/* Output routines for GCC for Hitachi Super-H.
- Copyright (C) 1993, 1994, 1995 Free Software Foundation, Inc.
+ Copyright (C) 1993, 1994, 1995, 1996 Free Software Foundation, Inc.
This file is part of GNU CC.
#include "regs.h"
#include "hard-reg-set.h"
#include "output.h"
+#include "insn-attr.h"
#define MSW (TARGET_LITTLE_ENDIAN ? 1 : 0)
#define LSW (TARGET_LITTLE_ENDIAN ? 0 : 1)
the compiler doesn't emit code to preserve all registers. */
static int pragma_trapa;
+/* This is set by #pragma nosave_low_regs. This is useful on the SH3,
+ which has a separate set of low regs for User and Supervisor modes.
+ This should only be used for the lowest level of interrupts. Higher levels
+ of interrupts must save the registers in case they themselves are
+ interrupted. */
+int pragma_nosave_low_regs;
+
/* This is used for communication between SETUP_INCOMING_VARARGS and
sh_expand_prologue. */
int current_function_anonymous_args;
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, PR_REGS, T_REGS, NO_REGS,
- MAC_REGS, MAC_REGS,
+ MAC_REGS, MAC_REGS, FPUL_REGS, GENERAL_REGS,
+ FP0_REGS,FP_REGS, FP_REGS, FP_REGS,
+ FP_REGS, FP_REGS, FP_REGS, FP_REGS,
+ FP_REGS, FP_REGS, FP_REGS, FP_REGS,
+ FP_REGS, FP_REGS, FP_REGS, FP_REGS,
};
/* Provide reg_class from a letter such as appears in the machine
enum reg_class reg_class_from_letter[] =
{
/* a */ NO_REGS, /* b */ NO_REGS, /* c */ NO_REGS, /* d */ NO_REGS,
- /* e */ NO_REGS, /* f */ NO_REGS, /* g */ NO_REGS, /* h */ NO_REGS,
+ /* e */ NO_REGS, /* f */ FP_REGS, /* g */ NO_REGS, /* h */ NO_REGS,
/* i */ NO_REGS, /* j */ NO_REGS, /* k */ NO_REGS, /* l */ PR_REGS,
/* m */ NO_REGS, /* n */ NO_REGS, /* o */ NO_REGS, /* p */ NO_REGS,
/* q */ NO_REGS, /* r */ NO_REGS, /* s */ NO_REGS, /* t */ T_REGS,
- /* u */ NO_REGS, /* v */ NO_REGS, /* w */ NO_REGS, /* x */ MAC_REGS,
- /* y */ NO_REGS, /* z */ R0_REGS
+ /* u */ NO_REGS, /* v */ NO_REGS, /* w */ FP0_REGS, /* x */ MAC_REGS,
+ /* y */ FPUL_REGS, /* z */ R0_REGS
};
\f
/* Print the operand address in x to the stream. */
rtx operands[];
enum machine_mode mode;
{
- /* Copy the source to a register if both operands aren't registers. */
- if (! reload_in_progress && ! reload_completed
- && ! register_operand (operands[0], mode)
- && ! register_operand (operands[1], mode))
- operands[1] = copy_to_mode_reg (mode, operands[1]);
+ if (! reload_in_progress && ! reload_completed)
+ {
+ /* Copy the source to a register if both operands aren't registers. */
+ if (! register_operand (operands[0], mode)
+ && ! register_operand (operands[1], mode))
+ operands[1] = copy_to_mode_reg (mode, operands[1]);
+
+ /* 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
+ && GET_CODE (operands[0]) == MEM
+ && GET_CODE (XEXP (operands[0], 0)) == PLUS
+ && GET_CODE (XEXP (XEXP (operands[0], 0), 1)) == REG)
+ operands[1] = copy_to_mode_reg (mode, operands[1]);
+ }
return 0;
}
{
rtx t_reg = gen_rtx (REG, SImode, T_REG);
enum rtx_code oldcode = code;
+ enum machine_mode mode;
/* First need a compare insn. */
switch (code)
sh_compare_op1 = tmp;
}
- sh_compare_op0 = force_reg (SImode, sh_compare_op0);
+ mode = GET_MODE (sh_compare_op0);
+ if (mode == VOIDmode)
+ mode = GET_MODE (sh_compare_op1);
+
+ sh_compare_op0 = force_reg (mode, sh_compare_op0);
if (code != EQ && code != NE
&& (sh_compare_op1 != const0_rtx
|| code == GTU || code == GEU || code == LTU || code == LEU))
- sh_compare_op1 = force_reg (SImode, sh_compare_op1);
+ sh_compare_op1 = force_reg (mode, sh_compare_op1);
+ /* ??? This should be `mode' not `SImode' in the compare, but that would
+ require fixing the branch patterns too. */
emit_insn (gen_rtx (SET, VOIDmode, t_reg,
gen_rtx (code, SImode, sh_compare_op0,
sh_compare_op1)));
{
if (code != EQ && code != NE)
{
+ enum machine_mode mode = GET_MODE (sh_compare_op0);
+ if (mode == VOIDmode)
+ mode = GET_MODE (sh_compare_op1);
+
/* Force args into regs, since we can't use constants here. */
- sh_compare_op0 = force_reg (SImode, sh_compare_op0);
+ sh_compare_op0 = force_reg (mode, sh_compare_op0);
if (sh_compare_op1 != const0_rtx
|| code == GTU || code == GEU || code == LTU || code == LEU)
- sh_compare_op1 = force_reg (SImode, sh_compare_op1);
+ sh_compare_op1 = force_reg (mode, sh_compare_op1);
}
operands[1] = sh_compare_op0;
operands[2] = sh_compare_op1;
rtx *operands;
{
int label = lf++;
+ int length = get_attr_length (insn);
+ int adjusted_length;
+
+ /* Undo the effects of ADJUST_INSN_LENGTH, so that we get the real
+ length. If NEXT_INSN (PREV_INSN (insn)) != insn, then the insn
+ is inside a sequence, and ADJUST_INSN_LENGTH was not called on
+ it. */
+ if (PREV_INSN (insn) == NULL
+ || NEXT_INSN (PREV_INSN (insn)) == insn)
+ {
+ adjusted_length = length;
+ ADJUST_INSN_LENGTH (insn, adjusted_length);
+ length -= (adjusted_length - length);
+ }
- switch (get_attr_length (insn))
+ switch (length)
{
case 2:
/* A branch with an unfilled delay slot. */
}
return "";
}
- return "bad";
-}
-\f
-/* A copy of the option structure defined in toplev.c. */
-
-struct option
-{
- char *string;
- int *variable;
- int on_value;
-};
-
-/* Output a single output option string NAME to FILE, without generating
- lines longer than MAX. */
-
-static int
-output_option (file, sep, type, name, indent, pos, max)
- FILE *file;
- char *sep;
- char *type;
- char *name;
- char *indent;
- int pos;
- int max;
-{
- if (strlen (sep) + strlen (type) + strlen (name) + pos > max)
- {
- fprintf (file, indent);
- return fprintf (file, "%s%s", type, name);
- }
- return pos + fprintf (file, "%s%s%s", sep, type, name);
-}
-
-/* A copy of the target_switches variable in toplev.c. */
-static struct
-{
- char *name;
- int value;
-} m_options[] = TARGET_SWITCHES;
-
-/* Output all options to the assembly language file. */
-
-static void
-output_options (file, f_options, f_len, W_options, W_len,
- pos, max, sep, indent, term)
- FILE *file;
- struct option *f_options;
- struct option *W_options;
- int f_len, W_len;
- int pos;
- int max;
- char *sep;
- char *indent;
- char *term;
-{
- register int j;
-
- if (optimize)
- pos = output_option (file, sep, "-O", "", indent, pos, max);
- if (write_symbols != NO_DEBUG)
- pos = output_option (file, sep, "-g", "", indent, pos, max);
- if (profile_flag)
- pos = output_option (file, sep, "-p", "", indent, pos, max);
- if (profile_block_flag)
- pos = output_option (file, sep, "-a", "", indent, pos, max);
-
- for (j = 0; j < f_len; j++)
- if (*f_options[j].variable == f_options[j].on_value)
- pos = output_option (file, sep, "-f", f_options[j].string,
- indent, pos, max);
-
- for (j = 0; j < W_len; j++)
- if (*W_options[j].variable == W_options[j].on_value)
- pos = output_option (file, sep, "-W", W_options[j].string,
- indent, pos, max);
-
- for (j = 0; j < sizeof m_options / sizeof m_options[0]; j++)
- if (m_options[j].name[0] != '\0'
- && m_options[j].value > 0
- && ((m_options[j].value & target_flags)
- == m_options[j].value))
- pos = output_option (file, sep, "-m", m_options[j].name,
- indent, pos, max);
-
- fprintf (file, term);
+ abort ();
}
-
+\f
/* Output to FILE the start of the assembler file. */
void
-output_file_start (file, f_options, f_len, W_options, W_len)
+output_file_start (file)
FILE *file;
- struct option *f_options;
- struct option *W_options;
- int f_len, W_len;
{
register int pos;
gcc2_compiled. symbol aren't in the text section. */
data_section ();
- pos = fprintf (file, "\n! Hitachi SH cc1 (%s) arguments:", version_string);
- output_options (file, f_options, f_len, W_options, W_len,
- pos, 75, " ", "\n! ", "\n\n");
-
if (TARGET_LITTLE_ENDIAN)
fprintf (file, "\t.little\n");
}
{16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2},
{16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}};
+/* Likewise, but for shift amounts < 16, up to three highmost bits
+ might be clobbered. This is typically used when combined with some
+ kind of sign or zero extension. */
+
+static char ext_shift_insns[] =
+ { 0,1,1,2,2,3,2,2,1,2,2,3,3,3,2,2,1,2,2,3,3,4,3,3,2,3,3,4,4,4,3,3};
+
+static short ext_shift_amounts[32][4] = {
+ {0}, {1}, {2}, {2, 1},
+ {2, 2}, {2, 1, 2}, {8, -2}, {8, -1},
+ {8}, {8, 1}, {8, 2}, {8, 1, 2},
+ {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1},
+ {16}, {16, 1}, {16, 2}, {16, 1, 2},
+ {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8},
+ {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2},
+ {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}};
+
+/* Assuming we have a value that has been sign-extended by at least one bit,
+ can we use the ext_shift_amounts with the last shift turned to an arithmetic shift
+ to shift it by N without data loss, and quicker than by other means? */
+#define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15)
+
/* This is used in length attributes in sh.md to help compute the length
of arbitrary constant shift instructions. */
/* Otherwise, return the true cost in instructions. */
if (GET_CODE (x) == ASHIFTRT)
- return ashiftrt_insns[value];
+ {
+ int cost = ashiftrt_insns[value];
+ /* If SH3, then we put the constant in a reg and use shad. */
+ if (TARGET_SH3 && cost > 3)
+ cost = 3;
+ return cost;
+ }
else
return shift_insns[value];
}
}
}
+/* Same for HImode */
+
+void
+gen_ashift_hi (type, n, reg)
+ int type;
+ int n;
+ rtx reg;
+{
+ /* Negative values here come from the shift_amounts array. */
+ if (n < 0)
+ {
+ if (type == ASHIFT)
+ type = LSHIFTRT;
+ else
+ type = ASHIFT;
+ n = -n;
+ }
+
+ switch (type)
+ {
+ case ASHIFTRT:
+ emit_insn (gen_ashrhi3_k (reg, reg, GEN_INT (n)));
+ break;
+ case LSHIFTRT:
+ if (n == 1)
+ emit_insn (gen_lshrhi3_m (reg, reg, GEN_INT (n)));
+ else
+ emit_insn (gen_lshrhi3_k (reg, reg, GEN_INT (n)));
+ break;
+ case ASHIFT:
+ emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n)));
+ break;
+ }
+}
+
/* Output RTL to split a constant shift into its component SH constant
shift instructions. */
-/* ??? For SH3, should reject constant shifts when slower than loading the
- shift count into a register? */
-
int
gen_shifty_op (code, operands)
int code;
int value = INTVAL (operands[2]);
int max, i;
+ /* Truncate the shift count in case it is out of bounds. */
+ value = value & 0x1f;
+
if (value == 31)
{
if (code == LSHIFTRT)
}
}
}
+ else if (value == 0)
+ {
+ /* This can happen when not optimizing. We must output something here
+ to prevent the compiler from aborting in final.c after the try_split
+ call. */
+ emit_insn (gen_nop ());
+ return;
+ }
max = shift_insns[value];
for (i = 0; i < max; i++)
gen_ashift (code, shift_amounts[value][i], operands[0]);
}
+
+/* Same as above, but optimized for values where the topmost bits don't
+ matter. */
+
+int
+gen_shifty_hi_op (code, operands)
+ int code;
+ rtx *operands;
+{
+ int value = INTVAL (operands[2]);
+ int max, i;
+ void (*gen_fun)();
+
+ /* This operation is used by and_shl for SImode values with a few
+ high bits known to be cleared. */
+ value &= 31;
+ if (value == 0)
+ {
+ emit_insn (gen_nop ());
+ return;
+ }
+
+ gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift;
+ if (code == ASHIFT)
+ {
+ max = ext_shift_insns[value];
+ for (i = 0; i < max; i++)
+ gen_fun (code, ext_shift_amounts[value][i], operands[0]);
+ }
+ else
+ /* When shifting right, emit the shifts in reverse order, so that
+ solitary negative values come first. */
+ for (i = ext_shift_insns[value] - 1; i >= 0; i--)
+ gen_fun (code, ext_shift_amounts[value][i], operands[0]);
+}
/* Output RTL for an arithmetic right shift. */
tree func_name;
int value;
- if (TARGET_SH3 && GET_CODE (operands[2]) != CONST_INT)
+ if (TARGET_SH3)
{
- rtx count = copy_to_mode_reg (SImode, operands[2]);
- emit_insn (gen_negsi2 (count, count));
- emit_insn (gen_ashrsi3_d (operands[0], operands[1], count));
- return 1;
+ if (GET_CODE (operands[2]) != CONST_INT)
+ {
+ rtx count = copy_to_mode_reg (SImode, operands[2]);
+ emit_insn (gen_negsi2 (count, count));
+ emit_insn (gen_ashrsi3_d (operands[0], operands[1], count));
+ return 1;
+ }
+ else if (ashiftrt_insns[INTVAL (operands[2])] > 3)
+ {
+ rtx count = force_reg (SImode, GEN_INT (- INTVAL (operands[2])));
+ emit_insn (gen_ashrsi3_d (operands[0], operands[1], count));
+ return 1;
+ }
}
if (GET_CODE (operands[2]) != CONST_INT)
return 0;
emit_move_insn (operands[0], gen_rtx (REG, SImode, 4));
return 1;
}
+
+/* Try to find a good way to implement the combiner pattern
+ [(set (match_operand:SI 0 "register_operand" "r")
+ (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r")
+ (match_operand:SI 2 "const_int_operand" "n"))
+ (match_operand:SI 3 "const_int_operand" "n"))) .
+ LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3.
+ return 0 for simple right / left or left/right shift combination.
+ return 1 for a combination of shifts with zero_extend.
+ return 2 for a combination of shifts with an AND that needs r0.
+ return 3 for a combination of shifts with an AND that needs an extra
+ scratch register, when the three highmost bits of the AND mask are clear.
+ return 4 for a combination of shifts with an AND that needs an extra
+ scratch register, when any of the three highmost bits of the AND mask
+ is set.
+ If ATTRP is set, store an initial right shift width in ATTRP[0],
+ and the instruction length in ATTRP[1] . These values are not valid
+ when returning 0.
+ When ATTRP is set and returning 1, ATTRP[2] gets set to the index into
+ shift_amounts for the last shift value that is to be used before the
+ sign extend. */
+int
+shl_and_kind (left_rtx, mask_rtx, attrp)
+ rtx left_rtx, mask_rtx;
+ int *attrp;
+{
+ unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2;
+ int left = INTVAL (left_rtx), right;
+ int best = 0;
+ int cost, best_cost = 10000;
+ int best_right = 0, best_len = 0;
+ int i;
+ int can_ext;
+
+ if (left < 0 || left > 31)
+ return 0;
+ if (GET_CODE (mask_rtx) == CONST_INT)
+ mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left;
+ else
+ mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left;
+ /* Can this be expressed as a right shift / left shift pair ? */
+ lsb = ((mask ^ (mask - 1)) >> 1) + 1;
+ right = exact_log2 (lsb);
+ mask2 = ~(mask + lsb - 1);
+ lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1;
+ /* mask has no zeroes but trailing zeroes <==> ! mask2 */
+ if (! mask2)
+ best_cost = shift_insns[right] + shift_insns[right + left];
+ /* mask has no trailing zeroes <==> ! right */
+ else if (! right && mask2 == ~(lsb2 - 1))
+ {
+ int late_right = exact_log2 (lsb2);
+ best_cost = shift_insns[left + late_right] + shift_insns[late_right];
+ }
+ /* Try to use zero extend */
+ if (mask2 == ~(lsb2 - 1))
+ {
+ int width, first;
+
+ for (width = 8; width <= 16; width += 8)
+ {
+ /* Can we zero-extend right away? */
+ if (lsb2 == (HOST_WIDE_INT)1 << width)
+ {
+ cost
+ = 1 + ext_shift_insns[right] + ext_shift_insns[left + right];
+ if (cost < best_cost)
+ {
+ best = 1;
+ best_cost = cost;
+ best_right = right;
+ best_len = cost;
+ if (attrp)
+ attrp[2] = -1;
+ }
+ continue;
+ }
+ /* ??? Could try to put zero extend into initial right shift,
+ or even shift a bit left before the right shift. */
+ /* Determine value of first part of left shift, to get to the
+ zero extend cut-off point. */
+ first = width - exact_log2 (lsb2) + right;
+ if (first >= 0 && right + left - first >= 0)
+ {
+ cost = ext_shift_insns[right] + ext_shift_insns[first] + 1
+ + ext_shift_insns[right + left - first];
+ if (cost < best_cost)
+ {
+ best = 1;
+ best_cost = cost;
+ best_right = right;
+ best_len = cost;
+ if (attrp)
+ attrp[2] = first;
+ }
+ }
+ }
+ }
+ /* Try to use r0 AND pattern */
+ for (i = 0; i <= 2; i++)
+ {
+ if (i > right)
+ break;
+ if (! CONST_OK_FOR_L (mask >> i))
+ continue;
+ cost = (i != 0) + 2 + ext_shift_insns[left + i];
+ if (cost < best_cost)
+ {
+ best = 2;
+ best_cost = cost;
+ best_right = i;
+ best_len = cost - 1;
+ }
+ }
+ /* Try to use a scratch register to hold the AND operand. */
+ can_ext = ((mask << left) & 0xe0000000) == 0;
+ for (i = 0; i <= 2; i++)
+ {
+ if (i > right)
+ break;
+ cost = (i != 0) + (CONST_OK_FOR_I (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);
+ }
+ }
+
+ if (attrp)
+ {
+ attrp[0] = best_right;
+ attrp[1] = best_len;
+ }
+ return best;
+}
+
+/* This is used in length attributes of the unnamed instructions
+ corresponding to shl_and_kind return values of 1 and 2. */
+int
+shl_and_length (insn)
+ rtx insn;
+{
+ rtx set_src, left_rtx, mask_rtx;
+ int attributes[3];
+
+ set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
+ left_rtx = XEXP (XEXP (set_src, 0), 1);
+ mask_rtx = XEXP (set_src, 1);
+ shl_and_kind (left_rtx, mask_rtx, attributes);
+ return attributes[1];
+}
+
+/* This is used in length attribute of the and_shl_scratch instruction. */
+
+int
+shl_and_scr_length (insn)
+ rtx insn;
+{
+ rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
+ int len = shift_insns[INTVAL (XEXP (set_src, 1))];
+ rtx op = XEXP (set_src, 0);
+ len += shift_insns[INTVAL (XEXP (op, 1))] + 1;
+ op = XEXP (XEXP (op, 0), 0);
+ return len + shift_insns[INTVAL (XEXP (op, 1))];
+}
+
+/* Generating rtl? */
+extern int rtx_equal_function_value_matters;
+
+/* Generate rtl for instructions for which shl_and_kind advised a particular
+ method of generating them, i.e. returned zero. */
+
+int
+gen_shl_and (dest, left_rtx, mask_rtx, source)
+ rtx dest, left_rtx, mask_rtx, source;
+{
+ int attributes[3];
+ unsigned HOST_WIDE_INT mask;
+ int kind = shl_and_kind (left_rtx, mask_rtx, attributes);
+ int right, total_shift;
+ int (*shift_gen_fun) PROTO((int, rtx*)) = gen_shifty_hi_op;
+
+ right = attributes[0];
+ total_shift = INTVAL (left_rtx) + right;
+ mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift;
+ switch (kind)
+ {
+ default:
+ return -1;
+ case 1:
+ {
+ int first = attributes[2];
+ rtx operands[3];
+
+ if (first < 0)
+ {
+ emit_insn ((mask << right) == 0xff
+ ? gen_zero_extendqisi2(dest,
+ gen_lowpart (QImode, source))
+ : gen_zero_extendhisi2(dest,
+ gen_lowpart (HImode, source)));
+ source = dest;
+ }
+ if (source != dest)
+ emit_insn (gen_movsi (dest, source));
+ operands[0] = dest;
+ if (right)
+ {
+ operands[2] = GEN_INT (right);
+ gen_shifty_hi_op (LSHIFTRT, operands);
+ }
+ if (first > 0)
+ {
+ operands[2] = GEN_INT (first);
+ gen_shifty_hi_op (ASHIFT, operands);
+ total_shift -= first;
+ mask <<= first;
+ }
+ if (first >= 0)
+ emit_insn (mask == 0xff
+ ? gen_zero_extendqisi2(dest, gen_lowpart (QImode, dest))
+ : gen_zero_extendhisi2(dest, gen_lowpart (HImode, dest)));
+ if (total_shift > 0)
+ {
+ operands[2] = GEN_INT (total_shift);
+ gen_shifty_hi_op (ASHIFT, operands);
+ }
+ break;
+ }
+ case 4:
+ shift_gen_fun = gen_shifty_op;
+ case 2:
+ case 3:
+ /* If the topmost bit that matters is set, set the topmost bits
+ that don't matter. This way, we might be able to get a shorter
+ signed constant. */
+ if (mask & ((HOST_WIDE_INT)1 << 31 - total_shift))
+ mask |= (HOST_WIDE_INT)~0 << (31 - total_shift);
+ /* Don't expand fine-grained when combining, because that will
+ make the pattern fail. */
+ if (rtx_equal_function_value_matters
+ || reload_in_progress || reload_completed)
+ {
+ rtx operands[3];
+
+ if (right)
+ {
+ emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right)));
+ source = dest;
+ }
+ emit_insn (gen_andsi3 (dest, source, GEN_INT (mask)));
+ if (total_shift)
+ {
+ operands[0] = dest;
+ operands[1] = dest;
+ operands[2] = GEN_INT (total_shift);
+ shift_gen_fun (ASHIFT, operands);
+ }
+ break;
+ }
+ else
+ {
+ int neg = 0;
+ if (kind != 4 && total_shift < 16)
+ {
+ neg = -ext_shift_amounts[total_shift][1];
+ if (neg > 0)
+ neg -= ext_shift_amounts[total_shift][2];
+ else
+ neg = 0;
+ }
+ emit_insn (gen_and_shl_scratch (dest, source,
+ GEN_INT (right),
+ GEN_INT (mask),
+ GEN_INT (total_shift + neg),
+ GEN_INT (neg)));
+ emit_insn (gen_movsi (dest, dest));
+ break;
+ }
+ }
+ return 0;
+}
+
+/* Try to find a good way to implement the combiner pattern
+ [(set (match_operand:SI 0 "register_operand" "=r")
+ (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r")
+ (match_operand:SI 2 "const_int_operand" "n")
+ (match_operand:SI 3 "const_int_operand" "n")
+ (const_int 0)))
+ (clobber (reg:SI 18))]
+ LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3.
+ return 0 for simple left / right shift combination.
+ return 1 for left shift / 8 bit sign extend / left shift.
+ return 2 for left shift / 16 bit sign extend / left shift.
+ return 3 for left shift / 8 bit sign extend / shift / sign extend.
+ return 4 for left shift / 16 bit sign extend / shift / sign extend.
+ return 5 for left shift / 16 bit sign extend / right shift
+ return 6 for < 8 bit sign extend / left shift.
+ return 7 for < 8 bit sign extend / left shift / single right shift.
+ If COSTP is nonzero, assign the calculated cost to *COSTP. */
+
+int
+shl_sext_kind (left_rtx, size_rtx, costp)
+ rtx left_rtx, size_rtx;
+ int *costp;
+{
+ int left, size, insize, ext;
+ int cost, best_cost;
+ int kind;
+
+ left = INTVAL (left_rtx);
+ size = INTVAL (size_rtx);
+ insize = size - left;
+ if (insize <= 0)
+ abort ();
+ /* Default to left / right shift. */
+ kind = 0;
+ best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size];
+ if (size <= 16)
+ {
+ /* 16 bit shift / sign extend / 16 bit shift */
+ cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size];
+ /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden
+ below, by alternative 3 or something even better. */
+ if (cost < best_cost)
+ {
+ kind = 5;
+ best_cost = cost;
+ }
+ }
+ /* Try a plain sign extend between two shifts. */
+ for (ext = 16; ext >= insize; ext -= 8)
+ {
+ if (ext <= size)
+ {
+ cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext];
+ if (cost < best_cost)
+ {
+ kind = ext / 8U;
+ best_cost = cost;
+ }
+ }
+ /* Check if we can do a sloppy shift with a final signed shift
+ restoring the sign. */
+ if (EXT_SHIFT_SIGNED (size - ext))
+ cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1;
+ /* If not, maybe it's still cheaper to do the second shift sloppy,
+ and do a final sign extend? */
+ else if (size <= 16)
+ cost = ext_shift_insns[ext - insize] + 1
+ + ext_shift_insns[size > ext ? size - ext : ext - size] + 1;
+ else
+ continue;
+ if (cost < best_cost)
+ {
+ kind = ext / 8U + 2;
+ best_cost = cost;
+ }
+ }
+ /* Check if we can sign extend in r0 */
+ if (insize < 8)
+ {
+ cost = 3 + shift_insns[left];
+ if (cost < best_cost)
+ {
+ kind = 6;
+ best_cost = cost;
+ }
+ /* Try the same with a final signed shift. */
+ if (left < 31)
+ {
+ cost = 3 + ext_shift_insns[left + 1] + 1;
+ if (cost < best_cost)
+ {
+ kind = 7;
+ best_cost = cost;
+ }
+ }
+ }
+ if (TARGET_SH3)
+ {
+ /* Try to use a dynamic shift. */
+ cost = shift_insns[32 - insize] + 3;
+ if (cost < best_cost)
+ {
+ kind = 0;
+ best_cost = cost;
+ }
+ }
+ if (costp)
+ *costp = cost;
+ return kind;
+}
+
+/* Function to be used in the length attribute of the instructions
+ implementing this pattern. */
+
+int
+shl_sext_length (insn)
+ rtx insn;
+{
+ rtx set_src, left_rtx, size_rtx;
+ int cost;
+
+ set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
+ left_rtx = XEXP (XEXP (set_src, 0), 1);
+ size_rtx = XEXP (set_src, 1);
+ shl_sext_kind (left_rtx, size_rtx, &cost);
+ return cost;
+}
+
+/* Generate rtl for this pattern */
+
+int
+gen_shl_sext (dest, left_rtx, size_rtx, source)
+ rtx dest, left_rtx, size_rtx, source;
+{
+ int kind;
+ int left, size, insize, cost;
+ rtx operands[3];
+
+ kind = shl_sext_kind (left_rtx, size_rtx, &cost);
+ left = INTVAL (left_rtx);
+ size = INTVAL (size_rtx);
+ insize = size - left;
+ switch (kind)
+ {
+ case 1:
+ case 2:
+ case 3:
+ case 4:
+ {
+ int ext = kind & 1 ? 8 : 16;
+ int shift2 = size - ext;
+
+ /* Don't expand fine-grained when combining, because that will
+ make the pattern fail. */
+ if (! rtx_equal_function_value_matters
+ && ! reload_in_progress && ! reload_completed)
+ {
+ emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx));
+ emit_insn (gen_movsi (dest, source));
+ break;
+ }
+ if (dest != source)
+ emit_insn (gen_movsi (dest, source));
+ operands[0] = dest;
+ if (ext - insize)
+ {
+ operands[2] = GEN_INT (ext - insize);
+ gen_shifty_hi_op (ASHIFT, operands);
+ }
+ emit_insn (kind & 1
+ ? gen_extendqisi2(dest, gen_lowpart (QImode, dest))
+ : gen_extendhisi2(dest, gen_lowpart (HImode, dest)));
+ if (kind <= 2)
+ {
+ if (shift2)
+ {
+ operands[2] = GEN_INT (shift2);
+ gen_shifty_op (ASHIFT, operands);
+ }
+ }
+ else
+ {
+ if (shift2 > 0)
+ {
+ if (EXT_SHIFT_SIGNED (shift2))
+ {
+ operands[2] = GEN_INT (shift2 + 1);
+ gen_shifty_op (ASHIFT, operands);
+ operands[2] = GEN_INT (1);
+ gen_shifty_op (ASHIFTRT, operands);
+ break;
+ }
+ operands[2] = GEN_INT (shift2);
+ gen_shifty_hi_op (ASHIFT, operands);
+ }
+ else if (shift2)
+ {
+ operands[2] = GEN_INT (-shift2);
+ gen_shifty_hi_op (LSHIFTRT, operands);
+ }
+ emit_insn (size <= 8
+ ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest))
+ : gen_extendhisi2 (dest, gen_lowpart (HImode, dest)));
+ }
+ break;
+ }
+ case 5:
+ {
+ int i = 16 - size;
+ emit_insn (gen_shl_sext_ext (dest, source, GEN_INT (16 - insize),
+ GEN_INT (16)));
+ /* Don't use gen_ashrsi3 because it generates new pseudos. */
+ while (--i >= 0)
+ gen_ashift (ASHIFTRT, 1, dest);
+ break;
+ }
+ case 6:
+ case 7:
+ /* Don't expand fine-grained when combining, because that will
+ make the pattern fail. */
+ if (! rtx_equal_function_value_matters
+ && ! reload_in_progress && ! reload_completed)
+ {
+ emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx));
+ emit_insn (gen_movsi (dest, source));
+ break;
+ }
+ emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1)));
+ emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1))));
+ emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1))));
+ operands[0] = dest;
+ operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx;
+ gen_shifty_op (ASHIFT, operands);
+ if (kind == 7)
+ emit_insn (gen_ashrsi3_k (dest, dest, GEN_INT (1)));
+ break;
+ default:
+ return -1;
+ }
+ return 0;
+}
\f
/* The SH cannot load a large constant into a register, constants have to
come from a pc relative load. The reference of a pc relative load
case HImode:
break;
case SImode:
+ case SFmode:
if (need_align)
{
need_align = 0;
scan = emit_label_after (p->label, scan);
scan = emit_insn_after (gen_consttable_4 (p->value), scan);
break;
+ case DFmode:
case DImode:
if (need_align)
{
broken_move (insn)
rtx insn;
{
- if (GET_CODE (insn) == INSN
- && GET_CODE (PATTERN (insn)) == SET
- /* We can load any 8 bit value if we don't care what the high
- order bits end up as. */
- && GET_MODE (SET_DEST (PATTERN (insn))) != QImode
- && CONSTANT_P (SET_SRC (PATTERN (insn)))
- && (GET_CODE (SET_SRC (PATTERN (insn))) != CONST_INT
- || ! CONST_OK_FOR_I (INTVAL (SET_SRC (PATTERN (insn))))))
- return 1;
+ if (GET_CODE (insn) == INSN)
+ {
+ rtx pat = PATTERN (insn);
+ if (GET_CODE (pat) == PARALLEL)
+ pat = XVECEXP (pat, 0, 0);
+ if (GET_CODE (pat) == SET
+ /* We can load any 8 bit value if we don't care what the high
+ order bits end up as. */
+ && GET_MODE (SET_DEST (pat)) != QImode
+ && CONSTANT_P (SET_SRC (pat))
+ && ! (GET_CODE (SET_SRC (pat)) == CONST_DOUBLE
+ && (fp_zero_operand (SET_SRC (pat))
+ || fp_one_operand (SET_SRC (pat)))
+ && GET_CODE (SET_DEST (pat)) == REG
+ && REGNO (SET_DEST (pat)) >= FIRST_FP_REG
+ && REGNO (SET_DEST (pat)) <= LAST_FP_REG)
+ && (GET_CODE (SET_SRC (pat)) != CONST_INT
+ || ! CONST_OK_FOR_I (INTVAL (SET_SRC (pat)))))
+ return 1;
+ }
return 0;
}
int found_si = 0;
rtx found_barrier = 0;
rtx found_mova = 0;
+ int si_limit;
+ int hi_limit;
/* For HImode: range is 510, add 4 because pc counts from address of
second instruction after this one, subtract 2 for the jump instruction
- that we may need to emit before the table. This gives 512.
+ that we may need to emit before the table, subtract 2 for the instruction
+ that fills the jump delay slot (in very rare cases, reorg will take an
+ instruction from after the constant pool or will leave the delay slot
+ empty). This gives 510.
For SImode: range is 1020, add 4 because pc counts from address of
second instruction after this one, subtract 2 in case pc is 2 byte
aligned, subtract 2 for the jump instruction that we may need to emit
- before the table. This gives 1020. */
- while (from && count_si < 1020 && count_hi < 512)
+ before the table, subtract 2 for the instruction that fills the jump
+ delay slot. This gives 1018. */
+
+ /* If not optimizing, then it is possible that the jump instruction we add
+ won't be shortened, and thus will have a length of 14 instead of 2.
+ We must adjust the limits downwards to account for this, giving a limit
+ of 1008 for SImode and 500 for HImode. */
+
+ if (optimize)
+ {
+ si_limit = 1018;
+ hi_limit = 510;
+ }
+ else
+ {
+ si_limit = 1008;
+ hi_limit = 500;
+ }
+
+ /* If not optimizing for space, then the constant pool will be
+ aligned to a 4 to 16 byte boundary. We must make room for that
+ alignment that by reducing the limits.
+ ??? It would be better to not align the constant pool, but
+ ASM_OUTPUT_ALIGN_CODE does not make any provision for basing the
+ alignment on the instruction. */
+
+ if (! TARGET_SMALLCODE)
+ {
+ if (TARGET_SH3 || TARGET_SH3E)
+ {
+ si_limit -= 14;
+ hi_limit -= 14;
+ }
+ else
+ {
+ si_limit -= 2;
+ hi_limit -= 2;
+ }
+ }
+
+ while (from && count_si < si_limit && count_hi < hi_limit)
{
int inc = get_attr_length (from);
if (broken_move (from))
{
- rtx src = SET_SRC (PATTERN (from));
-
- if (hi_const (src))
+ rtx pat = PATTERN (from);
+ rtx src = SET_SRC (pat);
+ rtx dst = SET_DEST (pat);
+ enum machine_mode mode = GET_MODE (dst);
+
+ /* We must explicitly check the mode, because sometimes the
+ front end will generate code to load unsigned constants into
+ HImode targets without properly sign extending them. */
+ if (mode == HImode || (mode == SImode && hi_const (src)))
{
found_hi = 1;
/* We put the short constants before the long constants, so
/* If we exceeded the range, then we must back up over the last
instruction we looked at. Otherwise, we just need to undo the
NEXT_INSN at the end of the loop. */
- if (count_hi > 512 || count_si > 1020)
+ if (count_hi > hi_limit || count_si > si_limit)
from = PREV_INSN (PREV_INSN (from));
else
from = PREV_INSN (from);
from = emit_jump_insn_after (gen_jump (label), from);
JUMP_LABEL (from) = label;
+ LABEL_NUSES (label) = 1;
found_barrier = emit_barrier_after (from);
emit_label_after (label, found_barrier);
}
return found_barrier;
}
+/* If the instruction INSN is implemented by a special function, and we can
+ positively find the register that is used to call the sfunc, and this
+ register is not used anywhere else in this instruction - except as the
+ destination of a set, return this register; else, return 0. */
+static rtx
+sfunc_uses_reg (insn)
+ rtx insn;
+{
+ int i;
+ rtx pattern, part, reg_part, reg;
+
+ if (GET_CODE (insn) != INSN)
+ return 0;
+ pattern = PATTERN (insn);
+ if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC)
+ return 0;
+
+ for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--)
+ {
+ part = XVECEXP (pattern, 0, i);
+ if (GET_CODE (part) == USE)
+ reg_part = part;
+ }
+ if (! reg_part)
+ return 0;
+ reg = XEXP (reg_part, 0);
+ for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--)
+ {
+ part = XVECEXP (pattern, 0, i);
+ if (part == reg_part)
+ continue;
+ if (reg_mentioned_p (reg, ((GET_CODE (part) == SET
+ && GET_CODE (SET_DEST (part)) == REG)
+ ? SET_SRC (part) : part)))
+ return 0;
+ }
+ return reg;
+}
+
+/* See if the only way in which INSN uses REG is by calling it, or by
+ setting it while calling it. Set *SET to a SET rtx if the register
+ is set by INSN. */
+
+static int
+noncall_uses_reg (reg, insn, set)
+ rtx reg;
+ rtx insn;
+ rtx *set;
+{
+ rtx pattern, reg2;
+
+ *set = NULL_RTX;
+
+ reg2 = sfunc_uses_reg (insn);
+ if (reg2 && REGNO (reg2) == REGNO (reg))
+ {
+ pattern = single_set (insn);
+ if (pattern
+ && GET_CODE (SET_DEST (pattern)) == REG
+ && REGNO (reg) == REGNO (SET_DEST (pattern)))
+ *set = pattern;
+ return 0;
+ }
+ if (GET_CODE (insn) != CALL_INSN)
+ {
+ /* We don't use rtx_equal_p because we don't care if the mode is
+ different. */
+ pattern = single_set (insn);
+ if (pattern
+ && GET_CODE (SET_DEST (pattern)) == REG
+ && REGNO (reg) == REGNO (SET_DEST (pattern)))
+ {
+ rtx par, part;
+ int i;
+
+ *set = pattern;
+ par = PATTERN (insn);
+ if (GET_CODE (par) == PARALLEL)
+ for (i = XVECLEN (par, 0) - 1; i >= 0; i--)
+ {
+ part = XVECEXP (par, 0, i);
+ if (GET_CODE (part) != SET && reg_mentioned_p (reg, part))
+ return 1;
+ }
+ return reg_mentioned_p (reg, SET_SRC (pattern));
+ }
+
+ return 1;
+ }
+
+ pattern = PATTERN (insn);
+
+ if (GET_CODE (pattern) == PARALLEL)
+ {
+ int i;
+
+ for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--)
+ if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i)))
+ return 1;
+ pattern = XVECEXP (pattern, 0, 0);
+ }
+
+ if (GET_CODE (pattern) == SET)
+ {
+ if (reg_mentioned_p (reg, SET_DEST (pattern)))
+ {
+ /* We don't use rtx_equal_p, because we don't care if the
+ mode is different. */
+ if (GET_CODE (SET_DEST (pattern)) != REG
+ || REGNO (reg) != REGNO (SET_DEST (pattern)))
+ return 1;
+
+ *set = pattern;
+ }
+
+ pattern = SET_SRC (pattern);
+ }
+
+ if (GET_CODE (pattern) != CALL
+ || GET_CODE (XEXP (pattern, 0)) != MEM
+ || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0)))
+ return 1;
+
+ return 0;
+}
+
/* Exported to toplev.c.
- Scan the function looking for move instructions which have to be changed to
- pc-relative loads and insert the literal tables. */
+ Do a final pass over the function, just before delayed branch
+ scheduling. */
void
machine_dependent_reorg (first)
{
rtx insn;
+ /* If relaxing, generate pseudo-ops to associate function calls with
+ the symbols they call. It does no harm to not generate these
+ pseudo-ops. However, when we can generate them, it enables to
+ linker to potentially relax the jsr to a bsr, and eliminate the
+ register load and, possibly, the constant pool entry. */
+
+ if (TARGET_RELAX)
+ {
+ /* Remove all REG_LABEL notes. We want to use them for our own
+ purposes. This works because none of the remaining passes
+ need to look at them.
+
+ ??? But it may break in the future. We should use a machine
+ dependent REG_NOTE, or some other approach entirely. */
+ for (insn = first; insn; insn = NEXT_INSN (insn))
+ {
+ if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
+ {
+ rtx note;
+
+ while ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)) != 0)
+ remove_note (insn, note);
+ }
+ }
+
+ for (insn = first; insn; insn = NEXT_INSN (insn))
+ {
+ rtx pattern, reg, link, set, scan, dies, label;
+ int rescan = 0, foundinsn = 0;
+
+ if (GET_CODE (insn) == CALL_INSN)
+ {
+ pattern = PATTERN (insn);
+
+ if (GET_CODE (pattern) == PARALLEL)
+ pattern = XVECEXP (pattern, 0, 0);
+ if (GET_CODE (pattern) == SET)
+ pattern = SET_SRC (pattern);
+
+ if (GET_CODE (pattern) != CALL
+ || GET_CODE (XEXP (pattern, 0)) != MEM)
+ continue;
+
+ reg = XEXP (XEXP (pattern, 0), 0);
+ }
+ else
+ {
+ reg = sfunc_uses_reg (insn);
+ if (! reg)
+ continue;
+ }
+
+ if (GET_CODE (reg) != REG)
+ continue;
+
+ /* This is a function call via REG. If the only uses of REG
+ between the time that it is set and the time that it dies
+ are in function calls, then we can associate all the
+ function calls with the setting of REG. */
+
+ for (link = LOG_LINKS (insn); link; link = XEXP (link, 1))
+ {
+ if (REG_NOTE_KIND (link) != 0)
+ continue;
+ set = single_set (XEXP (link, 0));
+ if (set && rtx_equal_p (reg, SET_DEST (set)))
+ {
+ link = XEXP (link, 0);
+ break;
+ }
+ }
+
+ if (! link)
+ {
+ /* ??? Sometimes global register allocation will have
+ deleted the insn pointed to by LOG_LINKS. Try
+ scanning backward to find where the register is set. */
+ for (scan = PREV_INSN (insn);
+ scan && GET_CODE (scan) != CODE_LABEL;
+ scan = PREV_INSN (scan))
+ {
+ if (GET_RTX_CLASS (GET_CODE (scan)) != 'i')
+ continue;
+
+ if (! reg_mentioned_p (reg, scan))
+ continue;
+
+ if (noncall_uses_reg (reg, scan, &set))
+ break;
+
+ if (set)
+ {
+ link = scan;
+ break;
+ }
+ }
+ }
+
+ if (! link)
+ continue;
+
+ /* The register is set at LINK. */
+
+ /* We can only optimize the function call if the register is
+ being set to a symbol. In theory, we could sometimes
+ optimize calls to a constant location, but the assembler
+ and linker do not support that at present. */
+ if (GET_CODE (SET_SRC (set)) != SYMBOL_REF
+ && GET_CODE (SET_SRC (set)) != LABEL_REF)
+ continue;
+
+ /* Scan forward from LINK to the place where REG dies, and
+ make sure that the only insns which use REG are
+ themselves function calls. */
+
+ /* ??? This doesn't work for call targets that were allocated
+ by reload, since there may not be a REG_DEAD note for the
+ register. */
+
+ dies = NULL_RTX;
+ for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan))
+ {
+ rtx scanset;
+
+ /* Don't try to trace forward past a CODE_LABEL if we haven't
+ seen INSN yet. Ordinarily, we will only find the setting insn
+ in LOG_LINKS if it is in the same basic block. However,
+ cross-jumping can insert code labels in between the load and
+ the call, and can result in situations where a single call
+ insn may have two targets depending on where we came from. */
+
+ if (GET_CODE (scan) == CODE_LABEL && ! foundinsn)
+ break;
+
+ if (GET_RTX_CLASS (GET_CODE (scan)) != 'i')
+ continue;
+
+ /* Don't try to trace forward past a JUMP. To optimize
+ safely, we would have to check that all the
+ instructions at the jump destination did not use REG. */
+
+ if (GET_CODE (scan) == JUMP_INSN)
+ break;
+
+ if (! reg_mentioned_p (reg, scan))
+ continue;
+
+ if (noncall_uses_reg (reg, scan, &scanset))
+ break;
+
+ if (scan == insn)
+ foundinsn = 1;
+
+ if (scan != insn
+ && (GET_CODE (scan) == CALL_INSN || sfunc_uses_reg (scan)))
+ {
+ /* There is a function call to this register other
+ than the one we are checking. If we optimize
+ this call, we need to rescan again below. */
+ rescan = 1;
+ }
+
+ /* ??? We shouldn't have to worry about SCANSET here.
+ We should just be able to check for a REG_DEAD note
+ on a function call. However, the REG_DEAD notes are
+ apparently not dependable around libcalls; c-torture
+ execute/920501-2 is a test case. If SCANSET is set,
+ then this insn sets the register, so it must have
+ died earlier. Unfortunately, this will only handle
+ the cases in which the register is, in fact, set in a
+ later insn. */
+
+ /* ??? We shouldn't have to use FOUNDINSN here.
+ However, the LOG_LINKS fields are apparently not
+ entirely reliable around libcalls;
+ newlib/libm/math/e_pow.c is a test case. Sometimes
+ an insn will appear in LOG_LINKS even though it is
+ not the most recent insn which sets the register. */
+
+ if (foundinsn
+ && (scanset
+ || find_reg_note (scan, REG_DEAD, reg)))
+ {
+ dies = scan;
+ break;
+ }
+ }
+
+ if (! dies)
+ {
+ /* Either there was a branch, or some insn used REG
+ other than as a function call address. */
+ continue;
+ }
+
+ /* Create a code label, and put it in a REG_LABEL note on
+ the insn which sets the register, and on each call insn
+ which uses the register. In final_prescan_insn we look
+ for the REG_LABEL notes, and output the appropriate label
+ or pseudo-op. */
+
+ label = gen_label_rtx ();
+ REG_NOTES (link) = gen_rtx (EXPR_LIST, REG_LABEL, label,
+ REG_NOTES (link));
+ REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, label,
+ REG_NOTES (insn));
+ if (rescan)
+ {
+ scan = link;
+ do
+ {
+ rtx reg2;
+
+ scan = NEXT_INSN (scan);
+ if (scan != insn
+ && ((GET_CODE (scan) == CALL_INSN
+ && reg_mentioned_p (reg, scan))
+ || ((reg2 = sfunc_uses_reg (scan))
+ && REGNO (reg2) == REGNO (reg))))
+ REG_NOTES (scan) = gen_rtx (EXPR_LIST, REG_LABEL,
+ label, REG_NOTES (scan));
+ }
+ while (scan != dies);
+ }
+ }
+ }
+
+ /* Scan the function looking for move instructions which have to be
+ changed to pc-relative loads and insert the literal tables. */
+
for (insn = first; insn; insn = NEXT_INSN (insn))
{
if (broken_move (insn))
{
if (broken_move (scan))
{
- rtx pat = PATTERN (scan);
- rtx src = SET_SRC (pat);
- rtx dst = SET_DEST (pat);
- enum machine_mode mode = GET_MODE (dst);
+ rtx *patp = &PATTERN (scan), pat = *patp;
+ rtx src, dst;
rtx lab;
rtx newinsn;
rtx newsrc;
+ enum machine_mode mode;
+
+ if (GET_CODE (pat) == PARALLEL)
+ patp = &XVECEXP (pat, 0, 0), pat = *patp;
+ src = SET_SRC (pat);
+ dst = SET_DEST (pat);
+ mode = GET_MODE (dst);
if (mode == SImode && hi_const (src))
{
newsrc = gen_rtx (MEM, mode,
gen_rtx (LABEL_REF, VOIDmode, lab));
RTX_UNCHANGING_P (newsrc) = 1;
- newinsn = emit_insn_after (gen_rtx (SET, VOIDmode,
- dst, newsrc), scan);
-
- delete_insn (scan);
- scan = newinsn;
+ *patp = gen_rtx (SET, VOIDmode, dst, newsrc);
+ INSN_CODE (scan) = -1;
}
}
dump_table (barrier);
}
/* Dump out instruction addresses, which is useful for debugging the
- constant pool table stuff. */
+ constant pool table stuff.
+
+ If relaxing, output the label and pseudo-ops used to link together
+ calls and the instruction which set the registers. */
/* ??? This is unnecessary, and probably should be deleted. This makes
the insn_addresses declaration above unnecessary. */
{
if (TARGET_DUMPISIZE)
fprintf (asm_out_file, "\n! at %04x\n", insn_addresses[INSN_UID (insn)]);
+
+ if (TARGET_RELAX)
+ {
+ rtx note;
+
+ note = find_reg_note (insn, REG_LABEL, NULL_RTX);
+ if (note)
+ {
+ rtx pattern;
+
+ pattern = PATTERN (insn);
+ if (GET_CODE (pattern) == PARALLEL)
+ pattern = XVECEXP (pattern, 0, 0);
+ if (GET_CODE (pattern) == CALL
+ || (GET_CODE (pattern) == SET
+ && (GET_CODE (SET_SRC (pattern)) == CALL
+ || get_attr_type (insn) == TYPE_SFUNC)))
+ fprintf (asm_out_file, "\t.uses L%d\n",
+ CODE_LABEL_NUMBER (XEXP (note, 0)));
+ else if (GET_CODE (pattern) == SET)
+ ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L",
+ CODE_LABEL_NUMBER (XEXP (note, 0)));
+ else
+ abort ();
+ }
+ }
}
/* Dump out any constants accumulated in the final pass. These will
static int extra_push;
-/* Adjust the stack and return the number of bytes taken to do it. */
+/* 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. */
static void
-output_stack_adjust (size, reg)
+output_stack_adjust (size, reg, temp)
int size;
rtx reg;
+ int temp;
{
if (size)
{
if (! CONST_OK_FOR_I (size))
{
- rtx reg = gen_rtx (REG, SImode, 3);
- emit_insn (gen_movsi (reg, val));
- val = reg;
+ /* Try to do it with two partial adjustments; however, must make
+ sure that the stack is properly aligned at all times, in case
+ an interrupt occurs between the two partial adjustments. */
+ if (CONST_OK_FOR_I (size / 2 & -4)
+ && CONST_OK_FOR_I (size - (size / 2 & -4)))
+ {
+ val = GEN_INT (size / 2 & -4);
+ emit_insn (gen_addsi3 (reg, reg, val));
+ val = GEN_INT (size - (size / 2 & -4));
+ }
+ else
+ {
+ rtx reg;
+
+ /* 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 (temp < 0)
+ abort ();
+ reg = gen_rtx (REG, SImode, temp);
+ emit_insn (gen_movsi (reg, val));
+ val = reg;
+ }
}
insn = gen_addsi3 (reg, reg, val);
int rn;
{
rtx x;
- x = emit_insn (gen_push (gen_rtx (REG, SImode, rn)));
+ if ((rn >= FIRST_FP_REG && rn <= LAST_FP_REG)
+ || rn == FPUL_REG)
+ x = emit_insn (gen_push_e (gen_rtx (REG, SFmode, rn)));
+ else
+ x = emit_insn (gen_push (gen_rtx (REG, SImode, rn)));
+
REG_NOTES (x) = gen_rtx (EXPR_LIST, REG_INC,
gen_rtx(REG, SImode, STACK_POINTER_REGNUM), 0);
}
int rn;
{
rtx x;
- x = emit_insn (gen_pop (gen_rtx (REG, SImode, rn)));
+ if ((rn >= FIRST_FP_REG && rn <= LAST_FP_REG)
+ || rn == FPUL_REG)
+ x = emit_insn (gen_pop_e (gen_rtx (REG, SFmode, rn)));
+ else
+ x = emit_insn (gen_pop (gen_rtx (REG, SImode, rn)));
+
REG_NOTES (x) = gen_rtx (EXPR_LIST, REG_INC,
gen_rtx(REG, SImode, STACK_POINTER_REGNUM), 0);
}
the number of bytes the insns take. */
static void
-push_regs (mask)
- int mask;
+push_regs (mask, mask2)
+ int mask, mask2;
{
int i;
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ for (i = 0; i < 32; i++)
if (mask & (1 << i))
push (i);
+ for (i = 32; i < FIRST_PSEUDO_REGISTER; i++)
+ if (mask2 & (1 << (i - 32)))
+ push (i);
}
/* Work out the registers which need to be saved, both as a mask and a
make sure that all the regs it clobbers are safe too. */
static int
-calc_live_regs (count_ptr)
+calc_live_regs (count_ptr, live_regs_mask2)
int *count_ptr;
+ int *live_regs_mask2;
{
int reg;
int live_regs_mask = 0;
int count = 0;
+ *live_regs_mask2 = 0;
for (reg = 0; reg < FIRST_PSEUDO_REGISTER; reg++)
{
if (pragma_interrupt && ! pragma_trapa)
{
/* Need to save all the regs ever live. */
if ((regs_ever_live[reg]
- || (call_used_regs[reg] && regs_ever_live[PR_REG]))
+ || (call_used_regs[reg]
+ && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG)
+ && regs_ever_live[PR_REG]))
&& reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM
+ && reg != RETURN_ADDRESS_POINTER_REGNUM
&& reg != T_REG && reg != GBR_REG)
{
- live_regs_mask |= 1 << reg;
+ if (reg >= 32)
+ *live_regs_mask2 |= 1 << (reg - 32);
+ else
+ live_regs_mask |= 1 << reg;
count++;
}
}
/* Only push those regs which are used and need to be saved. */
if (regs_ever_live[reg] && ! call_used_regs[reg])
{
- live_regs_mask |= (1 << reg);
+ if (reg >= 32)
+ *live_regs_mask2 |= 1 << (reg - 32);
+ else
+ live_regs_mask |= (1 << reg);
count++;
}
}
{
int live_regs_mask;
int d, i;
- live_regs_mask = calc_live_regs (&d);
+ int live_regs_mask2;
+ live_regs_mask = calc_live_regs (&d, &live_regs_mask2);
/* We have pretend args if we had an object sent partially in registers
and partially on the stack, e.g. a large structure. */
- output_stack_adjust (-current_function_pretend_args_size, stack_pointer_rtx);
+ output_stack_adjust (-current_function_pretend_args_size,
+ stack_pointer_rtx, 3);
extra_push = 0;
/* This is set by SETUP_VARARGS to indicate that this is a varargs
- routine. Clear it here so that the next function isn't affected. */
+ routine. Clear it here so that the next function isn't affected. */
if (current_function_anonymous_args)
{
current_function_anonymous_args = 0;
- /* Push arg regs as if they'd been provided by caller in stack. */
- for (i = 0; i < NPARM_REGS; i++)
- {
- int rn = NPARM_REGS + FIRST_PARM_REG - i - 1;
- if (i > (NPARM_REGS - current_function_args_info
- - current_function_varargs))
- break;
- push (rn);
- extra_push += 4;
- }
+ /* This is not used by the SH3E calling convention */
+ if (!TARGET_SH3E)
+ {
+ /* Push arg regs as if they'd been provided by caller in stack. */
+ for (i = 0; i < NPARM_REGS(SImode); i++)
+ {
+ int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1;
+ if (i > (NPARM_REGS(SImode)
+ - current_function_args_info.arg_count[(int) SH_ARG_INT]
+ - current_function_varargs))
+ break;
+ push (rn);
+ extra_push += 4;
+ }
+ }
}
- push_regs (live_regs_mask);
- output_stack_adjust (-get_frame_size (), stack_pointer_rtx);
+
+ push_regs (live_regs_mask, live_regs_mask2);
+
+ output_stack_adjust (-get_frame_size (), stack_pointer_rtx, 3);
if (frame_pointer_needed)
emit_insn (gen_movsi (frame_pointer_rtx, stack_pointer_rtx));
int live_regs_mask;
int d, i;
- live_regs_mask = calc_live_regs (&d);
+ int live_regs_mask2;
+ live_regs_mask = calc_live_regs (&d, &live_regs_mask2);
if (frame_pointer_needed)
{
to ensure that instruction scheduling won't move the stack pointer
adjust before instructions reading from the frame. This can fail
if there is an interrupt which then writes to the stack. */
- output_stack_adjust (get_frame_size (), frame_pointer_rtx);
+ output_stack_adjust (get_frame_size (), frame_pointer_rtx, 7);
emit_insn (gen_movsi (stack_pointer_rtx, frame_pointer_rtx));
}
else
- output_stack_adjust (get_frame_size (), stack_pointer_rtx);
+ output_stack_adjust (get_frame_size (), stack_pointer_rtx, 7);
/* Pop all the registers. */
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
{
int j = (FIRST_PSEUDO_REGISTER - 1) - i;
- if (live_regs_mask & (1 << j))
+ if (j < 32 && (live_regs_mask & (1 << j)))
+ pop (j);
+ else if (j >= 32 && (live_regs_mask2 & (1 << (j - 32))))
pop (j);
}
output_stack_adjust (extra_push + current_function_pretend_args_size,
- stack_pointer_rtx);
+ stack_pointer_rtx, 7);
}
/* Clear variables at function end. */
FILE *stream;
int size;
{
- pragma_interrupt = pragma_trapa = 0;
+ pragma_interrupt = pragma_trapa = pragma_nosave_low_regs = 0;
+}
+
+rtx
+sh_builtin_saveregs (arglist)
+ tree arglist;
+{
+ tree fntype = TREE_TYPE (current_function_decl);
+ /* First unnamed integer register. */
+ int first_intreg = current_function_args_info.arg_count[(int) SH_ARG_INT];
+ /* Number of integer registers we need to save. */
+ int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg);
+ /* First unnamed SFmode float reg */
+ int first_floatreg = current_function_args_info.arg_count[(int) SH_ARG_FLOAT];
+ /* Number of SFmode float regs to save. */
+ int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg);
+ int ptrsize = GET_MODE_SIZE (Pmode);
+ rtx valist, regbuf, fpregs;
+ int bufsize, regno;
+
+ /* Allocate block of memory for the regs. */
+ /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte?
+ Or can assign_stack_local accept a 0 SIZE argument? */
+ bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD);
+
+ regbuf = assign_stack_local (BLKmode, bufsize, 0);
+ MEM_IN_STRUCT_P (regbuf) = 1;
+
+ /* Save int args.
+ This is optimized to only save the regs that are necessary. Explicitly
+ named args need not be saved. */
+ if (n_intregs > 0)
+ move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg,
+ gen_rtx (MEM, BLKmode,
+ plus_constant (XEXP (regbuf, 0),
+ n_floatregs * UNITS_PER_WORD)),
+ n_intregs, n_intregs * UNITS_PER_WORD);
+
+ /* Save float args.
+ This is optimized to only save the regs that are necessary. Explicitly
+ named args need not be saved.
+ We explicitly build a pointer to the buffer because it halves the insn
+ count when not optimizing (otherwise the pointer is built for each reg
+ saved). */
+
+ fpregs = gen_reg_rtx (Pmode);
+ emit_move_insn (fpregs, XEXP (regbuf, 0));
+ for (regno = first_floatreg; regno < NPARM_REGS (SFmode); regno ++)
+ emit_move_insn (gen_rtx (MEM, SFmode,
+ plus_constant (fpregs,
+ GET_MODE_SIZE (SFmode)
+ * (regno - first_floatreg))),
+ gen_rtx (REG, SFmode,
+ BASE_ARG_REG (SFmode) + regno));
+
+ /* Return the address of the regbuf. */
+ return XEXP (regbuf, 0);
}
/* Define the offset between two registers, one to be eliminated, and
int total_saved_regs_space;
int total_auto_space = get_frame_size ();
- calc_live_regs (®s_saved);
+ int live_regs_mask, live_regs_mask2;
+ live_regs_mask = calc_live_regs (®s_saved, &live_regs_mask2);
+
total_saved_regs_space = (regs_saved) * 4;
if (from == ARG_POINTER_REGNUM && to == FRAME_POINTER_REGNUM)
if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
return 0;
+ if (from == RETURN_ADDRESS_POINTER_REGNUM
+ && (to == FRAME_POINTER_REGNUM || to == STACK_POINTER_REGNUM))
+ {
+ int i, n = 0;
+ for (i = PR_REG+1; i < 32; i++)
+ if (live_regs_mask & (1 << i))
+ n += 4;
+ for (i = 32; i < FIRST_PSEUDO_REGISTER; i++)
+ if (live_regs_mask2 & (1 << (i - 32)))
+ n += 4;
+ return n + total_auto_space;
+ }
+
abort ();
}
\f
compiler. */
int
-handle_pragma (file)
+handle_pragma (file, t)
FILE *file;
+ tree t;
{
- int c;
- char pbuf[200];
- int psize = 0;
+ int retval = 0;
+ register char *pname;
- c = getc (file);
- while (c == ' ' || c == '\t')
- c = getc (file);
+ if (TREE_CODE (t) != IDENTIFIER_NODE)
+ return 0;
- if (c == '\n' || c == EOF)
- return c;
+ pname = IDENTIFIER_POINTER (t);
+ if (strcmp (pname, "interrupt") == 0)
+ pragma_interrupt = retval = 1;
+ else if (strcmp (pname, "trapa") == 0)
+ pragma_interrupt = pragma_trapa = retval = 1;
+ else if (strcmp (pname, "nosave_low_regs") == 0)
+ pragma_nosave_low_regs = retval = 1;
- while (psize < sizeof (pbuf) - 1 && c != '\n')
- {
- pbuf[psize++] = c;
- if (psize == 9 && strncmp (pbuf, "interrupt", 9) == 0)
- {
- pragma_interrupt = 1;
- return ' ';
- }
- if (psize == 5 && strncmp (pbuf, "trapa", 5) == 0)
- {
- pragma_interrupt = pragma_trapa = 1;
- return ' ';
- }
- c = getc (file);
- }
- return c;
+ return retval;
}
\f
/* Predicates used by the templates. */
{
if (register_operand (op, mode))
{
+ int regno;
+
if (GET_CODE (op) == REG)
- return (REGNO (op) != T_REG
- && REGNO (op) != PR_REG
- && REGNO (op) != MACH_REG
- && REGNO (op) != MACL_REG);
- return 1;
+ regno = REGNO (op);
+ else if (GET_CODE (op) == SUBREG && GET_CODE (SUBREG_REG (op)) == REG)
+ regno = REGNO (SUBREG_REG (op));
+ else
+ return 1;
+
+ return (regno != T_REG && regno != PR_REG && regno != FPUL_REG
+ && regno != MACH_REG && regno != MACL_REG);
}
return 0;
}
return 0;
}
-\f
-/* Determine where to put an argument to a function.
- Value is zero to push the argument on the stack,
- or a hard register in which to store the argument.
-
- MODE is the argument's machine mode.
- TYPE is the data type of the argument (as a tree).
- This is null for libcalls where that information may
- not be available.
- CUM is a variable of type CUMULATIVE_ARGS which gives info about
- the preceding args and about the function being called.
- NAMED is nonzero if this argument is a named parameter
- (otherwise it is an extra parameter matching an ellipsis). */
-rtx
-sh_function_arg (cum, mode, type, named)
- CUMULATIVE_ARGS cum;
- enum machine_mode mode;
- tree type;
- int named;
+/* Nonzero if OP is a floating point value with value 0.0. */
+
+int
+fp_zero_operand (op)
+ rtx op;
{
- if (named)
- {
- int rr = (ROUND_REG (cum, mode));
+ REAL_VALUE_TYPE r;
- if (rr < NPARM_REGS)
- return ((type == 0 || ! TREE_ADDRESSABLE (type))
- ? gen_rtx (REG, mode, FIRST_PARM_REG + rr) : 0);
- }
- return 0;
+ if (GET_MODE (op) != SFmode)
+ return 0;
+
+ REAL_VALUE_FROM_CONST_DOUBLE (r, op);
+ return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r);
}
-/* For an arg passed partly in registers and partly in memory,
- this is the number of registers used.
- For args passed entirely in registers or entirely in memory, zero.
- Any arg that starts in the first 4 regs but won't entirely fit in them
- needs partial registers on the SH. */
+/* Nonzero if OP is a floating point value with value 1.0. */
int
-sh_function_arg_partial_nregs (cum, mode, type, named)
- CUMULATIVE_ARGS cum;
- enum machine_mode mode;
- tree type;
- int named;
+fp_one_operand (op)
+ rtx op;
{
- if (cum < NPARM_REGS)
- {
- if ((type == 0 || ! TREE_ADDRESSABLE (type))
- && (cum + (mode == BLKmode
- ? ROUND_ADVANCE (int_size_in_bytes (type))
- : ROUND_ADVANCE (GET_MODE_SIZE (mode))) - NPARM_REGS > 0))
- return NPARM_REGS - cum;
- }
- return 0;
+ REAL_VALUE_TYPE r;
+
+ if (GET_MODE (op) != SFmode)
+ return 0;
+
+ REAL_VALUE_FROM_CONST_DOUBLE (r, op);
+ return REAL_VALUES_EQUAL (r, dconst1);
}
\f
/* Return non-zero if REG is not used after INSN.
We assume REG is a reload reg, and therefore does
- not live past labels or calls or jumps. */
+ not live past labels. It may live past calls or jumps though. */
int
reg_unused_after (reg, insn)
rtx reg;
{
code = GET_CODE (insn);
+#if 0
+ /* If this is a label that existed before reload, then the register
+ if dead here. However, if this is a label added by reorg, then
+ the register may still be live here. We can't tell the difference,
+ so we just ignore labels completely. */
if (code == CODE_LABEL)
return 1;
+ /* else */
+#endif
+
+ if (code == JUMP_INSN)
+ return 0;
/* If this is a sequence, we must handle them all at once.
We could have for instance a call that sets the target register,
if (GET_CODE (this_insn) == CALL_INSN)
code = CALL_INSN;
+ else if (GET_CODE (this_insn) == JUMP_INSN)
+ {
+ if (INSN_ANNULLED_BRANCH_P (this_insn))
+ return 0;
+ code = JUMP_INSN;
+ }
if (set && reg_overlap_mentioned_p (reg, SET_SRC (set)))
return 0;
}
if (retval == 1)
return 1;
+ else if (code == JUMP_INSN)
+ return 0;
}
else if (GET_RTX_CLASS (code) == 'i')
{