static bool pa_commutative_p (rtx x, int outer_code);
static void copy_fp_args (rtx) ATTRIBUTE_UNUSED;
static int length_fp_args (rtx) ATTRIBUTE_UNUSED;
-static struct deferred_plabel *get_plabel (rtx) ATTRIBUTE_UNUSED;
static inline void pa_file_start_level (void) ATTRIBUTE_UNUSED;
static inline void pa_file_start_space (int) ATTRIBUTE_UNUSED;
static inline void pa_file_start_file (int) ATTRIBUTE_UNUSED;
static void pa_hpux64_gas_file_start (void) ATTRIBUTE_UNUSED;
static void pa_hpux64_hpas_file_start (void) ATTRIBUTE_UNUSED;
static void output_deferred_plabels (void);
+static void output_deferred_profile_counters (void) ATTRIBUTE_UNUSED;
#ifdef ASM_OUTPUT_EXTERNAL_REAL
static void pa_hpux_file_end (void);
#endif
static int pa_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
tree, bool);
static struct machine_function * pa_init_machine_status (void);
+static enum reg_class pa_secondary_reload (bool, rtx, enum reg_class,
+ enum machine_mode,
+ secondary_reload_info *);
/* Save the operands last given to a compare for use when we
#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM pa_tls_referenced_p
+#undef TARGET_SECONDARY_RELOAD
+#define TARGET_SECONDARY_RELOAD pa_secondary_reload
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
/* Parse the -mfixed-range= option string. */
pa_init_builtins (void)
{
#ifdef DONT_HAVE_FPUTC_UNLOCKED
- built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE;
- implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE;
+ built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] =
+ built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
+ implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED]
+ = implicit_built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
#endif
}
if (scratch_reg && reload_in_progress && GET_CODE (operand0) == MEM
&& ((tem = find_replacement (&XEXP (operand0, 0)))
!= XEXP (operand0, 0)))
- operand0 = gen_rtx_MEM (GET_MODE (operand0), tem);
+ operand0 = replace_equiv_address (operand0, tem);
if (scratch_reg && reload_in_progress && GET_CODE (operand1) == MEM
&& ((tem = find_replacement (&XEXP (operand1, 0)))
!= XEXP (operand1, 0)))
- operand1 = gen_rtx_MEM (GET_MODE (operand1), tem);
+ operand1 = replace_equiv_address (operand1, tem);
/* Handle secondary reloads for loads/stores of FP registers from
REG+D addresses where D does not fit in 5 or 14 bits, including
else
emit_move_insn (scratch_reg, XEXP (operand1, 0));
emit_insn (gen_rtx_SET (VOIDmode, operand0,
- gen_rtx_MEM (mode, scratch_reg)));
+ replace_equiv_address (operand1, scratch_reg)));
return 1;
}
else if (scratch_reg
}
else
emit_move_insn (scratch_reg, XEXP (operand0, 0));
- emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_MEM (mode, scratch_reg),
+ emit_insn (gen_rtx_SET (VOIDmode,
+ replace_equiv_address (operand0, scratch_reg),
operand1));
return 1;
}
&& CONSTANT_P (operand1)
&& fp_reg_operand (operand0, mode))
{
- rtx xoperands[2];
+ rtx const_mem, xoperands[2];
/* SCRATCH_REG will hold an address and maybe the actual data. We want
it in WORD_MODE regardless of what mode it was originally given
/* Force the constant into memory and put the address of the
memory location into scratch_reg. */
+ const_mem = force_const_mem (mode, operand1);
xoperands[0] = scratch_reg;
- xoperands[1] = XEXP (force_const_mem (mode, operand1), 0);
+ xoperands[1] = XEXP (const_mem, 0);
emit_move_sequence (xoperands, Pmode, 0);
/* Now load the destination register. */
emit_insn (gen_rtx_SET (mode, operand0,
- gen_rtx_MEM (mode, scratch_reg)));
+ replace_equiv_address (const_mem, scratch_reg)));
return 1;
}
/* Handle secondary reloads for SAR. These occur when trying to load
OPERAND0. */
scratch_reg = force_mode (GET_MODE (operand0), scratch_reg);
- emit_move_insn (scratch_reg, gen_rtx_MEM (GET_MODE (operand0),
- scratch_reg));
+ emit_move_insn (scratch_reg,
+ replace_equiv_address (operand1, scratch_reg));
}
else
{
&& (reload_completed || reload_in_progress)
&& flag_pic)
{
- operands[1] = force_const_mem (mode, operand1);
- operands[1] = legitimize_pic_address (XEXP (operands[1], 0),
+ rtx const_mem = force_const_mem (mode, operand1);
+ operands[1] = legitimize_pic_address (XEXP (const_mem, 0),
mode, temp);
- operands[1] = gen_rtx_MEM (mode, operands[1]);
+ operands[1] = replace_equiv_address (const_mem, operands[1]);
emit_move_sequence (operands, mode, temp);
}
else
case CONSTRUCTOR:
{
- register tree link;
- for (link = CONSTRUCTOR_ELTS (exp); link; link = TREE_CHAIN (link))
- if (TREE_VALUE (link) != 0)
- reloc |= reloc_needed (TREE_VALUE (link));
+ tree value;
+ unsigned HOST_WIDE_INT ix;
+
+ FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), ix, value)
+ if (value)
+ reloc |= reloc_needed (value);
}
break;
rtx tmpreg = gen_rtx_REG (Pmode, 1);
emit_move_insn (tmpreg, delta);
- emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
- dest = gen_rtx_MEM (word_mode, tmpreg);
- insn = emit_move_insn (dest, src);
+ insn = emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
if (DO_FRAME_NOTES)
{
REG_NOTES (insn)
= gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
- gen_rtx_SET (VOIDmode,
- gen_rtx_MEM (word_mode,
- gen_rtx_PLUS (word_mode, basereg,
- delta)),
- src),
+ gen_rtx_SET (VOIDmode, tmpreg,
+ gen_rtx_PLUS (Pmode, basereg, delta)),
REG_NOTES (insn));
+ RTX_FRAME_RELATED_P (insn) = 1;
}
+ dest = gen_rtx_MEM (word_mode, tmpreg);
+ insn = emit_move_insn (dest, src);
}
else
{
RTX_FRAME_RELATED_P (insn) = 1;
/* RTX_FRAME_RELATED_P must be set on each frame related set
- in a parallel with more than one element. Don't set
- RTX_FRAME_RELATED_P in the first set if reg is temporary
- register 1. The effect of this operation is recorded in
- the initial copy. */
- if (reg != 1)
- {
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1;
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
- }
- else
- {
- /* The first element of a PARALLEL is always processed if it is
- a SET. Thus, we need an expression list for this case. */
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
- gen_rtx_SET (VOIDmode, basereg,
- gen_rtx_PLUS (word_mode, basereg, delta)),
- REG_NOTES (insn));
- }
+ in a parallel with more than one element. */
+ RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1;
+ RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
}
}
emit_move_insn (tmpreg, delta);
insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
gen_rtx_PLUS (Pmode, tmpreg, basereg));
+ if (DO_FRAME_NOTES)
+ REG_NOTES (insn)
+ = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
+ gen_rtx_SET (VOIDmode, tmpreg,
+ gen_rtx_PLUS (Pmode, basereg, delta)),
+ REG_NOTES (insn));
}
else
{
/* The SAVE_SP flag is used to indicate that register %r3 is stored
at the beginning of the frame and that it is used as the frame
pointer for the frame. We do this because our current frame
- layout doesn't conform to that specified in the the HP runtime
+ layout doesn't conform to that specified in the HP runtime
documentation and we need a way to indicate to programs such as
GDB where %r3 is saved. The SAVE_SP flag was chosen because it
isn't used by HP compilers but is supported by the assembler.
frames. */
insn = emit_move_insn (tmpreg, frame_pointer_rtx);
if (DO_FRAME_NOTES)
- {
- /* We need to record the frame pointer save here since the
- new frame pointer is set in the following insn. */
- RTX_FRAME_RELATED_P (insn) = 1;
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
- gen_rtx_SET (VOIDmode,
- gen_rtx_MEM (word_mode, stack_pointer_rtx),
- frame_pointer_rtx),
- REG_NOTES (insn));
- }
+ RTX_FRAME_RELATED_P (insn) = 1;
insn = emit_move_insn (frame_pointer_rtx, stack_pointer_rtx);
if (DO_FRAME_NOTES)
return get_hard_reg_initial_val (word_mode, PIC_OFFSET_TABLE_REGNUM);
}
+#ifndef NO_DEFERRED_PROFILE_COUNTERS
+#define NO_DEFERRED_PROFILE_COUNTERS 0
+#endif
+
+/* Define heap vector type for funcdef numbers. */
+DEF_VEC_I(int);
+DEF_VEC_ALLOC_I(int,heap);
+
+/* Vector of funcdef numbers. */
+static VEC(int,heap) *funcdef_nos;
+
+/* Output deferred profile counters. */
+static void
+output_deferred_profile_counters (void)
+{
+ unsigned int i;
+ int align, n;
+
+ if (VEC_empty (int, funcdef_nos))
+ return;
+
+ data_section ();
+ align = MIN (BIGGEST_ALIGNMENT, LONG_TYPE_SIZE);
+ ASM_OUTPUT_ALIGN (asm_out_file, floor_log2 (align / BITS_PER_UNIT));
+
+ for (i = 0; VEC_iterate (int, funcdef_nos, i, n); i++)
+ {
+ targetm.asm_out.internal_label (asm_out_file, "LP", n);
+ assemble_integer (const0_rtx, LONG_TYPE_SIZE / BITS_PER_UNIT, align, 1);
+ }
+
+ VEC_free (int, heap, funcdef_nos);
+}
+
void
hppa_profile_hook (int label_no)
{
emit_insn (gen_load_offset_label_address (gen_rtx_REG (SImode, 25),
reg, begin_label_rtx, label_rtx));
-#ifndef NO_PROFILE_COUNTERS
+#if !NO_DEFERRED_PROFILE_COUNTERS
{
rtx count_label_rtx, addr, r24;
char count_label_name[16];
+ VEC_safe_push (int, heap, funcdef_nos, label_no);
ASM_GENERATE_INTERNAL_LABEL (count_label_name, "LP", label_no);
count_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (count_label_name));
}
#undef aputs
-static struct deferred_plabel *
-get_plabel (rtx symbol)
+/* Search the deferred plabel list for SYMBOL and return its internal
+ label. If an entry for SYMBOL is not found, a new entry is created. */
+
+rtx
+get_deferred_plabel (rtx symbol)
{
const char *fname = XSTR (symbol, 0);
size_t i;
mark_referenced (id);
}
- return &deferred_plabels[i];
+ return deferred_plabels[i].internal_label;
}
static void
fputc ('\n', asm_out_file);
}
\f
-/* Return the class of any secondary reload register that is needed to
- move IN into a register in class CLASS using mode MODE.
-
- Profiling has showed this routine and its descendants account for
- a significant amount of compile time (~7%). So it has been
- optimized to reduce redundant computations and eliminate useless
- function calls.
-
- It might be worthwhile to try and make this a leaf function too. */
-
-enum reg_class
-secondary_reload_class (enum reg_class class, enum machine_mode mode, rtx in)
+static enum reg_class
+pa_secondary_reload (bool in_p, rtx x, enum reg_class class,
+ enum machine_mode mode, secondary_reload_info *sri)
{
- int regno, is_symbolic;
+ int is_symbolic, regno;
- /* Trying to load a constant into a FP register during PIC code
- generation will require %r1 as a scratch register. */
- if (flag_pic
- && GET_MODE_CLASS (mode) == MODE_INT
- && FP_REG_CLASS_P (class)
- && (GET_CODE (in) == CONST_INT || GET_CODE (in) == CONST_DOUBLE))
- return R1_REGS;
-
- /* Profiling showed the PA port spends about 1.3% of its compilation
- time in true_regnum from calls inside secondary_reload_class. */
+ /* Handle the easy stuff first. */
+ if (class == R1_REGS)
+ return NO_REGS;
- if (GET_CODE (in) == REG)
+ if (REG_P (x))
{
- regno = REGNO (in);
- if (regno >= FIRST_PSEUDO_REGISTER)
- regno = true_regnum (in);
+ regno = REGNO (x);
+ if (class == BASE_REG_CLASS && regno < FIRST_PSEUDO_REGISTER)
+ return NO_REGS;
}
- else if (GET_CODE (in) == SUBREG)
- regno = true_regnum (in);
else
regno = -1;
/* If we have something like (mem (mem (...)), we can safely assume the
inner MEM will end up in a general register after reloading, so there's
no need for a secondary reload. */
- if (GET_CODE (in) == MEM
- && GET_CODE (XEXP (in, 0)) == MEM)
+ if (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == MEM)
return NO_REGS;
+ /* Trying to load a constant into a FP register during PIC code
+ generation requires %r1 as a scratch register. */
+ if (flag_pic
+ && GET_MODE_CLASS (mode) == MODE_INT
+ && FP_REG_CLASS_P (class)
+ && (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE))
+ {
+ gcc_assert (mode == SImode || mode == DImode);
+ sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
+ : CODE_FOR_reload_indi_r1);
+ return NO_REGS;
+ }
+
+ /* Profiling showed the PA port spends about 1.3% of its compilation
+ time in true_regnum from calls inside pa_secondary_reload_class. */
+ if (regno >= FIRST_PSEUDO_REGISTER || GET_CODE (x) == SUBREG)
+ regno = true_regnum (x);
+
/* Handle out of range displacement for integer mode loads/stores of
FP registers. */
if (((regno >= FIRST_PSEUDO_REGISTER || regno == -1)
&& GET_MODE_CLASS (mode) == MODE_INT
&& FP_REG_CLASS_P (class))
|| (class == SHIFT_REGS && (regno <= 0 || regno >= 32)))
- return GENERAL_REGS;
+ {
+ sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
+ return NO_REGS;
+ }
/* A SAR<->FP register copy requires a secondary register (GPR) as
well as secondary memory. */
if (regno >= 0 && regno < FIRST_PSEUDO_REGISTER
&& ((REGNO_REG_CLASS (regno) == SHIFT_REGS && FP_REG_CLASS_P (class))
- || (class == SHIFT_REGS && FP_REG_CLASS_P (REGNO_REG_CLASS (regno)))))
- return GENERAL_REGS;
+ || (class == SHIFT_REGS
+ && FP_REG_CLASS_P (REGNO_REG_CLASS (regno)))))
+ {
+ sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
+ return NO_REGS;
+ }
- if (GET_CODE (in) == HIGH)
- in = XEXP (in, 0);
+ /* Secondary reloads of symbolic operands require %r1 as a scratch
+ register when we're generating PIC code and the operand isn't
+ readonly. */
+ if (GET_CODE (x) == HIGH)
+ x = XEXP (x, 0);
/* Profiling has showed GCC spends about 2.6% of its compilation
- time in symbolic_operand from calls inside secondary_reload_class.
-
- We use an inline copy and only compute its return value once to avoid
- useless work. */
- switch (GET_CODE (in))
+ time in symbolic_operand from calls inside pa_secondary_reload_class.
+ So, we use an inline copy to avoid useless work. */
+ switch (GET_CODE (x))
{
- rtx tmp;
+ rtx op;
case SYMBOL_REF:
+ is_symbolic = !SYMBOL_REF_TLS_MODEL (x);
+ break;
case LABEL_REF:
is_symbolic = 1;
break;
case CONST:
- tmp = XEXP (in, 0);
- is_symbolic = ((GET_CODE (XEXP (tmp, 0)) == SYMBOL_REF
- || GET_CODE (XEXP (tmp, 0)) == LABEL_REF)
- && GET_CODE (XEXP (tmp, 1)) == CONST_INT);
+ op = XEXP (x, 0);
+ is_symbolic = (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
+ && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
+ || GET_CODE (XEXP (op, 0)) == LABEL_REF)
+ && GET_CODE (XEXP (op, 1)) == CONST_INT);
break;
-
default:
is_symbolic = 0;
break;
}
- if (!flag_pic
- && is_symbolic
- && read_only_operand (in, VOIDmode))
- return NO_REGS;
-
- if (class != R1_REGS && is_symbolic)
- return R1_REGS;
+ if (is_symbolic && (flag_pic || !read_only_operand (x, VOIDmode)))
+ {
+ gcc_assert (mode == SImode || mode == DImode);
+ sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
+ : CODE_FOR_reload_indi_r1);
+ }
return NO_REGS;
}
/* Args grow down. Not handled by generic routines. */
u = fold_convert (valist_type, size_in_bytes (type));
- t = build (MINUS_EXPR, valist_type, valist, u);
+ t = build2 (MINUS_EXPR, valist_type, valist, u);
/* Copied from va-pa.h, but we probably don't need to align to
word size, since we generate and preserve that invariant. */
u = build_int_cst (valist_type, (size > 4 ? -8 : -4));
- t = build (BIT_AND_EXPR, valist_type, t, u);
+ t = build2 (BIT_AND_EXPR, valist_type, t, u);
- t = build (MODIFY_EXPR, valist_type, valist, t);
+ t = build2 (MODIFY_EXPR, valist_type, valist, t);
ofs = (8 - size) % 4;
if (ofs != 0)
{
u = fold_convert (valist_type, size_int (ofs));
- t = build (PLUS_EXPR, valist_type, t, u);
+ t = build2 (PLUS_EXPR, valist_type, t, u);
}
t = fold_convert (ptr, t);
- t = build_fold_indirect_ref (t);
+ t = build_va_arg_indirect_ref (t);
if (indirect)
- t = build_fold_indirect_ref (t);
+ t = build_va_arg_indirect_ref (t);
return t;
}
zero for cmpb, we must ensure that we use cmpb for the comparison. */
if (GET_MODE (operands[1]) == DImode && operands[2] == const0_rtx)
operands[2] = gen_rtx_REG (DImode, 0);
+ if (GET_MODE (operands[2]) == DImode && operands[1] == const0_rtx)
+ operands[1] = gen_rtx_REG (DImode, 0);
/* If this is a long branch with its delay slot unfilled, set `nullify'
as it can nullify the delay slot and save a nop. */
/* ??? As far as I can tell, the HP linker doesn't support the
long pc-relative sequence described in the 64-bit runtime
architecture. So, we use a slightly longer indirect call. */
- struct deferred_plabel *p = get_plabel (call_dest);
-
- xoperands[0] = p->internal_label;
+ xoperands[0] = get_deferred_plabel (call_dest);
xoperands[1] = gen_label_rtx ();
/* If this isn't a sibcall, we put the load of %r27 into the
essentially an inline implementation of $$dyncall.
We don't actually try to call $$dyncall as this is
as difficult as calling the function itself. */
- struct deferred_plabel *p = get_plabel (call_dest);
-
- xoperands[0] = p->internal_label;
+ xoperands[0] = get_deferred_plabel (call_dest);
xoperands[1] = gen_label_rtx ();
/* Since the call is indirect, FP arguments in registers
pa_commutative_p (rtx x, int outer_code)
{
return (COMMUTATIVE_P (x)
- && ((outer_code != UNKNOWN && outer_code != MEM)
+ && (TARGET_NO_SPACE_REGS
+ || (outer_code != UNKNOWN && outer_code != MEM)
|| GET_CODE (x) != PLUS));
}
unsigned int i;
extern_symbol *p;
+ if (!NO_DEFERRED_PROFILE_COUNTERS)
+ output_deferred_profile_counters ();
+
output_deferred_plabels ();
for (i = 0; VEC_iterate (extern_symbol, extern_symbols, i, p); i++)