (GEN_MOV (const_reg, GEN_INT (size)));
insn = emit_fn (GEN_ADD3 (reg, reg, const_reg));
}
- if (emit_fn == frame_insn)
+ if (! epilogue_p)
REG_NOTES (insn)
= (gen_rtx_EXPR_LIST
(REG_FRAME_RELATED_EXPR,
/* Generate code to push the regs specified in the mask. */
static void
-push_regs (mask)
- HOST_WIDE_INT *mask;
+push_regs (mask, interrupt_handler)
+ HARD_REG_SET *mask;
+ int interrupt_handler;
{
int i;
+ int skip_fpscr = 0;
/* Push PR last; this gives better latencies after the prologue, and
candidates for the return delay slot when there are no general
registers pushed. */
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
- if (i != PR_REG && mask[i / 32] & (1 << (i % 32)))
- push (i);
- if (mask[PR_REG / 32] & (1 << (PR_REG % 32)))
+ {
+ /* If this is an interrupt handler, and the SZ bit varies,
+ and we have to push any floating point register, we need
+ to switch to the correct precision first. */
+ if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD
+ && hard_regs_intersect_p (mask, ®_class_contents[DF_REGS]))
+ {
+ HARD_REG_SET unsaved;
+
+ push (FPSCR_REG);
+ COMPL_HARD_REG_SET(unsaved, *mask);
+ fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved);
+ skip_fpscr = 1;
+ }
+ if (i != PR_REG
+ && (i != FPSCR_REG || ! skip_fpscr)
+ && TEST_HARD_REG_BIT (*mask, i))
+ push (i);
+ }
+ if (TEST_HARD_REG_BIT (*mask, PR_REG))
push (PR_REG);
}
+/* Calculate how much extra space is needed to save all callee-saved
+ target registers.
+ LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */
+
+static int
+shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask)
+{
+ int reg;
+ int stack_space = 0;
+ int interrupt_handler = sh_cfun_interrupt_handler_p ();
+
+ for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--)
+ if ((! call_used_regs[reg] || interrupt_handler)
+ && ! TEST_HARD_REG_BIT (*live_regs_mask, reg))
+ /* Leave space to save this target register on the stack,
+ in case target register allocation wants to use it. */
+ stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg));
+ return stack_space;
+}
+
+/* Decide whether we should reserve space for callee-save target registers,
+ in case target register allocation wants to use them. REGS_SAVED is
+ the space, in bytes, that is already required for register saves.
+ LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */
+
+static int
+shmedia_reserve_space_for_target_registers_p (int regs_saved,
+ HARD_REG_SET *live_regs_mask)
+{
+ if (optimize_size)
+ return 0;
+ return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved;
+}
+
+/* Decide how much space to reserve for callee-save target registers
+ in case target register allocation wants to use them.
+ LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */
+
+static int
+shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask)
+{
+ if (shmedia_space_reserved_for_target_registers)
+ return shmedia_target_regs_stack_space (live_regs_mask);
+ else
+ return 0;
+}
+
/* Work out the registers which need to be saved, both as a mask and a
- count of saved words.
+ count of saved words. Return the count.
If doing a pragma interrupt function, then push all regs used by the
function, and if we call another function (we can tell by looking at PR),
make sure that all the regs it clobbers are safe too. */
-static void
-calc_live_regs (count_ptr, live_regs_mask)
- int *count_ptr;
- HOST_WIDE_INT *live_regs_mask;
+static int
+calc_live_regs (live_regs_mask)
+ HARD_REG_SET *live_regs_mask;
{
int reg;
int count;
int interrupt_handler;
- int pr_live;
+ int pr_live, has_call;
interrupt_handler = sh_cfun_interrupt_handler_p ();
- for (count = 0; 32 * count < FIRST_PSEUDO_REGISTER; count++)
- live_regs_mask[count] = 0;
+ CLEAR_HARD_REG_SET (*live_regs_mask);
+ if (TARGET_SH4 && TARGET_FMOVD && interrupt_handler
+ && regs_ever_live[FPSCR_REG])
+ target_flags &= ~FPU_SINGLE_BIT;
/* If we can save a lot of saves by switching to double mode, do that. */
- if (TARGET_SH4 && TARGET_FMOVD && TARGET_FPU_SINGLE)
+ else if (TARGET_SH4 && TARGET_FMOVD && TARGET_FPU_SINGLE)
for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2)
if (regs_ever_live[reg] && regs_ever_live[reg+1]
&& (! call_used_regs[reg] || (interrupt_handler && ! pragma_trapa))
the initial value can become the PR_MEDIA_REG hard register, as seen for
execute/20010122-1.c:test9. */
if (TARGET_SHMEDIA)
- pr_live = regs_ever_live[PR_MEDIA_REG];
+ /* ??? this function is called from initial_elimination_offset, hence we
+ can't use the result of sh_media_register_for_return here. */
+ pr_live = sh_pr_n_sets ();
else
{
rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG);
? (GET_CODE (pr_initial) != REG
|| REGNO (pr_initial) != (PR_REG))
: regs_ever_live[PR_REG]);
+ /* For Shcompact, if not optimizing, we end up with a memory reference
+ using the return address pointer for __builtin_return_address even
+ though there is no actual need to put the PR register on the stack. */
+ pr_live |= regs_ever_live[RETURN_ADDRESS_POINTER_REGNUM];
}
/* Force PR to be live if the prologue has to call the SHmedia
argument decoder or register saver. */
& ~ CALL_COOKIE_RET_TRAMP (1))
|| current_function_has_nonlocal_label))
pr_live = 1;
+ has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live;
for (count = 0, reg = FIRST_PSEUDO_REGISTER - 1; reg >= 0; reg--)
{
if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)
(regs_ever_live[reg]
|| (call_used_regs[reg]
&& (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG)
- && pr_live))
+ && has_call)
+ || (has_call && REGISTER_NATURAL_MODE (reg) == SImode
+ && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg))))
&& reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM
&& reg != RETURN_ADDRESS_POINTER_REGNUM
&& reg != T_REG && reg != GBR_REG
(TARGET_SHCOMPACT
&& flag_pic
&& current_function_args_info.call_cookie
- && reg == PIC_OFFSET_TABLE_REGNUM)
+ && reg == (int) PIC_OFFSET_TABLE_REGNUM)
|| (regs_ever_live[reg] && ! call_used_regs[reg])
|| (current_function_calls_eh_return
- && (reg == EH_RETURN_DATA_REGNO (0)
- || reg == EH_RETURN_DATA_REGNO (1)
- || reg == EH_RETURN_DATA_REGNO (2)
- || reg == EH_RETURN_DATA_REGNO (3)))))
+ && (reg == (int) EH_RETURN_DATA_REGNO (0)
+ || reg == (int) EH_RETURN_DATA_REGNO (1)
+ || reg == (int) EH_RETURN_DATA_REGNO (2)
+ || reg == (int) EH_RETURN_DATA_REGNO (3)))))
{
- live_regs_mask[reg / 32] |= 1 << (reg % 32);
+ SET_HARD_REG_BIT (*live_regs_mask, reg);
count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg));
if ((TARGET_SH4 || TARGET_SH5) && TARGET_FMOVD
{
if (! TARGET_FPU_SINGLE && ! regs_ever_live[reg ^ 1])
{
- live_regs_mask[(reg ^ 1) / 32] |= 1 << ((reg ^ 1) % 32);
+ SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1));
count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1));
}
}
}
}
}
+ /* If we have a target register optimization pass after prologue / epilogue
+ threading, we need to assume all target registers will be live even if
+ they aren't now. */
+ if (flag_branch_target_load_optimize2
+ && TARGET_SAVE_ALL_TARGET_REGS
+ && shmedia_space_reserved_for_target_registers)
+ for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--)
+ if ((! call_used_regs[reg] || interrupt_handler)
+ && ! TEST_HARD_REG_BIT (*live_regs_mask, reg))
+ {
+ SET_HARD_REG_BIT (*live_regs_mask, reg);
+ count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg));
+ }
+ /* If this is an interrupt handler, we don't have any call-clobbered
+ registers we can conveniently use for target register save/restore.
+ Make sure we save at least one general purpose register when we need
+ to save target registers. */
+ if (interrupt_handler
+ && hard_regs_intersect_p (live_regs_mask,
+ ®_class_contents[TARGET_REGS])
+ && ! hard_regs_intersect_p (live_regs_mask,
+ ®_class_contents[GENERAL_REGS]))
+ {
+ SET_HARD_REG_BIT (*live_regs_mask, R0_REG);
+ count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG));
+ }
- *count_ptr = count;
+ return count;
}
/* Code to generate prologue and epilogue sequences */
if (! current_function_is_leaf)
return -1;
+ if (lookup_attribute ("interrupt_handler",
+ DECL_ATTRIBUTES (current_function_decl)))
+ return -1;
tr0_used = flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM];
return -1;
}
+/* The maximum registers we need to save are:
+ - 62 general purpose registers (r15 is stack pointer, r63 is zero)
+ - 32 floating point registers (for each pair, we save none,
+ one single precision value, or a double precision value).
+ - 8 target registers
+ - add 1 entry for a delimiter. */
+#define MAX_SAVED_REGS (62+32+8)
+
+typedef struct save_entry_s
+{
+ unsigned char reg;
+ unsigned char mode;
+ short offset;
+} save_entry;
+
+#define MAX_TEMPS 4
+
+/* There will be a delimiter entry with VOIDmode both at the start and the
+ end of a filled in schedule. The end delimiter has the offset of the
+ save with the smallest (i.e. most negative) offset. */
+typedef struct save_schedule_s
+{
+ save_entry entries[MAX_SAVED_REGS + 2];
+ int temps[MAX_TEMPS+1];
+} save_schedule;
+
+/* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero,
+ use reverse order. Returns the last entry written to (not counting
+ the delimiter). OFFSET_BASE is a number to be added to all offset
+ entries. */
+
+static save_entry *
+sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule,
+ int offset_base)
+{
+ int align, i;
+ save_entry *entry = schedule->entries;
+ int tmpx = 0;
+ int offset;
+
+ if (! current_function_interrupt)
+ for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++)
+ if (call_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG
+ && ! FUNCTION_ARG_REGNO_P (i)
+ && i != FIRST_RET_REG
+ && ! (current_function_needs_context && i == STATIC_CHAIN_REGNUM)
+ && ! (current_function_calls_eh_return
+ && (i == EH_RETURN_STACKADJ_REGNO
+ || ((unsigned)i <= EH_RETURN_DATA_REGNO (0)
+ && (unsigned)i >= EH_RETURN_DATA_REGNO (3)))))
+ schedule->temps[tmpx++] = i;
+ entry->reg = -1;
+ entry->mode = VOIDmode;
+ entry->offset = offset_base;
+ entry++;
+ /* We loop twice: first, we save 8-byte aligned registers in the
+ higher addresses, that are known to be aligned. Then, we
+ proceed to saving 32-bit registers that don't need 8-byte
+ alignment.
+ If this is an interrupt function, all registers that need saving
+ need to be saved in full. moreover, we need to postpone saving
+ target registers till we have saved some general purpose registers
+ we can then use as scratch registers. */
+ offset = offset_base;
+ for (align = 1; align >= 0; align--)
+ {
+ for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
+ if (TEST_HARD_REG_BIT (*live_regs_mask, i))
+ {
+ enum machine_mode mode = REGISTER_NATURAL_MODE (i);
+ int reg = i;
+
+ if (current_function_interrupt)
+ {
+ if (TARGET_REGISTER_P (i))
+ continue;
+ if (GENERAL_REGISTER_P (i))
+ mode = DImode;
+ }
+ if (mode == SFmode && (i % 2) == 1
+ && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i)
+ && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1))))
+ {
+ mode = DFmode;
+ i--;
+ reg--;
+ }
+
+ /* If we're doing the aligned pass and this is not aligned,
+ or we're doing the unaligned pass and this is aligned,
+ skip it. */
+ if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0)
+ != align)
+ continue;
+
+ if (current_function_interrupt
+ && GENERAL_REGISTER_P (i)
+ && tmpx < MAX_TEMPS)
+ schedule->temps[tmpx++] = i;
+
+ offset -= GET_MODE_SIZE (mode);
+ entry->reg = i;
+ entry->mode = mode;
+ entry->offset = offset;
+ entry++;
+ }
+ if (align && current_function_interrupt)
+ for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--)
+ if (TEST_HARD_REG_BIT (*live_regs_mask, i))
+ {
+ offset -= GET_MODE_SIZE (DImode);
+ entry->reg = i;
+ entry->mode = DImode;
+ entry->offset = offset;
+ entry++;
+ }
+ }
+ entry->reg = -1;
+ entry->mode = VOIDmode;
+ entry->offset = offset;
+ schedule->temps[tmpx] = -1;
+ return entry - 1;
+}
+
void
sh_expand_prologue ()
{
- HOST_WIDE_INT live_regs_mask[(FIRST_PSEUDO_REGISTER + 31) / 32];
+ HARD_REG_SET live_regs_mask;
int d, i;
int d_rounding = 0;
int save_flags = target_flags;
and partially on the stack, e.g. a large structure. */
output_stack_adjust (-current_function_pretend_args_size
- current_function_args_info.stack_regs * 8,
- stack_pointer_rtx, TARGET_SH5 ? 0 : 1, frame_insn);
+ stack_pointer_rtx, 0, NULL);
extra_push = 0;
rtx insn = emit_move_insn (gen_rtx_REG (DImode, tr),
gen_rtx_REG (DImode, PR_MEDIA_REG));
+ /* ??? We should suppress saving pr when we don't need it, but this
+ is tricky because of builtin_return_address. */
+
/* If this function only exits with sibcalls, this copy
will be flagged as dead. */
REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
if (sp_switch)
emit_insn (gen_sp_switch_1 ());
- calc_live_regs (&d, live_regs_mask);
+ d = calc_live_regs (&live_regs_mask);
/* ??? Maybe we could save some switching if we can move a mode switch
that already happens to be at the function start into the prologue. */
- if (target_flags != save_flags)
+ if (target_flags != save_flags && ! current_function_interrupt)
emit_insn (gen_toggle_sz ());
if (TARGET_SH5)
{
- int i;
- int offset;
- int align;
- rtx r0 = gen_rtx_REG (Pmode, R0_REG);
+ int offset_base, offset;
+ rtx r0 = NULL_RTX;
int offset_in_r0 = -1;
int sp_in_r0 = 0;
-
- if (d % (STACK_BOUNDARY / BITS_PER_UNIT))
+ int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask);
+ int total_size, save_size;
+ save_schedule schedule;
+ save_entry *entry;
+ int *tmp_pnt;
+
+ if (call_used_regs[R0_REG] && ! fixed_regs[R0_REG]
+ && ! current_function_interrupt)
+ r0 = gen_rtx_REG (Pmode, R0_REG);
+
+ /* D is the actual number of bytes that we need for saving registers,
+ however, in initial_elimination_offset we have committed to using
+ an additional TREGS_SPACE amount of bytes - in order to keep both
+ addresses to arguments supplied by the caller and local variables
+ valid, we must keep this gap. Place it between the incoming
+ arguments and the actually saved registers in a bid to optimize
+ locality of reference. */
+ total_size = d + tregs_space;
+ total_size += rounded_frame_size (total_size);
+ save_size = total_size - rounded_frame_size (d);
+ if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT))
d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT)
- - d % (STACK_BOUNDARY / BITS_PER_UNIT));
-
- offset = d + d_rounding;
- output_stack_adjust (-offset, stack_pointer_rtx, 1, frame_insn);
-
- /* We loop twice: first, we save 8-byte aligned registers in the
- higher addresses, that are known to be aligned. Then, we
- proceed to saving 32-bit registers that don't need 8-byte
- alignment. */
- /* Note that if you change this code in a way that affects where
- the return register is saved, you have to update not only
- sh_expand_epilogue, but also sh_set_return_address. */
- for (align = 1; align >= 0; align--)
- for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
- if (live_regs_mask[i/32] & (1 << (i % 32)))
- {
- enum machine_mode mode = REGISTER_NATURAL_MODE (i);
- int reg = i;
- rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX;
-
- if (mode == SFmode && (i % 2) == 1
- && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i)
- && (live_regs_mask[(i ^ 1) / 32] & (1 << ((i ^ 1) % 32))))
- {
- mode = DFmode;
- i--;
- reg--;
- }
-
- /* If we're doing the aligned pass and this is not aligned,
- or we're doing the unaligned pass and this is aligned,
- skip it. */
- if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT)
- == 0) != align)
- continue;
+ - save_size % (STACK_BOUNDARY / BITS_PER_UNIT));
+
+ /* If adjusting the stack in a single step costs nothing extra, do so.
+ I.e. either if a single addi is enough, or we need a movi anyway,
+ and we don't exceed the maximum offset range (the test for the
+ latter is conservative for simplicity). */
+ if (TARGET_SHMEDIA
+ && (CONST_OK_FOR_I10 (-total_size)
+ || (! CONST_OK_FOR_I10 (-(save_size + d_rounding))
+ && total_size <= 2044)))
+ d_rounding = total_size - save_size;
+
+ offset_base = d + d_rounding;
+
+ output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx,
+ 0, NULL);
+
+ sh5_schedule_saves (&live_regs_mask, &schedule, offset_base);
+ tmp_pnt = schedule.temps;
+ for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++)
+ {
+ enum machine_mode mode = entry->mode;
+ int reg = entry->reg;
+ rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX;
- offset -= GET_MODE_SIZE (mode);
+ offset = entry->offset;
- reg_rtx = gen_rtx_REG (mode, reg);
+ reg_rtx = gen_rtx_REG (mode, reg);
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- GEN_INT (offset)));
+ mem_rtx = gen_rtx_MEM (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
- GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_pre_dec);
+ GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_pre_dec);
- mem_rtx = NULL_RTX;
+ if (! r0)
+ abort ();
+ mem_rtx = NULL_RTX;
- try_pre_dec:
- do
- if (HAVE_PRE_DECREMENT
- && (offset_in_r0 - offset == GET_MODE_SIZE (mode)
- || mem_rtx == NULL_RTX
- || i == PR_REG || SPECIAL_REGISTER_P (i)))
- {
- pre_dec = gen_rtx_MEM (mode,
- gen_rtx_PRE_DEC (Pmode, r0));
+ try_pre_dec:
+ do
+ if (HAVE_PRE_DECREMENT
+ && (offset_in_r0 - offset == GET_MODE_SIZE (mode)
+ || mem_rtx == NULL_RTX
+ || reg == PR_REG || SPECIAL_REGISTER_P (reg)))
+ {
+ pre_dec = gen_rtx_MEM (mode,
+ gen_rtx_PRE_DEC (Pmode, r0));
- GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (pre_dec, 0),
- pre_dec_ok);
+ GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (pre_dec, 0),
+ pre_dec_ok);
- pre_dec = NULL_RTX;
+ pre_dec = NULL_RTX;
- break;
+ break;
- pre_dec_ok:
- mem_rtx = NULL_RTX;
- offset += GET_MODE_SIZE (mode);
- }
- while (0);
+ pre_dec_ok:
+ mem_rtx = NULL_RTX;
+ offset += GET_MODE_SIZE (mode);
+ }
+ while (0);
- if (mem_rtx != NULL_RTX)
- goto addr_ok;
+ if (mem_rtx != NULL_RTX)
+ goto addr_ok;
- if (offset_in_r0 == -1)
- {
- emit_move_insn (r0, GEN_INT (offset));
- offset_in_r0 = offset;
- }
- else if (offset != offset_in_r0)
+ if (offset_in_r0 == -1)
+ {
+ emit_move_insn (r0, GEN_INT (offset));
+ offset_in_r0 = offset;
+ }
+ else if (offset != offset_in_r0)
+ {
+ emit_move_insn (r0,
+ gen_rtx_PLUS
+ (Pmode, r0,
+ GEN_INT (offset - offset_in_r0)));
+ offset_in_r0 += offset - offset_in_r0;
+ }
+
+ if (pre_dec != NULL_RTX)
+ {
+ if (! sp_in_r0)
{
emit_move_insn (r0,
gen_rtx_PLUS
- (Pmode, r0,
- GEN_INT (offset - offset_in_r0)));
- offset_in_r0 += offset - offset_in_r0;
+ (Pmode, r0, stack_pointer_rtx));
+ sp_in_r0 = 1;
}
-
- if (pre_dec != NULL_RTX)
- {
- if (! sp_in_r0)
- {
- emit_move_insn (r0,
- gen_rtx_PLUS
- (Pmode, r0, stack_pointer_rtx));
- sp_in_r0 = 1;
- }
- offset -= GET_MODE_SIZE (mode);
- offset_in_r0 -= GET_MODE_SIZE (mode);
+ offset -= GET_MODE_SIZE (mode);
+ offset_in_r0 -= GET_MODE_SIZE (mode);
- mem_rtx = pre_dec;
- }
- else if (sp_in_r0)
- mem_rtx = gen_rtx_MEM (mode, r0);
- else
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- r0));
-
- /* We must not use an r0-based address for target-branch
- registers or for special registers without pre-dec
- memory addresses, since we store their values in r0
- first. */
- if (TARGET_REGISTER_P (i)
- || ((i == PR_REG || SPECIAL_REGISTER_P (i))
- && mem_rtx != pre_dec))
- abort ();
-
- addr_ok:
- if (TARGET_REGISTER_P (i)
- || ((i == PR_REG || SPECIAL_REGISTER_P (i))
- && mem_rtx != pre_dec))
- {
- rtx r0mode = gen_rtx_REG (GET_MODE (reg_rtx), R0_REG);
+ mem_rtx = pre_dec;
+ }
+ else if (sp_in_r0)
+ mem_rtx = gen_rtx_MEM (mode, r0);
+ else
+ mem_rtx = gen_rtx_MEM (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ r0));
+
+ /* We must not use an r0-based address for target-branch
+ registers or for special registers without pre-dec
+ memory addresses, since we store their values in r0
+ first. */
+ if (TARGET_REGISTER_P (reg)
+ || ((reg == PR_REG || SPECIAL_REGISTER_P (reg))
+ && mem_rtx != pre_dec))
+ abort ();
- emit_move_insn (r0mode, reg_rtx);
+ addr_ok:
+ if (TARGET_REGISTER_P (reg)
+ || ((reg == PR_REG || SPECIAL_REGISTER_P (reg))
+ && mem_rtx != pre_dec))
+ {
+ rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt);
+
+ emit_move_insn (tmp_reg, reg_rtx);
+ if (REGNO (tmp_reg) == R0_REG)
+ {
offset_in_r0 = -1;
sp_in_r0 = 0;
-
- reg_rtx = r0mode;
+ if (refers_to_regno_p (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0))
+ abort ();
}
- emit_move_insn (mem_rtx, reg_rtx);
+ if (*++tmp_pnt <= 0)
+ tmp_pnt = schedule.temps;
+
+ reg_rtx = tmp_reg;
}
+ {
+ rtx insn;
+
+ /* Mark as interesting for dwarf cfi generator */
+ insn = emit_move_insn (mem_rtx, reg_rtx);
+ RTX_FRAME_RELATED_P (insn) = 1;
+
+ if (TARGET_SHCOMPACT && (offset_in_r0 != -1))
+ {
+ rtx reg_rtx = gen_rtx_REG (mode, reg);
+ rtx set, note_rtx;
+ rtx mem_rtx = gen_rtx_MEM (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
+
+ set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx);
+ note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set,
+ REG_NOTES (insn));
+ REG_NOTES (insn) = note_rtx;
+ }
+ }
+ }
- if (offset != d_rounding)
+ if (entry->offset != d_rounding)
abort ();
}
else
- push_regs (live_regs_mask);
+ push_regs (&live_regs_mask, current_function_interrupt);
if (flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM])
{
if (SHMEDIA_REGS_STACK_ADJUST ())
{
emit_move_insn (gen_rtx_REG (Pmode, R0_REG),
- gen_rtx_SYMBOL_REF (Pmode,
- TARGET_FPU_ANY
- ? "__GCC_push_shmedia_regs"
- : "__GCC_push_shmedia_regs_nofpu"));
+ function_symbol (TARGET_FPU_ANY
+ ? "__GCC_push_shmedia_regs"
+ : "__GCC_push_shmedia_regs_nofpu"));
/* This must NOT go through the PLT, otherwise mach and macl
may be clobbered. */
emit_insn (gen_shmedia_save_restore_regs_compact
(GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ())));
}
- if (target_flags != save_flags)
+ if (target_flags != save_flags && ! current_function_interrupt)
{
rtx insn = emit_insn (gen_toggle_sz ());
target_flags = save_flags;
output_stack_adjust (-rounded_frame_size (d) + d_rounding,
- stack_pointer_rtx, TARGET_SH5 ? 0 : 1, frame_insn);
+ stack_pointer_rtx, 0, NULL);
if (frame_pointer_needed)
frame_insn (GEN_MOV (frame_pointer_rtx, stack_pointer_rtx));
/* This must NOT go through the PLT, otherwise mach and macl
may be clobbered. */
emit_move_insn (gen_rtx_REG (Pmode, R0_REG),
- gen_rtx_SYMBOL_REF (Pmode,
- "__GCC_shcompact_incoming_args"));
+ function_symbol ("__GCC_shcompact_incoming_args"));
emit_insn (gen_shcompact_incoming_args ());
}
}
void
sh_expand_epilogue ()
{
- HOST_WIDE_INT live_regs_mask[(FIRST_PSEUDO_REGISTER + 31) / 32];
+ HARD_REG_SET live_regs_mask;
int d, i;
int d_rounding = 0;
int save_flags = target_flags;
- int frame_size;
+ int frame_size, save_size;
+ int fpscr_deferred = 0;
+
+ d = calc_live_regs (&live_regs_mask);
+
+ save_size = d;
+ frame_size = rounded_frame_size (d);
+
+ if (TARGET_SH5)
+ {
+ int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask);
+ int total_size;
+ if (d % (STACK_BOUNDARY / BITS_PER_UNIT))
+ d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT)
+ - d % (STACK_BOUNDARY / BITS_PER_UNIT));
- calc_live_regs (&d, live_regs_mask);
+ total_size = d + tregs_space;
+ total_size += rounded_frame_size (total_size);
+ save_size = total_size - frame_size;
- if (TARGET_SH5 && d % (STACK_BOUNDARY / BITS_PER_UNIT))
- d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT)
- - d % (STACK_BOUNDARY / BITS_PER_UNIT));
+ /* If adjusting the stack in a single step costs nothing extra, do so.
+ I.e. either if a single addi is enough, or we need a movi anyway,
+ and we don't exceed the maximum offset range (the test for the
+ latter is conservative for simplicity). */
+ if (TARGET_SHMEDIA
+ && ! frame_pointer_needed
+ && (CONST_OK_FOR_I10 (total_size)
+ || (! CONST_OK_FOR_I10 (save_size + d_rounding)
+ && total_size <= 2044)))
+ d_rounding = frame_size;
- frame_size = rounded_frame_size (d) - d_rounding;
+ frame_size -= d_rounding;
+ }
if (frame_pointer_needed)
{
- output_stack_adjust (frame_size, frame_pointer_rtx, 7, emit_insn);
+ output_stack_adjust (frame_size, frame_pointer_rtx, 1, &live_regs_mask);
/* We must avoid moving the stack pointer adjustment past code
which reads from the local frame, else an interrupt could
occur after the SP adjustment and clobber data in the local
frame. */
emit_insn (gen_blockage ());
- output_stack_adjust (frame_size, stack_pointer_rtx, 7, emit_insn);
+ output_stack_adjust (frame_size, stack_pointer_rtx, 1, &live_regs_mask);
}
if (SHMEDIA_REGS_STACK_ADJUST ())
{
emit_move_insn (gen_rtx_REG (Pmode, R0_REG),
- gen_rtx_SYMBOL_REF (Pmode,
- TARGET_FPU_ANY
- ? "__GCC_pop_shmedia_regs"
- : "__GCC_pop_shmedia_regs_nofpu"));
+ function_symbol (TARGET_FPU_ANY
+ ? "__GCC_pop_shmedia_regs"
+ : "__GCC_pop_shmedia_regs_nofpu"));
/* This must NOT go through the PLT, otherwise mach and macl
may be clobbered. */
emit_insn (gen_shmedia_save_restore_regs_compact
/* Pop all the registers. */
- if (target_flags != save_flags)
+ if (target_flags != save_flags && ! current_function_interrupt)
emit_insn (gen_toggle_sz ());
if (TARGET_SH5)
{
- int offset = d_rounding;
+ int offset_base, offset;
int offset_in_r0 = -1;
int sp_in_r0 = 0;
- int align;
rtx r0 = gen_rtx_REG (Pmode, R0_REG);
- int tmp_regno = R20_REG;
-
- /* We loop twice: first, we save 8-byte aligned registers in the
- higher addresses, that are known to be aligned. Then, we
- proceed to saving 32-bit registers that don't need 8-byte
- alignment. */
- for (align = 0; align <= 1; align++)
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
- if (live_regs_mask[i/32] & (1 << (i % 32)))
- {
- enum machine_mode mode = REGISTER_NATURAL_MODE (i);
- int reg = i;
- rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn;
-
- if (mode == SFmode && (i % 2) == 0
- && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i)
- && (live_regs_mask[(i ^ 1) / 32] & (1 << ((i ^ 1) % 32))))
- {
- mode = DFmode;
- i++;
- }
+ save_schedule schedule;
+ save_entry *entry;
+ int *tmp_pnt;
+
+ entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding);
+ offset_base = -entry[1].offset + d_rounding;
+ tmp_pnt = schedule.temps;
+ for (; entry->mode != VOIDmode; entry--)
+ {
+ enum machine_mode mode = entry->mode;
+ int reg = entry->reg;
+ rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn;
- /* If we're doing the aligned pass and this is not aligned,
- or we're doing the unaligned pass and this is aligned,
- skip it. */
- if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT)
- == 0) != align)
- continue;
+ offset = offset_base + entry->offset;
+ reg_rtx = gen_rtx_REG (mode, reg);
- reg_rtx = gen_rtx_REG (mode, reg);
+ mem_rtx = gen_rtx_MEM (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- GEN_INT (offset)));
+ GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_post_inc);
- GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_post_inc);
+ mem_rtx = NULL_RTX;
- mem_rtx = NULL_RTX;
+ try_post_inc:
+ do
+ if (HAVE_POST_INCREMENT
+ && (offset == offset_in_r0
+ || (offset + GET_MODE_SIZE (mode) != d + d_rounding
+ && mem_rtx == NULL_RTX)
+ || reg == PR_REG || SPECIAL_REGISTER_P (reg)))
+ {
+ post_inc = gen_rtx_MEM (mode,
+ gen_rtx_POST_INC (Pmode, r0));
- try_post_inc:
- do
- if (HAVE_POST_INCREMENT
- && (offset == offset_in_r0
- || (offset + GET_MODE_SIZE (mode) != d + d_rounding
- && mem_rtx == NULL_RTX)
- || i == PR_REG || SPECIAL_REGISTER_P (i)))
- {
- post_inc = gen_rtx_MEM (mode,
- gen_rtx_POST_INC (Pmode, r0));
+ GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (post_inc, 0),
+ post_inc_ok);
- GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (post_inc, 0),
- post_inc_ok);
+ post_inc = NULL_RTX;
- post_inc = NULL_RTX;
+ break;
+
+ post_inc_ok:
+ mem_rtx = NULL_RTX;
+ }
+ while (0);
+
+ if (mem_rtx != NULL_RTX)
+ goto addr_ok;
- break;
-
- post_inc_ok:
- mem_rtx = NULL_RTX;
- }
- while (0);
+ if (offset_in_r0 == -1)
+ {
+ emit_move_insn (r0, GEN_INT (offset));
+ offset_in_r0 = offset;
+ }
+ else if (offset != offset_in_r0)
+ {
+ emit_move_insn (r0,
+ gen_rtx_PLUS
+ (Pmode, r0,
+ GEN_INT (offset - offset_in_r0)));
+ offset_in_r0 += offset - offset_in_r0;
+ }
- if (mem_rtx != NULL_RTX)
- goto addr_ok;
-
- if (offset_in_r0 == -1)
- {
- emit_move_insn (r0, GEN_INT (offset));
- offset_in_r0 = offset;
- }
- else if (offset != offset_in_r0)
+ if (post_inc != NULL_RTX)
+ {
+ if (! sp_in_r0)
{
emit_move_insn (r0,
gen_rtx_PLUS
- (Pmode, r0,
- GEN_INT (offset - offset_in_r0)));
- offset_in_r0 += offset - offset_in_r0;
+ (Pmode, r0, stack_pointer_rtx));
+ sp_in_r0 = 1;
}
-
- if (post_inc != NULL_RTX)
- {
- if (! sp_in_r0)
- {
- emit_move_insn (r0,
- gen_rtx_PLUS
- (Pmode, r0, stack_pointer_rtx));
- sp_in_r0 = 1;
- }
-
- mem_rtx = post_inc;
+
+ mem_rtx = post_inc;
- offset_in_r0 += GET_MODE_SIZE (mode);
- }
- else if (sp_in_r0)
- mem_rtx = gen_rtx_MEM (mode, r0);
- else
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- r0));
-
- if ((i == PR_REG || SPECIAL_REGISTER_P (i))
- && mem_rtx != post_inc)
- abort ();
-
- addr_ok:
- if ((i == PR_REG || SPECIAL_REGISTER_P (i))
- && mem_rtx != post_inc)
- {
- insn = emit_move_insn (r0, mem_rtx);
- mem_rtx = r0;
- }
- else if (TARGET_REGISTER_P (i))
- {
- rtx tmp_reg = gen_rtx_REG (mode, tmp_regno);
-
- /* Give the scheduler a bit of freedom by using R20..R23
- in a round-robin fashion. Don't use R1 here because
- we want to use it for EH_RETURN_STACKADJ_RTX. */
- insn = emit_move_insn (tmp_reg, mem_rtx);
- mem_rtx = tmp_reg;
- if (++tmp_regno > R23_REG)
- tmp_regno = R20_REG;
- }
+ offset_in_r0 += GET_MODE_SIZE (mode);
+ }
+ else if (sp_in_r0)
+ mem_rtx = gen_rtx_MEM (mode, r0);
+ else
+ mem_rtx = gen_rtx_MEM (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ r0));
- insn = emit_move_insn (reg_rtx, mem_rtx);
+ if ((reg == PR_REG || SPECIAL_REGISTER_P (reg))
+ && mem_rtx != post_inc)
+ abort ();
- offset += GET_MODE_SIZE (mode);
+ addr_ok:
+ if ((reg == PR_REG || SPECIAL_REGISTER_P (reg))
+ && mem_rtx != post_inc)
+ {
+ insn = emit_move_insn (r0, mem_rtx);
+ mem_rtx = r0;
+ }
+ else if (TARGET_REGISTER_P (reg))
+ {
+ rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt);
+
+ /* Give the scheduler a bit of freedom by using up to
+ MAX_TEMPS registers in a round-robin fashion. */
+ insn = emit_move_insn (tmp_reg, mem_rtx);
+ mem_rtx = tmp_reg;
+ if (*++tmp_pnt < 0)
+ tmp_pnt = schedule.temps;
}
- if (offset != d + d_rounding)
- abort ();
+ insn = emit_move_insn (reg_rtx, mem_rtx);
+ if (reg == PR_MEDIA_REG && sh_media_register_for_return () >= 0)
+ /* This is dead, unless we return with a sibcall. */
+ REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
+ const0_rtx,
+ REG_NOTES (insn));
+ }
- goto finish;
+ if (entry->offset + offset_base != d + d_rounding)
+ abort ();
}
- else
- d = 0;
- if (live_regs_mask[PR_REG / 32] & (1 << (PR_REG % 32)))
- pop (PR_REG);
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ else /* ! TARGET_SH5 */
{
- int j = (FIRST_PSEUDO_REGISTER - 1) - i;
+ save_size = 0;
+ if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG))
+ pop (PR_REG);
+ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ {
+ int j = (FIRST_PSEUDO_REGISTER - 1) - i;
+
+ if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD
+ && hard_regs_intersect_p (&live_regs_mask,
+ ®_class_contents[DF_REGS]))
+ fpscr_deferred = 1;
+ else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j))
+ pop (j);
+ if (j == FIRST_FP_REG && fpscr_deferred)
+ pop (FPSCR_REG);
- if (j != PR_REG && live_regs_mask[j / 32] & (1 << (j % 32)))
- pop (j);
+ }
}
- finish:
- if (target_flags != save_flags)
+ if (target_flags != save_flags && ! current_function_interrupt)
emit_insn (gen_toggle_sz ());
target_flags = save_flags;
output_stack_adjust (extra_push + current_function_pretend_args_size
- + d + d_rounding
+ + save_size + d_rounding
+ current_function_args_info.stack_regs * 8,
- stack_pointer_rtx, 7, emit_insn);
+ stack_pointer_rtx, 1, NULL);
if (current_function_calls_eh_return)
emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx,
/* PR_REG will never be live in SHmedia mode, and we don't need to
USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG
by the return pattern. */
- if (live_regs_mask[PR_REG / 32] & (1 << (PR_REG % 32)))
+ if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG))
emit_insn (gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, PR_REG)));
}
sh_set_return_address (ra, tmp)
rtx ra, tmp;
{
- HOST_WIDE_INT live_regs_mask[(FIRST_PSEUDO_REGISTER + 31) / 32];
+ HARD_REG_SET live_regs_mask;
int d;
- int d_rounding = 0;
int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG;
int pr_offset;
- calc_live_regs (&d, live_regs_mask);
+ d = calc_live_regs (&live_regs_mask);
/* If pr_reg isn't life, we can set it (or the register given in
sh_media_register_for_return) directly. */
- if ((live_regs_mask[pr_reg / 32] & (1 << (pr_reg % 32))) == 0)
+ if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg))
{
rtx rr;
if (TARGET_SH5)
{
- int i;
int offset;
- int align;
+ save_schedule schedule;
+ save_entry *entry;
- if (d % (STACK_BOUNDARY / BITS_PER_UNIT))
- d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT)
- - d % (STACK_BOUNDARY / BITS_PER_UNIT));
-
- offset = 0;
-
- /* We loop twice: first, we save 8-byte aligned registers in the
- higher addresses, that are known to be aligned. Then, we
- proceed to saving 32-bit registers that don't need 8-byte
- alignment. */
- for (align = 0; align <= 1; align++)
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
- if (live_regs_mask[i/32] & (1 << (i % 32)))
- {
- enum machine_mode mode = REGISTER_NATURAL_MODE (i);
-
- if (mode == SFmode && (i % 2) == 0
- && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i)
- && (live_regs_mask[(i ^ 1) / 32] & (1 << ((i ^ 1) % 32))))
- {
- mode = DFmode;
- i++;
- }
-
- /* If we're doing the aligned pass and this is not aligned,
- or we're doing the unaligned pass and this is aligned,
- skip it. */
- if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT)
- == 0) != align)
- continue;
-
- if (i == pr_reg)
- goto found;
-
- offset += GET_MODE_SIZE (mode);
- }
+ entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0);
+ offset = entry[1].offset;
+ for (; entry->mode != VOIDmode; entry--)
+ if (entry->reg == pr_reg)
+ goto found;
/* We can't find pr register. */
abort ();
found:
- pr_offset = (rounded_frame_size (d) - d_rounding + offset
+ offset = entry->offset - offset;
+ pr_offset = (rounded_frame_size (d) + offset
+ SHMEDIA_REGS_STACK_ADJUST ());
}
else
- pr_offset = rounded_frame_size (d) - d_rounding;
+ pr_offset = rounded_frame_size (d);
emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset)));
emit_insn (GEN_ADD3 (tmp, tmp, frame_pointer_rtx));
move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg,
adjust_address (regbuf, BLKmode,
n_floatregs * UNITS_PER_WORD),
- n_intregs, n_intregs * UNITS_PER_WORD);
+ n_intregs);
if (TARGET_SHMEDIA)
/* Return the address of the regbuf. */
HOST_WIDE_INT size, rsize;
tree tmp, pptr_type_node;
rtx addr_rtx, r;
- rtx result;
+ rtx result_ptr, result = NULL_RTX;
int pass_by_ref = MUST_PASS_IN_STACK (TYPE_MODE (type), type);
+ rtx lab_over;
size = int_size_in_bytes (type);
rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack;
tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack;
int pass_as_float;
- rtx lab_false, lab_over;
+ rtx lab_false;
f_next_o = TYPE_FIELDS (va_list_type_node);
f_next_o_limit = TREE_CHAIN (f_next_o);
next_stack = build (COMPONENT_REF, TREE_TYPE (f_next_stack),
valist, f_next_stack);
+ /* Structures with a single member with a distinct mode are passed
+ like their member. This is relevant if the latter has a REAL_TYPE
+ or COMPLEX_TYPE type. */
+ if (TREE_CODE (type) == RECORD_TYPE
+ && TYPE_FIELDS (type)
+ && TREE_CODE (TYPE_FIELDS (type)) == FIELD_DECL
+ && (TREE_CODE (TREE_TYPE (TYPE_FIELDS (type))) == REAL_TYPE
+ || TREE_CODE (TREE_TYPE (TYPE_FIELDS (type))) == COMPLEX_TYPE)
+ && TREE_CHAIN (TYPE_FIELDS (type)) == NULL_TREE)
+ type = TREE_TYPE (TYPE_FIELDS (type));
if (TARGET_SH4)
{
pass_as_float = ((TREE_CODE (type) == REAL_TYPE && size <= 8)
lab_false = gen_label_rtx ();
lab_over = gen_label_rtx ();
+ tmp = make_tree (pptr_type_node, addr_rtx);
+ valist = build1 (INDIRECT_REF, ptr_type_node, tmp);
+
if (pass_as_float)
{
int first_floatreg
if (r != addr_rtx)
emit_move_insn (addr_rtx, r);
+#ifdef FUNCTION_ARG_SCmode_WART
+ if (TYPE_MODE (type) == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN)
+ {
+ rtx addr, real, imag, result_value, slot;
+ tree subtype = TREE_TYPE (type);
+
+ addr = std_expand_builtin_va_arg (valist, subtype);
+#ifdef POINTERS_EXTEND_UNSIGNED
+ if (GET_MODE (addr) != Pmode)
+ addr = convert_memory_address (Pmode, addr);
+#endif
+ imag = gen_rtx_MEM (TYPE_MODE (type), addr);
+ set_mem_alias_set (imag, get_varargs_alias_set ());
+
+ addr = std_expand_builtin_va_arg (valist, subtype);
+#ifdef POINTERS_EXTEND_UNSIGNED
+ if (GET_MODE (addr) != Pmode)
+ addr = convert_memory_address (Pmode, addr);
+#endif
+ real = gen_rtx_MEM (TYPE_MODE (type), addr);
+ set_mem_alias_set (real, get_varargs_alias_set ());
+
+ result_value = gen_rtx_CONCAT (SCmode, real, imag);
+ /* ??? this interface is stupid - why require a pointer? */
+ result = gen_reg_rtx (Pmode);
+ slot = assign_stack_temp (SCmode, 8, 0);
+ emit_move_insn (slot, result_value);
+ emit_move_insn (result, XEXP (slot, 0));
+ }
+#endif /* FUNCTION_ARG_SCmode_WART */
+
emit_jump_insn (gen_jump (lab_over));
emit_barrier ();
emit_label (lab_false);
emit_move_insn (addr_rtx, r);
}
- emit_label (lab_over);
-
- tmp = make_tree (pptr_type_node, addr_rtx);
- valist = build1 (INDIRECT_REF, ptr_type_node, tmp);
+ if (! result)
+ emit_label (lab_over);
}
/* ??? In va-sh.h, there had been code to make values larger than
size 8 indirect. This does not match the FUNCTION_ARG macros. */
- result = std_expand_builtin_va_arg (valist, type);
+ result_ptr = std_expand_builtin_va_arg (valist, type);
+ if (result)
+ {
+ emit_move_insn (result, result_ptr);
+ emit_label (lab_over);
+ }
+ else
+ result = result_ptr;
+
if (pass_by_ref)
{
#ifdef POINTERS_EXTEND_UNSIGNED
int total_auto_space;
int save_flags = target_flags;
int copy_flags;
+ HARD_REG_SET live_regs_mask;
- HOST_WIDE_INT live_regs_mask[(FIRST_PSEUDO_REGISTER + 31) / 32];
- calc_live_regs (®s_saved, live_regs_mask);
+ shmedia_space_reserved_for_target_registers = false;
+ regs_saved = calc_live_regs (&live_regs_mask);
regs_saved += SHMEDIA_REGS_STACK_ADJUST ();
+
+ if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask))
+ {
+ shmedia_space_reserved_for_target_registers = true;
+ regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask);
+ }
+
if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT))
regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT)
- regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT));
{
if (TARGET_SH5)
{
- int i, n = total_saved_regs_space;
- int align;
+ int n = total_saved_regs_space;
int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG;
+ save_schedule schedule;
+ save_entry *entry;
n += total_auto_space;
/* If it wasn't saved, there's not much we can do. */
- if ((live_regs_mask[pr_reg / 32] & (1 << (pr_reg % 32))) == 0)
+ if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg))
return n;
target_flags = copy_flags;
- /* We loop twice: first, check 8-byte aligned registers,
- that are stored in the higher addresses, that are known
- to be aligned. Then, check 32-bit registers that don't
- need 8-byte alignment. */
- for (align = 1; align >= 0; align--)
- for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
- if (live_regs_mask[i/32] & (1 << (i % 32)))
- {
- enum machine_mode mode = REGISTER_NATURAL_MODE (i);
-
- if (mode == SFmode && (i % 2) == 1
- && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i)
- && (live_regs_mask[(i ^ 1) / 32]
- & (1 << ((i ^ 1) % 32))))
- {
- mode = DFmode;
- i--;
- }
-
- /* If we're doing the aligned pass and this is not aligned,
- or we're doing the unaligned pass and this is aligned,
- skip it. */
- if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT)
- == 0) != align)
- continue;
-
- n -= GET_MODE_SIZE (mode);
-
- if (i == pr_reg)
- {
- target_flags = save_flags;
- return n;
- }
- }
-
+ sh5_schedule_saves (&live_regs_mask, &schedule, n);
+ for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++)
+ if (entry->reg == pr_reg)
+ {
+ target_flags = save_flags;
+ return entry->offset;
+ }
abort ();
}
else
abort ();
}
\f
-/* Handle machine specific pragmas to be semi-compatible with Hitachi
+/* Handle machine specific pragmas to be semi-compatible with Renesas
compiler. */
void
if (TARGET_SHMEDIA)
{
/* FIXME: We should be checking whether the CONST_INT fits in a
- CONST_OK_FOR_J here, but this causes reload_cse to crash when
+ CONST_OK_FOR_I16 here, but this causes reload_cse to crash when
attempting to transform a sequence of two 64-bit sets of the
same register from literal constants into a set and an add,
when the difference is too wide for an add. */
if (GET_CODE (op) == CONST_INT
- || EXTRA_CONSTRAINT_S (op))
+ || EXTRA_CONSTRAINT_C16 (op))
return 1;
else
return 0;
}
- else if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_I (INTVAL (op)))
+ else if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_I08 (INTVAL (op)))
return 1;
return 0;
if (arith_reg_operand (op, mode))
return 1;
- if (EXTRA_CONSTRAINT_U (op))
+ if (EXTRA_CONSTRAINT_Z (op))
return 1;
return 0;
enum machine_mode mode;
{
return (arith_reg_operand (op, mode)
- || (GET_CODE (op) == CONST_INT && CONST_OK_FOR_O (INTVAL (op))));
+ || (GET_CODE (op) == CONST_INT && CONST_OK_FOR_I06 (INTVAL (op))));
}
/* Returns 1 if OP is a valid source operand for a logical operation. */
if (TARGET_SHMEDIA)
{
- if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_P (INTVAL (op)))
+ if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_I10 (INTVAL (op)))
return 1;
else
return 0;
}
- else if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_L (INTVAL (op)))
+ else if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_K08 (INTVAL (op)))
return 1;
return 0;
if (TARGET_SHMEDIA
&& mode == DImode
&& GET_CODE (op) == CONST_INT
- && (INTVAL (op) == (unsigned) 0xffffffff
- || INTVAL (op) == (HOST_WIDE_INT) -1 << 32))
+ && CONST_OK_FOR_J16 (INTVAL (op)))
return 1;
return 0;
rtx op;
enum machine_mode mode ATTRIBUTE_UNUSED;
{
- const char *str;
-
if (GET_CODE (op) != SYMBOL_REF)
return 0;
-
- str = XSTR (op, 0);
- STRIP_DATALABEL_ENCODING(str, str);
- if (! TLS_SYMNAME_P (str))
- return 0;
-
- switch (str[1])
- {
- case 'G':
- return TLS_MODEL_GLOBAL_DYNAMIC;
- case 'L':
- return TLS_MODEL_LOCAL_DYNAMIC;
- case 'i':
- return TLS_MODEL_INITIAL_EXEC;
- case 'l':
- return TLS_MODEL_LOCAL_EXEC;
- }
- return 0;
+ return SYMBOL_REF_TLS_MODEL (op);
}
int
return 0;
if ((GET_MODE (op) == DImode || GET_MODE (op) == VOIDmode)
- && EXTRA_CONSTRAINT_T (op))
+ && EXTRA_CONSTRAINT_Csy (op))
return ! reload_completed;
return target_reg_operand (op, mode);
enum attr_fp_mode fp_mode = mode;
rtx addr_reg = get_free_reg (regs_live);
- if (fp_mode == (enum attr_fp_mode) NORMAL_MODE (FP_MODE))
+ if (fp_mode == (enum attr_fp_mode) ACTUAL_NORMAL_MODE (FP_MODE))
emit_insn (gen_fpu_switch1 (addr_reg));
else
emit_insn (gen_fpu_switch0 (addr_reg));
return orig;
if (GET_CODE (orig) == LABEL_REF
- || (GET_CODE (orig) == SYMBOL_REF
- && (CONSTANT_POOL_ADDRESS_P (orig)
- /* SYMBOL_REF_FLAG is set on static symbols. */
- || SYMBOL_REF_FLAG (orig))))
+ || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig)))
{
if (reg == 0)
reg = gen_reg_rtx (Pmode);
return (TARGET_SHMEDIA && (reload_in_progress || reload_completed));
}
-static bool
-sh_ms_bitfield_layout_p (record_type)
- tree record_type ATTRIBUTE_UNUSED;
+static int
+sh_target_reg_class (void)
{
- return TARGET_SH5;
+ return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS;
}
-/* If using PIC, mark a SYMBOL_REF for a non-global symbol so that we
- may access it using GOTOFF instead of GOT. */
-
-static void
-sh_encode_section_info (decl, first)
- tree decl;
- int first;
+static bool
+sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen)
{
- rtx rtl, symbol;
-
- if (DECL_P (decl))
- rtl = DECL_RTL (decl);
- else
- rtl = TREE_CST_RTL (decl);
- if (GET_CODE (rtl) != MEM)
- return;
- symbol = XEXP (rtl, 0);
- if (GET_CODE (symbol) != SYMBOL_REF)
- return;
-
- if (flag_pic)
- SYMBOL_REF_FLAG (symbol) = (*targetm.binds_local_p) (decl);
-
- if (TREE_CODE (decl) == VAR_DECL && DECL_THREAD_LOCAL (decl))
- {
- const char *symbol_str, *orig_str;
- bool is_local;
- enum tls_model kind;
- char encoding;
- char *newstr;
- size_t len, dlen;
-
- orig_str = XSTR (symbol, 0);
- is_local = (*targetm.binds_local_p) (decl);
-
- if (! flag_pic)
- {
- if (is_local)
- kind = TLS_MODEL_LOCAL_EXEC;
- else
- kind = TLS_MODEL_INITIAL_EXEC;
- }
- else if (is_local)
- kind = TLS_MODEL_LOCAL_DYNAMIC;
- else
- kind = TLS_MODEL_GLOBAL_DYNAMIC;
- if (kind < flag_tls_default)
- kind = flag_tls_default;
-
- STRIP_DATALABEL_ENCODING (symbol_str, orig_str);
- dlen = symbol_str - orig_str;
-
- encoding = " GLil"[kind];
- if (TLS_SYMNAME_P (symbol_str))
- {
- if (encoding == symbol_str[1])
- return;
- /* Handle the changes from initial-exec to local-exec and
- from global-dynamic to local-dynamic. */
- if ((encoding == 'l' && symbol_str[1] == 'i')
- || (encoding == 'L' && symbol_str[1] == 'G'))
- symbol_str += 2;
- else
- abort ();
- }
-
- len = strlen (symbol_str);
- newstr = alloca (dlen + len + 3);
- if (dlen)
- memcpy (newstr, orig_str, dlen);
- newstr[dlen + 0] = SH_TLS_ENCODING[0];
- newstr[dlen + 1] = encoding;
- memcpy (newstr + dlen + 2, symbol_str, len + 1);
-
- XSTR (symbol, 0) = ggc_alloc_string (newstr, dlen + len + 2);
- }
-
- if (TARGET_SH5 && first && TREE_CODE (decl) != FUNCTION_DECL)
- XEXP (rtl, 0) = gen_datalabel_ref (symbol);
+ return (shmedia_space_reserved_for_target_registers
+ && (! after_prologue_epilogue_gen || TARGET_SAVE_ALL_TARGET_REGS));
}
-/* Undo the effects of the above. */
-
-static const char *
-sh_strip_name_encoding (str)
- const char *str;
+static bool
+sh_ms_bitfield_layout_p (record_type)
+ tree record_type ATTRIBUTE_UNUSED;
{
- STRIP_DATALABEL_ENCODING (str, str);
- STRIP_TLS_ENCODING (str, str);
- str += *str == '*';
- return str;
+ return TARGET_SH5;
}
-
\f
/*
On the SH1..SH4, the trampoline looks like
if (TARGET_HARVARD)
{
if (TARGET_USERMODE)
- emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__ic_invalidate"),
+ emit_library_call (function_symbol ("__ic_invalidate"),
0, VOIDmode, 1, tramp, SImode);
else
emit_insn (gen_ic_invalidate_line (tramp));
{
return (decl
&& (! TARGET_SHCOMPACT
- || current_function_args_info.stack_regs == 0));
+ || current_function_args_info.stack_regs == 0)
+ && ! sh_cfun_interrupt_handler_p ());
}
\f
/* Machine specific built-in functions. */
const struct builtin_description *d;
memset (shared, 0, sizeof shared);
- for (d = bdesc; d - bdesc < (int) (sizeof bdesc / sizeof bdesc[0]); d++)
+ for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++)
{
tree type, arg_type;
int signature = d->signature;
if (dstclass == T_REGS || dstclass == PR_REGS)
return 10;
+ if (dstclass == MAC_REGS && srcclass == MAC_REGS)
+ return 4;
+
if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD
&& REGCLASS_HAS_FP_REG (srcclass)
&& REGCLASS_HAS_FP_REG (dstclass))
return 4;
+ if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS)
+ || (dstclass== MAC_REGS && REGCLASS_HAS_FP_REG (srcclass)))
+ return 9;
+
if ((REGCLASS_HAS_FP_REG (dstclass)
&& REGCLASS_HAS_GENERAL_REG (srcclass))
|| (REGCLASS_HAS_GENERAL_REG (dstclass)
int structure_value_byref = 0;
rtx this, this_value, sibcall, insns, funexp;
tree funtype = TREE_TYPE (function);
- int simple_add
- = (TARGET_SHMEDIA ? CONST_OK_FOR_J (delta) : CONST_OK_FOR_I (delta));
+ int simple_add = CONST_OK_FOR_ADD (delta);
int did_load = 0;
rtx scratch0, scratch1, scratch2;
reload_completed = 1;
+ epilogue_completed = 1;
no_new_pseudos = 1;
current_function_uses_only_leaf_regs = 1;
- emit_note (NULL, NOTE_INSN_PROLOGUE_END);
+ emit_note (NOTE_INSN_PROLOGUE_END);
/* Find the "this" pointer. We have such a wide range of ABIs for the
SH that it's best to do this completely machine independently.
emit_move_insn (scratch1, GEN_INT (vcall_offset));
offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1);
}
- else if (TARGET_SHMEDIA
- ? CONST_OK_FOR_J (vcall_offset)
- : CONST_OK_FOR_I (vcall_offset))
+ else if (CONST_OK_FOR_ADD (vcall_offset))
{
emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset)));
offset_addr = scratch0;
use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this);
emit_barrier ();
- /* Run just enough of rest_of_compilation to do scheduling and get
+ /* Run just enough of rest_of_compilation to do scheduling and get
the insns emitted. Note that use_thunk calls
assemble_start_function and assemble_end_function. */
+
+ insn_locators_initialize ();
insns = get_insns ();
if (optimize > 0 && flag_schedule_insns_after_reload)
schedule_insns (rtl_dump_file);
}
- MACHINE_DEPENDENT_REORG (insns);
+ sh_reorg ();
if (optimize > 0 && flag_delayed_branch)
dbr_schedule (insns, rtl_dump_file);
}
reload_completed = 0;
+ epilogue_completed = 0;
no_new_pseudos = 0;
}
+rtx
+function_symbol (const char *name)
+{
+ rtx sym = gen_rtx_SYMBOL_REF (Pmode, name);
+ SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION;
+ return sym;
+}
+
+/* Find the number of a general purpose register in S. */
+static int
+scavenge_reg (HARD_REG_SET *s)
+{
+ int r;
+ for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++)
+ if (TEST_HARD_REG_BIT (*s, r))
+ return r;
+ return -1;
+}
+
+rtx
+sh_get_pr_initial_val (void)
+{
+ rtx val;
+
+ /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the
+ PR register on SHcompact, because it might be clobbered by the prologue.
+ We check first if that is known to be the case. */
+ if (TARGET_SHCOMPACT
+ && ((current_function_args_info.call_cookie
+ & ~ CALL_COOKIE_RET_TRAMP (1))
+ || current_function_has_nonlocal_label))
+ return gen_rtx_MEM (SImode, return_address_pointer_rtx);
+
+ /* If we haven't finished rtl generation, there might be a nonlocal label
+ that we haven't seen yet.
+ ??? get_hard_reg_initial_val fails if it is called while no_new_pseudos
+ is set, unless it has been called before for the same register. And even
+ then, we end in trouble if we didn't use the register in the same
+ basic block before. So call get_hard_reg_initial_val now and wrap it
+ in an unspec if we might need to replace it. */
+ val
+ = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG);
+ if (TARGET_SHCOMPACT && rtx_equal_function_value_matters)
+ return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA);
+ return val;
+}
+
#include "gt-sh.h"
OSDN Git Service
