/* Subroutines for insn-output.c for Intel X86.
- Copyright (C) 1988, 92, 94, 95, 96, 97, 1998 Free Software Foundation, Inc.
+ Copyright (C) 1988, 92, 94-98, 1999 Free Software Foundation, Inc.
This file is part of GNU CC.
17 /* cost of a divide/mod */
};
+struct processor_costs k6_cost = {
+ 1, /* cost of an add instruction */
+ 1, /* cost of a lea instruction */
+ 1, /* variable shift costs */
+ 1, /* constant shift costs */
+ 2, /* cost of starting a multiply */
+ 0, /* cost of multiply per each bit set */
+ 18 /* cost of a divide/mod */
+};
+
struct processor_costs *ix86_cost = &pentium_cost;
+/* Processor feature/optimization bitmasks. */
+#define m_386 (1<<PROCESSOR_I386)
+#define m_486 (1<<PROCESSOR_I486)
+#define m_PENT (1<<PROCESSOR_PENTIUM)
+#define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
+#define m_K6 (1<<PROCESSOR_K6)
+
+const int x86_use_leave = m_386 | m_K6;
+const int x86_push_memory = m_386 | m_K6;
+const int x86_zero_extend_with_and = m_486 | m_PENT;
+const int x86_movx = m_386 | m_PPRO | m_K6;
+const int x86_double_with_add = ~(m_386 | m_PENT | m_PPRO);
+const int x86_use_bit_test = m_386;
+const int x86_unroll_strlen = m_486 | m_PENT | m_PPRO;
+const int x86_use_q_reg = m_PENT | m_PPRO | m_K6;
+const int x86_use_any_reg = m_486;
+const int x86_cmove = m_PPRO;
+const int x86_deep_branch = m_PPRO| m_K6;
+
#define AT_BP(mode) (gen_rtx_MEM ((mode), frame_pointer_rtx))
extern FILE *asm_out_file;
SIREG, DIREG, INDEX_REGS, GENERAL_REGS,
/* FP registers */
FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS,
- FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
+ FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
/* arg pointer */
INDEX_REGS
};
{PROCESSOR_I686_STRING, PROCESSOR_PENTIUMPRO, &pentiumpro_cost,
0, 0},
{PROCESSOR_PENTIUMPRO_STRING, PROCESSOR_PENTIUMPRO,
- &pentiumpro_cost, 0, 0}};
+ &pentiumpro_cost, 0, 0},
+ {PROCESSOR_K6_STRING, PROCESSOR_K6, &k6_cost, 0, 0}};
int ptt_size = sizeof (processor_target_table) / sizeof (struct ptt);
for (i = 0; (ch = i386_reg_alloc_order[i]) != '\0'; i++)
{
int regno = 0;
-
+
switch (ch)
{
case 'a': regno = 0; break;
if (ix86_cpu_string == 0)
ix86_cpu_string = PROCESSOR_DEFAULT_STRING;
}
-
+
for (i = 0; i < ptt_size; i++)
if (! strcmp (ix86_arch_string, processor_target_table[i].name))
{
{
ix86_cpu = processor_target_table[j].processor;
ix86_cost = processor_target_table[j].cost;
- if (i > j && (int) ix86_arch >= (int) PROCESSOR_PENTIUMPRO)
+ if (i > j && (int) ix86_arch >= (int) PROCESSOR_K6)
error ("-mcpu=%s does not support -march=%s",
ix86_cpu_string, ix86_arch_string);
for (i = order = 0; (ch = i386_reg_alloc_order[i]) != '\0'; i++)
{
int regno = 0;
-
+
switch (ch)
{
case 'a': regno = 0; break;
case REG:
return i386_aligned_reg_p (REGNO (op));
-
+
default:
break;
}
tree args;
{
if (TREE_CODE (type) != FUNCTION_TYPE
+ && TREE_CODE (type) != METHOD_TYPE
&& TREE_CODE (type) != FIELD_DECL
&& TREE_CODE (type) != TYPE_DECL)
return 0;
int
i386_comp_type_attributes (type1, type2)
- tree type1 ATTRIBUTE_UNUSED;
- tree type2 ATTRIBUTE_UNUSED;
+ tree type1;
+ tree type2;
{
+ /* Check for mismatch of non-default calling convention. */
+ char *rtdstr = TARGET_RTD ? "cdecl" : "stdcall";
+
+ if (TREE_CODE (type1) != FUNCTION_TYPE)
+ return 1;
+
+ /* Check for mismatched return types (cdecl vs stdcall). */
+ if (!lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type1))
+ != !lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type2)))
+ return 0;
return 1;
}
tree fundecl;
tree funtype;
int size;
-{
+{
int rtd = TARGET_RTD && (!fundecl || TREE_CODE (fundecl) != IDENTIFIER_NODE);
/* Cdecl functions override -mrtd, and never pop the stack. */
if (! lookup_attribute ("cdecl", TYPE_ATTRIBUTES (funtype))) {
-
+
/* Stdcall functions will pop the stack if not variable args. */
if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (funtype)))
rtd = 1;
-
+
if (rtd
&& (TYPE_ARG_TYPES (funtype) == NULL_TREE
|| (TREE_VALUE (tree_last (TYPE_ARG_TYPES (funtype)))
== void_type_node)))
return size;
}
-
+
/* Lose any fake structure return argument. */
if (aggregate_value_p (TREE_TYPE (funtype)))
return GET_MODE_SIZE (Pmode);
-
+
return 0;
}
return 0;
}
\f
-/* Output an insn whose source is a 386 integer register. SRC is the
- rtx for the register, and TEMPLATE is the op-code template. SRC may
- be either SImode or DImode.
-
- The template will be output with operands[0] as SRC, and operands[1]
- as a pointer to the top of the 386 stack. So a call from floatsidf2
- would look like this:
-
- output_op_from_reg (operands[1], AS1 (fild%z0,%1));
-
- where %z0 corresponds to the caller's operands[1], and is used to
- emit the proper size suffix.
-
- ??? Extend this to handle HImode - a 387 can load and store HImode
- values directly. */
-
-void
-output_op_from_reg (src, template)
- rtx src;
- char *template;
-{
- rtx xops[4];
- int size = GET_MODE_SIZE (GET_MODE (src));
-
- xops[0] = src;
- xops[1] = AT_SP (Pmode);
- xops[2] = GEN_INT (size);
- xops[3] = stack_pointer_rtx;
-
- if (size > UNITS_PER_WORD)
- {
- rtx high;
-
- if (size > 2 * UNITS_PER_WORD)
- {
- high = gen_rtx_REG (SImode, REGNO (src) + 2);
- output_asm_insn (AS1 (push%L0,%0), &high);
- }
-
- high = gen_rtx_REG (SImode, REGNO (src) + 1);
- output_asm_insn (AS1 (push%L0,%0), &high);
- }
-
- output_asm_insn (AS1 (push%L0,%0), &src);
- output_asm_insn (template, xops);
- output_asm_insn (AS2 (add%L3,%2,%3), xops);
-}
-\f
/* Output an insn to pop an value from the 387 top-of-stack to 386
register DEST. The 387 register stack is popped if DIES is true. If
the mode of DEST is an integer mode, a `fist' integer store is done,
{
if (dies)
output_asm_insn (AS1 (fistp%z3,%y0), xops);
+ else if (GET_MODE (xops[3]) == DImode && ! dies)
+ {
+ /* There is no DImode version of this without a stack pop, so
+ we must emulate it. It doesn't matter much what the second
+ instruction is, because the value being pushed on the FP stack
+ is not used except for the following stack popping store.
+ This case can only happen without optimization, so it doesn't
+ matter that it is inefficient. */
+ output_asm_insn (AS1 (fistp%z3,%0), xops);
+ output_asm_insn (AS1 (fild%z3,%0), xops);
+ }
else
output_asm_insn (AS1 (fist%z3,%y0), xops);
}
output_asm_insn (AS1 (pop%L0,%0), &dest);
else
{
- xops[0] = adj_offsettable_operand (xops[0], 4);
+ xops[0] = adj_offsettable_operand (xops[0], 4);
xops[3] = dest;
output_asm_insn (AS2 (mov%L0,%0,%3), xops);
}
output_asm_insn (AS1 (pop%L0,%0), &dest);
else
{
- xops[0] = adj_offsettable_operand (xops[0], 4);
+ xops[0] = adj_offsettable_operand (xops[0], 4);
output_asm_insn (AS2 (mov%L0,%0,%3), xops);
}
}
middlehalf[0] = operands[0];
latehalf[0] = operands[0];
}
-
+
if (optype1 == REGOP)
{
middlehalf[1] = gen_rtx_REG (SImode, REGNO (operands[1]) + 1);
char *push;
rtx xops[2];
} tmp_info[MAX_TMPS];
-
+
rtx src = operands[1];
int max_tmps = 0;
int offset = 0;
return "";
}
\f
-/* Output the appropriate code to move data between two memory locations */
-
-char *
-output_move_memory (operands, insn, length, tmp_start, n_operands)
- rtx operands[];
- rtx insn;
- int length;
- int tmp_start;
- int n_operands;
-{
- struct
- {
- char *load;
- char *store;
- rtx xops[3];
- } tmp_info[MAX_TMPS];
-
- rtx dest = operands[0];
- rtx src = operands[1];
- rtx qi_tmp = NULL_RTX;
- int max_tmps = 0;
- int offset = 0;
- int i, num_tmps;
- rtx xops[3];
-
- if (GET_CODE (dest) == MEM
- && GET_CODE (XEXP (dest, 0)) == PRE_INC
- && XEXP (XEXP (dest, 0), 0) == stack_pointer_rtx)
- return output_move_pushmem (operands, insn, length, tmp_start, n_operands);
-
- if (! offsettable_memref_p (src))
- fatal_insn ("Source is not offsettable", insn);
-
- if (! offsettable_memref_p (dest))
- fatal_insn ("Destination is not offsettable", insn);
-
- /* Figure out which temporary registers we have available */
- for (i = tmp_start; i < n_operands; i++)
- {
- if (GET_CODE (operands[i]) == REG)
- {
- if ((length & 1) != 0 && qi_tmp == 0 && QI_REG_P (operands[i]))
- qi_tmp = operands[i];
-
- if (reg_overlap_mentioned_p (operands[i], dest))
- fatal_insn ("Temporary register overlaps the destination", insn);
-
- if (reg_overlap_mentioned_p (operands[i], src))
- fatal_insn ("Temporary register overlaps the source", insn);
-
- tmp_info[max_tmps++].xops[2] = operands[i];
- if (max_tmps == MAX_TMPS)
- break;
- }
- }
-
- if (max_tmps == 0)
- fatal_insn ("No scratch registers were found to do memory->memory moves",
- insn);
-
- if ((length & 1) != 0)
- {
- if (qi_tmp == 0)
- fatal_insn ("No byte register found when moving odd # of bytes.",
- insn);
- }
-
- while (length > 1)
- {
- for (num_tmps = 0; num_tmps < max_tmps; num_tmps++)
- {
- if (length >= 4)
- {
- tmp_info[num_tmps].load = AS2(mov%L0,%1,%2);
- tmp_info[num_tmps].store = AS2(mov%L0,%2,%0);
- tmp_info[num_tmps].xops[0]
- = adj_offsettable_operand (dest, offset);
- tmp_info[num_tmps].xops[1]
- = adj_offsettable_operand (src, offset);
-
- offset += 4;
- length -= 4;
- }
-
- else if (length >= 2)
- {
- tmp_info[num_tmps].load = AS2(mov%W0,%1,%2);
- tmp_info[num_tmps].store = AS2(mov%W0,%2,%0);
- tmp_info[num_tmps].xops[0]
- = adj_offsettable_operand (dest, offset);
- tmp_info[num_tmps].xops[1]
- = adj_offsettable_operand (src, offset);
-
- offset += 2;
- length -= 2;
- }
- else
- break;
- }
-
- for (i = 0; i < num_tmps; i++)
- output_asm_insn (tmp_info[i].load, tmp_info[i].xops);
-
- for (i = 0; i < num_tmps; i++)
- output_asm_insn (tmp_info[i].store, tmp_info[i].xops);
- }
-
- if (length == 1)
- {
- xops[0] = adj_offsettable_operand (dest, offset);
- xops[1] = adj_offsettable_operand (src, offset);
- xops[2] = qi_tmp;
- output_asm_insn (AS2(mov%B0,%1,%2), xops);
- output_asm_insn (AS2(mov%B0,%2,%0), xops);
- }
-
- return "";
-}
-\f
int
standard_80387_constant_p (x)
rtx x;
/* Note that on the 80387, other constants, such as pi,
are much slower to load as standard constants
than to load from doubles in memory! */
+ /* ??? Not true on K6: all constants are equal cost. */
#endif
return 0;
}
}
+/* Return nonzero if OP is a constant shift count small enough to
+ encode into an lea instruction. */
+
+int
+small_shift_operand (op, mode)
+ rtx op;
+ enum machine_mode mode ATTRIBUTE_UNUSED;
+{
+ return (GET_CODE (op) == CONST_INT && INTVAL (op) > 0 && INTVAL (op) < 4);
+}
+
/* Test for a valid operand for a call instruction.
Don't allow the arg pointer register or virtual regs
since they may change into reg + const, which the patterns
emit_move_insn (temp, operands[1]);
operands[1] = temp;
return TRUE;
- }
+ }
}
if (!ix86_binary_operator_ok (code, mode, operands))
emit_move_insn (temp, operands[1]);
operands[1] = temp;
return TRUE;
- }
+ }
if (modified && ! ix86_binary_operator_ok (code, mode, operands))
return FALSE;
prologue_node = make_node (FUNCTION_DECL);
DECL_RESULT (prologue_node) = 0;
-#ifdef ASM_DECLARE_FUNCTION_NAME
- ASM_DECLARE_FUNCTION_NAME (file, pic_label_name, prologue_node);
-#endif
+
+ /* This used to call ASM_DECLARE_FUNCTION_NAME() but since it's an
+ internal (non-global) label that's being emitted, it didn't make
+ sense to have .type information for local labels. This caused
+ the SCO OpenServer 5.0.4 ELF assembler grief (why are you giving
+ me debug info for a label that you're declaring non-global?) this
+ was changed to call ASM_OUTPUT_LABEL() instead. */
+
+
+ ASM_OUTPUT_LABEL (file, pic_label_name);
output_asm_insn ("movl (%1),%0", xops);
output_asm_insn ("ret", xops);
}
pic_label_rtx = 0;
return;
}
-
+
ix86_prologue (0);
}
{
if (! TARGET_SCHEDULE_PROLOGUE)
return;
-
+
ix86_prologue (1);
}
if (do_rtl)
{
emit_insn (gen_prologue_get_pc (xops[0], xops[1]));
- emit_insn (gen_prologue_set_got (xops[0],
+ emit_insn (gen_prologue_set_got (xops[0],
gen_rtx (SYMBOL_REF, Pmode,
- "$_GLOBAL_OFFSET_TABLE_"),
+ "$_GLOBAL_OFFSET_TABLE_"),
xops[1]));
}
else
{
xops[0] = pic_offset_table_rtx;
xops[1] = gen_label_rtx ();
-
+
if (do_rtl)
{
/* We can't put a raw CODE_LABEL into the RTL, and we can't emit
else
{
output_asm_insn (AS1 (call,%P1), xops);
- ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L",
+ ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L",
CODE_LABEL_NUMBER (xops[1]));
output_asm_insn (AS1 (pop%L0,%0), xops);
output_asm_insn ("addl $_GLOBAL_OFFSET_TABLE_+[.-%P1],%0", xops);
}
- }
+ }
/* When -fpic, we must emit a scheduling barrier, so that the instruction
that restores %ebx (which is PIC_OFFSET_TABLE_REGNUM), does not get
emit_insn (gen_blockage ());
}
+/* Compute the size of local storage taking into consideration the
+ desired stack alignment which is to be maintained. Also determine
+ the number of registers saved below the local storage. */
+
+HOST_WIDE_INT
+ix86_compute_frame_size (size, nregs_on_stack)
+ HOST_WIDE_INT size;
+ int *nregs_on_stack;
+{
+ int limit;
+ int nregs;
+ int regno;
+ int padding;
+ int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table
+ || current_function_uses_const_pool);
+ HOST_WIDE_INT total_size;
+
+ limit = frame_pointer_needed
+ ? FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM;
+
+ nregs = 0;
+
+ for (regno = limit - 1; regno >= 0; regno--)
+ if ((regs_ever_live[regno] && ! call_used_regs[regno])
+ || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used))
+ nregs++;
+
+ padding = 0;
+ total_size = size + (nregs * UNITS_PER_WORD);
+
+#ifdef PREFERRED_STACK_BOUNDARY
+ {
+ int offset;
+ int preferred_alignment = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT;
+
+ offset = 4;
+ if (frame_pointer_needed)
+ offset += UNITS_PER_WORD;
+
+ total_size += offset;
+
+ padding = ((total_size + preferred_alignment - 1)
+ & -preferred_alignment) - total_size;
+
+ if (padding < (((offset + preferred_alignment - 1)
+ & -preferred_alignment) - offset))
+ padding += preferred_alignment;
+ }
+#endif
+
+ if (nregs_on_stack)
+ *nregs_on_stack = nregs;
+
+ return size + padding;
+}
+
static void
ix86_prologue (do_rtl)
int do_rtl;
rtx xops[4];
int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table
|| current_function_uses_const_pool);
- long tsize = get_frame_size ();
+ HOST_WIDE_INT tsize = ix86_compute_frame_size (get_frame_size (), (int *)0);
rtx insn;
int cfa_offset = INCOMING_FRAME_SP_OFFSET, cfa_store_offset = cfa_offset;
-
+
xops[0] = stack_pointer_rtx;
xops[1] = frame_pointer_rtx;
xops[2] = GEN_INT (tsize);
else
{
- output_asm_insn ("push%L1 %1", xops);
+ output_asm_insn ("push%L1 %1", xops);
#ifdef INCOMING_RETURN_ADDR_RTX
if (dwarf2out_do_frame ())
{
}
#endif
- output_asm_insn (AS2 (mov%L0,%0,%1), xops);
+ output_asm_insn (AS2 (mov%L0,%0,%1), xops);
#ifdef INCOMING_RETURN_ADDR_RTX
if (dwarf2out_do_frame ())
dwarf2out_def_cfa ("", FRAME_POINTER_REGNUM, cfa_offset);
insn = emit_insn (gen_prologue_set_stack_ptr (xops[2]));
RTX_FRAME_RELATED_P (insn) = 1;
}
- else
+ else
{
output_asm_insn (AS2 (sub%L0,%2,%0), xops);
#ifdef INCOMING_RETURN_ADDR_RTX
#endif
}
}
- else
+ else
{
xops[3] = gen_rtx_REG (SImode, 0);
if (do_rtl)
}
}
+#ifdef SUBTARGET_PROLOGUE
+ SUBTARGET_PROLOGUE;
+#endif
+
if (pic_reg_used)
load_pic_register (do_rtl);
/* Return 1 if it is appropriate to emit `ret' instructions in the
body of a function. Do this only if the epilogue is simple, needing a
couple of insns. Prior to reloading, we can't tell how many registers
- must be saved, so return 0 then. Return 0 if there is no frame
+ must be saved, so return 0 then. Return 0 if there is no frame
marker to de-allocate.
If NON_SAVING_SETJMP is defined and true, then it is not possible
return;
}
-/* Restore function stack, frame, and registers. */
+/* Restore function stack, frame, and registers. */
void
ix86_expand_epilogue ()
int do_rtl;
{
register int regno;
- register int nregs, limit;
- int offset;
+ register int limit;
+ int nregs;
rtx xops[3];
int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table
|| current_function_uses_const_pool);
- long tsize = get_frame_size ();
-
- /* Compute the number of registers to pop */
-
- limit = (frame_pointer_needed ? FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM);
-
- nregs = 0;
-
- for (regno = limit - 1; regno >= 0; regno--)
- if ((regs_ever_live[regno] && ! call_used_regs[regno])
- || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used))
- nregs++;
+ int sp_valid = !frame_pointer_needed || current_function_sp_is_unchanging;
+ HOST_WIDE_INT offset;
+ HOST_WIDE_INT tsize = ix86_compute_frame_size (get_frame_size (), &nregs);
- /* sp is often unreliable so we must go off the frame pointer.
+ /* sp is often unreliable so we may have to go off the frame pointer. */
- In reality, we may not care if sp is unreliable, because we can restore
- the register relative to the frame pointer. In theory, since each move
- is the same speed as a pop, and we don't need the leal, this is faster.
- For now restore multiple registers the old way. */
-
- offset = - tsize - (nregs * UNITS_PER_WORD);
+ offset = -(tsize + nregs * UNITS_PER_WORD);
xops[2] = stack_pointer_rtx;
if (flag_pic || profile_flag || profile_block_flag)
emit_insn (gen_blockage ());
- if (nregs > 1 || ! frame_pointer_needed)
+ /* If we're only restoring one register and sp is not valid then
+ using a move instruction to restore the register since it's
+ less work than reloading sp and popping the register. Otherwise,
+ restore sp (if necessary) and pop the registers. */
+
+ limit = frame_pointer_needed
+ ? FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM;
+
+ if (nregs > 1 || sp_valid)
{
- if (frame_pointer_needed)
+ if ( !sp_valid )
{
xops[0] = adj_offsettable_operand (AT_BP (QImode), offset);
if (do_rtl)
else if (tsize)
{
- /* If there is no frame pointer, we must still release the frame. */
- xops[0] = GEN_INT (tsize);
+ /* Intel's docs say that for 4 or 8 bytes of stack frame one should
+ use `pop' and not `add'. */
+ int use_pop = tsize == 4;
- if (do_rtl)
- emit_insn (gen_rtx (SET, VOIDmode, xops[2],
- gen_rtx (PLUS, SImode, xops[2], xops[0])));
+ /* Use two pops only for the Pentium processors. */
+ if (tsize == 8 && !TARGET_386 && !TARGET_486)
+ {
+ rtx retval = current_function_return_rtx;
+
+ xops[1] = gen_rtx_REG (SImode, 1); /* %edx */
+
+ /* This case is a bit more complex. Since we cannot pop into
+ %ecx twice we need a second register. But this is only
+ available if the return value is not of DImode in which
+ case the %edx register is not available. */
+ use_pop = (retval == NULL
+ || ! reg_overlap_mentioned_p (xops[1], retval));
+ }
+
+ if (use_pop)
+ {
+ xops[0] = gen_rtx_REG (SImode, 2); /* %ecx */
+
+ if (do_rtl)
+ {
+ /* We have to prevent the two pops here from being scheduled.
+ GCC otherwise would try in some situation to put other
+ instructions in between them which has a bad effect. */
+ emit_insn (gen_blockage ());
+ emit_insn (gen_pop (xops[0]));
+ if (tsize == 8)
+ emit_insn (gen_pop (xops[1]));
+ }
+ else
+ {
+ output_asm_insn ("pop%L0 %0", xops);
+ if (tsize == 8)
+ output_asm_insn ("pop%L1 %1", xops);
+ }
+ }
else
- output_asm_insn (AS2 (add%L2,%0,%2), xops);
+ {
+ /* If there is no frame pointer, we must still release the frame. */
+ xops[0] = GEN_INT (tsize);
+
+ if (do_rtl)
+ emit_insn (gen_rtx (SET, VOIDmode, xops[2],
+ gen_rtx (PLUS, SImode, xops[2], xops[0])));
+ else
+ output_asm_insn (AS2 (add%L2,%0,%2), xops);
+ }
}
#ifdef FUNCTION_BLOCK_PROFILER_EXIT
output_asm_insn ("jmp %*%0", xops);
}
}
- else
+ else
{
if (do_rtl)
emit_jump_insn (gen_return_pop_internal (xops[1]));
} while (0)
int
+legitimate_pic_address_disp_p (disp)
+ register rtx disp;
+{
+ if (GET_CODE (disp) != CONST)
+ return 0;
+ disp = XEXP (disp, 0);
+
+ if (GET_CODE (disp) == PLUS)
+ {
+ if (GET_CODE (XEXP (disp, 1)) != CONST_INT)
+ return 0;
+ disp = XEXP (disp, 0);
+ }
+
+ if (GET_CODE (disp) != UNSPEC
+ || XVECLEN (disp, 0) != 1)
+ return 0;
+
+ /* Must be @GOT or @GOTOFF. */
+ if (XINT (disp, 1) != 6
+ && XINT (disp, 1) != 7)
+ return 0;
+
+ if (GET_CODE (XVECEXP (disp, 0, 0)) != SYMBOL_REF
+ && GET_CODE (XVECEXP (disp, 0, 0)) != LABEL_REF)
+ return 0;
+
+ return 1;
+}
+
+int
legitimate_address_p (mode, addr, strict)
enum machine_mode mode;
register rtx addr;
}
if (GET_CODE (addr) == REG || GET_CODE (addr) == SUBREG)
- base = addr;
+ base = addr;
else if (GET_CODE (addr) == PLUS)
{
return FALSE;
}
+ if (GET_MODE (base) != Pmode)
+ {
+ ADDR_INVALID ("Base is not in Pmode.\n", base);
+ return FALSE;
+ }
+
if ((strict && ! REG_OK_FOR_BASE_STRICT_P (base))
|| (! strict && ! REG_OK_FOR_BASE_NONSTRICT_P (base)))
{
return FALSE;
}
+ if (GET_MODE (indx) != Pmode)
+ {
+ ADDR_INVALID ("Index is not in Pmode.\n", indx);
+ return FALSE;
+ }
+
if ((strict && ! REG_OK_FOR_INDEX_STRICT_P (indx))
|| (! strict && ! REG_OK_FOR_INDEX_NONSTRICT_P (indx)))
{
}
}
- /* Validate displacement
- Constant pool addresses must be handled special. They are
- considered legitimate addresses, but only if not used with regs.
- When printed, the output routines know to print the reference with the
- PIC reg, even though the PIC reg doesn't appear in the RTL. */
+ /* Validate displacement. */
if (disp)
{
- if (GET_CODE (disp) == SYMBOL_REF
- && CONSTANT_POOL_ADDRESS_P (disp)
- && base == 0
- && indx == 0)
- ;
-
- else if (!CONSTANT_ADDRESS_P (disp))
+ if (!CONSTANT_ADDRESS_P (disp))
{
ADDR_INVALID ("Displacement is not valid.\n", disp);
return FALSE;
return FALSE;
}
- else if (flag_pic && SYMBOLIC_CONST (disp)
- && base != pic_offset_table_rtx
- && (indx != pic_offset_table_rtx || scale != NULL_RTX))
+ if (flag_pic && SYMBOLIC_CONST (disp))
{
- ADDR_INVALID ("Displacement is an invalid pic reference.\n", disp);
- return FALSE;
+ if (! legitimate_pic_address_disp_p (disp))
+ {
+ ADDR_INVALID ("Displacement is an invalid PIC construct.\n",
+ disp);
+ return FALSE;
+ }
+
+ if (base != pic_offset_table_rtx
+ && (indx != pic_offset_table_rtx || scale != NULL_RTX))
+ {
+ ADDR_INVALID ("PIC displacement against invalid base.\n", disp);
+ return FALSE;
+ }
}
- else if (HALF_PIC_P () && HALF_PIC_ADDRESS_P (disp)
- && (base != NULL_RTX || indx != NULL_RTX))
+ else if (HALF_PIC_P ())
{
- ADDR_INVALID ("Displacement is an invalid half-pic reference.\n",
- disp);
- return FALSE;
+ if (! HALF_PIC_ADDRESS_P (disp)
+ || (base != NULL_RTX || indx != NULL_RTX))
+ {
+ ADDR_INVALID ("Displacement is an invalid half-pic reference.\n",
+ disp);
+ return FALSE;
+ }
}
}
/* Return a legitimate reference for ORIG (an address) using the
register REG. If REG is 0, a new pseudo is generated.
- There are three types of references that must be handled:
+ There are two types of references that must be handled:
1. Global data references must load the address from the GOT, via
the PIC reg. An insn is emitted to do this load, and the reg is
returned.
- 2. Static data references must compute the address as an offset
- from the GOT, whose base is in the PIC reg. An insn is emitted to
- compute the address into a reg, and the reg is returned. Static
- data objects have SYMBOL_REF_FLAG set to differentiate them from
- global data objects.
-
- 3. Constant pool addresses must be handled special. They are
- considered legitimate addresses, but only if not used with regs.
- When printed, the output routines know to print the reference with the
- PIC reg, even though the PIC reg doesn't appear in the RTL.
+ 2. Static data references, constant pool addresses, and code labels
+ compute the address as an offset from the GOT, whose base is in
+ the PIC reg. Static data objects have SYMBOL_REF_FLAG set to
+ differentiate them from global data objects. The returned
+ address is the PIC reg + an unspec constant.
GO_IF_LEGITIMATE_ADDRESS rejects symbolic references unless the PIC
- reg also appears in the address (except for constant pool references,
- noted above).
-
- "switch" statements also require special handling when generating
- PIC code. See comments by the `casesi' insn in i386.md for details. */
+ reg also appears in the address. */
rtx
legitimize_pic_address (orig, reg)
{
rtx addr = orig;
rtx new = orig;
+ rtx base;
- if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
+ if (GET_CODE (addr) == LABEL_REF
+ || (GET_CODE (addr) == SYMBOL_REF
+ && (CONSTANT_POOL_ADDRESS_P (addr)
+ || SYMBOL_REF_FLAG (addr))))
{
- if (GET_CODE (addr) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (addr))
- reg = new = orig;
- else
- {
- if (reg == 0)
- reg = gen_reg_rtx (Pmode);
+ /* This symbol may be referenced via a displacement from the PIC
+ base address (@GOTOFF). */
- if ((GET_CODE (addr) == SYMBOL_REF && SYMBOL_REF_FLAG (addr))
- || GET_CODE (addr) == LABEL_REF)
- new = gen_rtx (PLUS, Pmode, pic_offset_table_rtx, orig);
- else
- new = gen_rtx_MEM (Pmode,
- gen_rtx (PLUS, Pmode, pic_offset_table_rtx, orig));
+ current_function_uses_pic_offset_table = 1;
+ new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, addr), 7);
+ new = gen_rtx_CONST (VOIDmode, new);
+ new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new);
+ if (reg != 0)
+ {
emit_move_insn (reg, new);
+ new = reg;
}
- current_function_uses_pic_offset_table = 1;
- return reg;
}
-
- else if (GET_CODE (addr) == CONST || GET_CODE (addr) == PLUS)
+ else if (GET_CODE (addr) == SYMBOL_REF)
{
- rtx base;
+ /* This symbol must be referenced via a load from the
+ Global Offset Table (@GOT). */
- if (GET_CODE (addr) == CONST)
- {
- addr = XEXP (addr, 0);
- if (GET_CODE (addr) != PLUS)
- abort ();
- }
-
- if (XEXP (addr, 0) == pic_offset_table_rtx)
- return orig;
+ current_function_uses_pic_offset_table = 1;
+ new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, addr), 6);
+ new = gen_rtx_CONST (VOIDmode, new);
+ new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new);
+ new = gen_rtx_MEM (Pmode, new);
+ RTX_UNCHANGING_P (new) = 1;
if (reg == 0)
reg = gen_reg_rtx (Pmode);
-
- base = legitimize_pic_address (XEXP (addr, 0), reg);
- addr = legitimize_pic_address (XEXP (addr, 1),
- base == reg ? NULL_RTX : reg);
-
- if (GET_CODE (addr) == CONST_INT)
- return plus_constant (base, INTVAL (addr));
-
- if (GET_CODE (addr) == PLUS && CONSTANT_P (XEXP (addr, 1)))
+ emit_move_insn (reg, new);
+ new = reg;
+ }
+ else
+ {
+ if (GET_CODE (addr) == CONST)
{
- base = gen_rtx (PLUS, Pmode, base, XEXP (addr, 0));
- addr = XEXP (addr, 1);
+ addr = XEXP (addr, 0);
+ if (GET_CODE (addr) == UNSPEC)
+ {
+ /* Check that the unspec is one of the ones we generate? */
+ }
+ else if (GET_CODE (addr) != PLUS)
+ abort();
}
+ if (GET_CODE (addr) == PLUS)
+ {
+ rtx op0 = XEXP (addr, 0), op1 = XEXP (addr, 1);
+
+ /* Check first to see if this is a constant offset from a @GOTOFF
+ symbol reference. */
+ if ((GET_CODE (op0) == LABEL_REF
+ || (GET_CODE (op0) == SYMBOL_REF
+ && (CONSTANT_POOL_ADDRESS_P (op0)
+ || SYMBOL_REF_FLAG (op0))))
+ && GET_CODE (op1) == CONST_INT)
+ {
+ current_function_uses_pic_offset_table = 1;
+ new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, op0), 7);
+ new = gen_rtx_PLUS (VOIDmode, new, op1);
+ new = gen_rtx_CONST (VOIDmode, new);
+ new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new);
+
+ if (reg != 0)
+ {
+ emit_move_insn (reg, new);
+ new = reg;
+ }
+ }
+ else
+ {
+ base = legitimize_pic_address (XEXP (addr, 0), reg);
+ new = legitimize_pic_address (XEXP (addr, 1),
+ base == reg ? NULL_RTX : reg);
- return gen_rtx (PLUS, Pmode, base, addr);
+ if (GET_CODE (new) == CONST_INT)
+ new = plus_constant (base, INTVAL (new));
+ else
+ {
+ if (GET_CODE (new) == PLUS && CONSTANT_P (XEXP (new, 1)))
+ {
+ base = gen_rtx_PLUS (Pmode, base, XEXP (new, 0));
+ new = XEXP (new, 1);
+ }
+ new = gen_rtx_PLUS (Pmode, base, new);
+ }
+ }
+ }
}
return new;
}
rtx temp = reload_in_progress ? operands[0] : gen_reg_rtx (Pmode);
if (GET_CODE (operands[0]) == MEM && SYMBOLIC_CONST (operands[1]))
- operands[1] = force_reg (SImode, operands[1]);
+ operands[1] = force_reg (Pmode, operands[1]);
else
operands[1] = legitimize_pic_address (operands[1], temp);
}
&& (log = (unsigned)exact_log2 (INTVAL (XEXP (x, 1)))) < 4)
{
changed = 1;
- x = gen_rtx (MULT, Pmode, force_reg (Pmode, XEXP (x, 0)),
- GEN_INT (1 << log));
+ x = gen_rtx_MULT (Pmode, force_reg (Pmode, XEXP (x, 0)),
+ GEN_INT (1 << log));
}
if (GET_CODE (x) == PLUS)
break;
case SYMBOL_REF:
- case LABEL_REF:
- if (GET_CODE (x) == SYMBOL_REF)
- assemble_name (file, XSTR (x, 0));
- else
- {
- ASM_GENERATE_INTERNAL_LABEL (buf, "L",
- CODE_LABEL_NUMBER (XEXP (x, 0)));
- assemble_name (asm_out_file, buf);
- }
-
- if (code == 'X')
- ; /* No suffix, dammit. */
- else if (GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x))
- fprintf (file, "@GOTOFF(%%ebx)");
- else if (code == 'P')
- fprintf (file, "@PLT");
- else if (GET_CODE (x) == LABEL_REF)
- fprintf (file, "@GOTOFF");
- else if (! SYMBOL_REF_FLAG (x))
- fprintf (file, "@GOT");
- else
- fprintf (file, "@GOTOFF");
-
+ assemble_name (file, XSTR (x, 0));
+ if (code == 'P' && ! SYMBOL_REF_FLAG (x))
+ fputs ("@PLT", file);
break;
+ case LABEL_REF:
+ x = XEXP (x, 0);
+ /* FALLTHRU */
case CODE_LABEL:
ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x));
assemble_name (asm_out_file, buf);
if (GET_CODE (XEXP (x, 0)) == CONST_INT)
{
output_pic_addr_const (file, XEXP (x, 0), code);
- if (INTVAL (XEXP (x, 1)) >= 0)
- fprintf (file, "+");
+ fprintf (file, "+");
output_pic_addr_const (file, XEXP (x, 1), code);
}
- else
+ else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
{
output_pic_addr_const (file, XEXP (x, 1), code);
- if (INTVAL (XEXP (x, 0)) >= 0)
- fprintf (file, "+");
+ fprintf (file, "+");
output_pic_addr_const (file, XEXP (x, 0), code);
}
+ else
+ abort ();
break;
case MINUS:
output_pic_addr_const (file, XEXP (x, 1), code);
break;
+ case UNSPEC:
+ if (XVECLEN (x, 0) != 1)
+ abort ();
+ output_pic_addr_const (file, XVECEXP (x, 0, 0), code);
+ switch (XINT (x, 1))
+ {
+ case 6:
+ fputs ("@GOT", file);
+ break;
+ case 7:
+ fputs ("@GOTOFF", file);
+ break;
+ case 8:
+ fputs ("@PLT", file);
+ break;
+ default:
+ output_operand_lossage ("invalid UNSPEC as operand");
+ break;
+ }
+ break;
+
default:
output_operand_lossage ("invalid expression as operand");
}
}
\f
+static void
+put_jump_code (code, reverse, file)
+ enum rtx_code code;
+ int reverse;
+ FILE *file;
+{
+ int flags = cc_prev_status.flags;
+ int ieee = (TARGET_IEEE_FP && (flags & CC_IN_80387));
+ const char *suffix;
+
+ if (flags & CC_Z_IN_NOT_C)
+ switch (code)
+ {
+ case EQ:
+ fputs (reverse ? "c" : "nc", file);
+ return;
+
+ case NE:
+ fputs (reverse ? "nc" : "c", file);
+ return;
+
+ default:
+ abort ();
+ }
+ if (ieee)
+ {
+ switch (code)
+ {
+ case LE:
+ suffix = reverse ? "ae" : "b";
+ break;
+ case GT:
+ case LT:
+ case GE:
+ suffix = reverse ? "ne" : "e";
+ break;
+ case EQ:
+ suffix = reverse ? "ne" : "e";
+ break;
+ case NE:
+ suffix = reverse ? "e" : "ne";
+ break;
+ default:
+ abort ();
+ }
+ fputs (suffix, file);
+ return;
+ }
+ if (flags & CC_TEST_AX)
+ abort();
+ if ((flags & CC_NO_OVERFLOW) && (code == LE || code == GT))
+ abort ();
+ if (reverse)
+ code = reverse_condition (code);
+ switch (code)
+ {
+ case EQ:
+ suffix = "e";
+ break;
+
+ case NE:
+ suffix = "ne";
+ break;
+
+ case GT:
+ suffix = flags & CC_IN_80387 ? "a" : "g";
+ break;
+
+ case GTU:
+ suffix = "a";
+ break;
+
+ case LT:
+ if (flags & CC_NO_OVERFLOW)
+ suffix = "s";
+ else
+ suffix = flags & CC_IN_80387 ? "b" : "l";
+ break;
+
+ case LTU:
+ suffix = "b";
+ break;
+
+ case GE:
+ if (flags & CC_NO_OVERFLOW)
+ suffix = "ns";
+ else
+ suffix = flags & CC_IN_80387 ? "ae" : "ge";
+ break;
+
+ case GEU:
+ suffix = "ae";
+ break;
+
+ case LE:
+ suffix = flags & CC_IN_80387 ? "be" : "le";
+ break;
+
+ case LEU:
+ suffix = "be";
+ break;
+
+ default:
+ abort ();
+ }
+ fputs (suffix, file);
+}
+
/* Append the correct conditional move suffix which corresponds to CODE. */
static void
if (mode == MODE_INT)
switch (code)
{
- case NE:
+ case NE:
if (cc_prev_status.flags & CC_Z_IN_NOT_C)
fputs ("b", file);
else
else if (mode == MODE_FLOAT)
switch (code)
{
- case NE:
+ case NE:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "ne", file);
return;
- case EQ:
+ case EQ:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "e", file);
return;
- case GE:
+ case GE:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "nb", file);
return;
- case GT:
+ case GT:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "nbe", file);
return;
- case LE:
+ case LE:
fputs (ieee ? (reverse_cc ? "nb" : "b") : "be", file);
return;
- case LT:
+ case LT:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "b", file);
return;
- case GEU:
+ case GEU:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "nb", file);
return;
- case GTU:
+ case GTU:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "nbe", file);
return;
- case LEU:
+ case LEU:
fputs (ieee ? (reverse_cc ? "nb" : "b") : "be", file);
return;
- case LTU:
+ case LTU:
fputs (ieee ? (reverse_cc ? "ne" : "e") : "b", file);
return;
default:
C -- print opcode suffix for set/cmov insn.
c -- like C, but print reversed condition
F -- print opcode suffix for fcmov insn.
- f -- like C, but print reversed condition
+ f -- like F, but print reversed condition
+ D -- print the opcode suffix for a jump
+ d -- like D, but print reversed condition
R -- print the prefix for register names.
z -- print the opcode suffix for the size of the current operand.
* -- print a star (in certain assembler syntax)
w -- print the operand as if it's a "word" (HImode) even if it isn't.
- c -- don't print special prefixes before constant operands.
J -- print the appropriate jump operand.
s -- print a shift double count, followed by the assemblers argument
delimiter.
case GTU: fputs ("jne", file); return;
case LEU: fputs ("je", file); return;
case LTU: fputs ("#branch never", file); return;
-
+
/* no matching branches for GT nor LE */
-
+
default:
abort ();
}
return;
+ case 'D':
+ put_jump_code (GET_CODE (x), 0, file);
+ return;
+
+ case 'd':
+ put_jump_code (GET_CODE (x), 1, file);
+ return;
+
/* This is used by the conditional move instructions. */
case 'C':
put_condition_code (GET_CODE (x), 0, MODE_INT, file);
REAL_VALUE_TO_DECIMAL (r, "%.22e", dstr);
fprintf (file, "%s", dstr);
}
- else
+ else
{
if (code != 'P')
{
switch (GET_CODE (addr))
{
case REG:
+ /* ESI addressing makes instruction vector decoded on the K6. We can
+ avoid this by ESI+0 addressing. */
+ if (REGNO_REG_CLASS (REGNO (addr)) == SIREG
+ && ix86_cpu == PROCESSOR_K6 && !optimize_size)
+ output_addr_const (file, const0_rtx);
ADDR_BEG (file);
fprintf (file, "%se", RP);
fputs (hi_reg_name[REGNO (addr)], file);
ireg = XEXP (addr, 0);
}
- output_addr_const (file, const0_rtx);
- PRINT_B_I_S (NULL_RTX, ireg, scale, file);
+ /* (reg,reg,) is shorter than (,reg,2). */
+ if (scale == 2)
+ {
+ PRINT_B_I_S (ireg, ireg, 1, file);
+ }
+ else
+ {
+ output_addr_const (file, const0_rtx);
+ PRINT_B_I_S (NULL_RTX, ireg, scale, file);
+ }
}
break;
if (REG_P (SET_DEST (exp))
&& (REG_P (SET_SRC (exp)) || GET_CODE (SET_SRC (exp)) == MEM
|| GET_RTX_CLASS (GET_CODE (SET_SRC (exp))) == '<'
- || (GET_CODE (SET_SRC (exp)) == IF_THEN_ELSE
- && GET_MODE_CLASS (GET_MODE (SET_DEST (exp))) == MODE_INT)))
+ || GET_CODE (SET_SRC (exp)) == IF_THEN_ELSE))
{
if (cc_status.value1
&& reg_overlap_mentioned_p (SET_DEST (exp), cc_status.value1))
while (num--)
{
rtx op = operands[num];
- if (GET_CODE (op) == REG)
+ if (! reload_completed)
+ {
+ lo_half[num] = gen_lowpart (SImode, op);
+ hi_half[num] = gen_highpart (SImode, op);
+ }
+ else if (GET_CODE (op) == REG)
{
lo_half[num] = gen_rtx_REG (SImode, REGNO (op));
hi_half[num] = gen_rtx_REG (SImode, REGNO (op) + 1);
switch (GET_CODE (operands[3]))
{
case PLUS:
- if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
- || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
- base_op = "fiadd";
- else
- base_op = "fadd";
+ base_op = "fadd";
break;
case MINUS:
- if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
- || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
- base_op = "fisub";
- else
- base_op = "fsub";
+ base_op = "fsub";
break;
case MULT:
- if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
- || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
- base_op = "fimul";
- else
- base_op = "fmul";
+ base_op = "fmul";
break;
case DIV:
- if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
- || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
- base_op = "fidiv";
- else
- base_op = "fdiv";
+ base_op = "fdiv";
break;
default:
if (GET_CODE (operands[2]) == MEM)
return strcat (buf, AS1 (%z2,%2));
- if (NON_STACK_REG_P (operands[1]))
- {
- output_op_from_reg (operands[1], strcat (buf, AS1 (%z0,%1)));
- return "";
- }
-
- else if (NON_STACK_REG_P (operands[2]))
- {
- output_op_from_reg (operands[2], strcat (buf, AS1 (%z0,%1)));
- return "";
- }
+ if (! STACK_REG_P (operands[1]) || ! STACK_REG_P (operands[2]))
+ abort ();
if (find_regno_note (insn, REG_DEAD, REGNO (operands[2])))
{
if (GET_CODE (operands[2]) == MEM)
return strcat (buf, AS1 (%z2,%2));
- if (NON_STACK_REG_P (operands[1]))
- {
- output_op_from_reg (operands[1], strcat (buf, AS1 (r%z0,%1)));
- return "";
- }
-
- else if (NON_STACK_REG_P (operands[2]))
- {
- output_op_from_reg (operands[2], strcat (buf, AS1 (%z0,%1)));
- return "";
- }
-
if (! STACK_REG_P (operands[1]) || ! STACK_REG_P (operands[2]))
abort ();
return AS1 (fldc%W2,%2);
}
\f
+/* Output code for INSN to extend a float. OPERANDS are the insn
+ operands. The output may be DFmode or XFmode and the input operand
+ may be SFmode or DFmode. Operands 2 and 3 are scratch memory and
+ are only necessary if operands 0 or 1 are non-stack registers. */
+
+void
+output_float_extend (insn, operands)
+ rtx insn;
+ rtx *operands;
+{
+ int stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0;
+ rtx xops[2];
+
+ if (! STACK_TOP_P (operands[0]) && ! STACK_TOP_P (operands[1]))
+ abort ();
+
+ if (STACK_TOP_P (operands[0]) && STACK_TOP_P (operands[1]) && stack_top_dies)
+ return;
+
+ if (STACK_TOP_P (operands[0]) )
+ {
+ if (NON_STACK_REG_P (operands[1]))
+ {
+ if (GET_MODE (operands[1]) == SFmode)
+ output_asm_insn (AS2 (mov%L0,%1,%2), operands);
+ else
+ {
+ xops[0] = operands[2];
+ xops[1] = operands[1];
+ output_asm_insn (output_move_double (xops), xops);
+ }
+ }
+
+ xops[0] = NON_STACK_REG_P (operands[1]) ? operands[2] : operands[1];
+
+ output_asm_insn (AS1 (fld%z0,%y0), xops);
+ }
+ else
+ {
+ xops[0] = NON_STACK_REG_P (operands[0]) ? operands[3] : operands[0];
+
+ if (stack_top_dies
+ || (GET_CODE (xops[0]) == MEM && GET_MODE (xops[0]) == XFmode))
+ {
+ output_asm_insn (AS1 (fstp%z0,%y0), xops);
+ if (! stack_top_dies)
+ output_asm_insn (AS1 (fld%z0,%y0), xops);
+ }
+ else
+ output_asm_insn (AS1 (fst%z0,%y0), xops);
+
+ if (NON_STACK_REG_P (operands[0]))
+ {
+ xops[0] = operands[0];
+ xops[1] = operands[3];
+ output_asm_insn (output_move_double (xops), xops);
+ }
+ }
+}
+\f
/* Output code for INSN to compare OPERANDS. The two operands might
not have the same mode: one might be within a FLOAT or FLOAT_EXTEND
expression. If the compare is in mode CCFPEQmode, use an opcode that
operands[1] = tmp;
cc_status.flags |= CC_REVERSED;
}
-
+
if (! STACK_TOP_P (operands[0]))
abort ();
{
static char buf[100];
- /* Decide if this is the integer or float compare opcode, or the
- unordered float compare. */
+ /* Decide if this is a float compare or an unordered float compare. */
if (unordered_compare)
strcpy (buf, (cc_status.flags & CC_FCOMI) ? "fucomi" : "fucom");
- else if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_FLOAT)
- strcpy (buf, (cc_status.flags & CC_FCOMI) ? "fcomi" : "fcom");
else
- strcpy (buf, "ficom");
+ strcpy (buf, (cc_status.flags & CC_FCOMI) ? "fcomi" : "fcom");
/* Modify the opcode if the 387 stack is to be popped. */
if (stack_top_dies)
strcat (buf, "p");
- if (NON_STACK_REG_P (operands[1]))
- output_op_from_reg (operands[1], strcat (buf, AS1 (%z0,%1)));
- else if (cc_status.flags & CC_FCOMI)
+ if (cc_status.flags & CC_FCOMI)
{
output_asm_insn (strcat (buf, AS2 (%z1,%y1,%0)), operands);
return "";
if (!(cc_status.flags & CC_REVERSED))
{
next = next_cc0_user (insn);
-
+
if (GET_CODE (next) == JUMP_INSN
&& GET_CODE (PATTERN (next)) == SET
&& SET_DEST (PATTERN (next)) == pc_rtx
copy->integrated = orig->integrated;
/* intel1 */
copy->is_spill_rtx = orig->is_spill_rtx;
-
+
format_ptr = GET_RTX_FORMAT (GET_CODE (copy));
for (i = 0; i < GET_RTX_LENGTH (GET_CODE (copy)); i++)
\f
/* Try to rewrite a memory address to make it valid */
-void
+void
rewrite_address (mem_rtx)
rtx mem_rtx;
{
obfree (storage);
}
- /* This part is utilized by loop.c.
+ /* This part is utilized by loop.c.
If the address contains PLUS (reg,const) and this pattern is invalid
in this case - try to rewrite the address to make it valid. */
storage = oballoc (0);
|| GET_MODE (SET_DEST (PATTERN (insn))) == XFmode)
&& GET_CODE (SET_DEST (PATTERN (insn))) == REG
&& REGNO (SET_DEST (PATTERN (insn))) >= FIRST_FLOAT_REG
- && GET_CODE (SET_SRC (insn)) != MEM)
+ && GET_CODE (SET_SRC (PATTERN (insn))) != MEM)
return 1;
return 0;
agi_dependent (insn, dep_insn)
rtx insn, dep_insn;
{
+ int push = 0, push_dep = 0;
if (GET_CODE (dep_insn) == INSN
&& GET_CODE (PATTERN (dep_insn)) == SET
- && GET_CODE (SET_DEST (PATTERN (dep_insn))) == REG)
- return reg_mentioned_in_mem (SET_DEST (PATTERN (dep_insn)), insn);
+ && GET_CODE (SET_DEST (PATTERN (dep_insn))) == REG
+ && reg_mentioned_in_mem (SET_DEST (PATTERN (dep_insn)), insn))
+ return 1;
+
+ if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SET
+ && GET_CODE (SET_DEST (PATTERN (insn))) == MEM
+ && push_operand (SET_DEST (PATTERN (insn)),
+ GET_MODE (SET_DEST (PATTERN (insn)))))
+ push = 1;
if (GET_CODE (dep_insn) == INSN && GET_CODE (PATTERN (dep_insn)) == SET
&& GET_CODE (SET_DEST (PATTERN (dep_insn))) == MEM
&& push_operand (SET_DEST (PATTERN (dep_insn)),
GET_MODE (SET_DEST (PATTERN (dep_insn)))))
- return reg_mentioned_in_mem (stack_pointer_rtx, insn);
+ push_dep = 1;
+
+ /* CPUs contain special hardware to allow two pushes. */
+ if (push && push_dep)
+ return 0;
+
+ /* Push operation implicitly change stack pointer causing AGI stalls. */
+ if (push_dep && reg_mentioned_in_mem (stack_pointer_rtx, insn))
+ return 1;
+
+ /* Push also implicitly read stack pointer. */
+ if (push && modified_in_p (stack_pointer_rtx, dep_insn))
+ return 1;
return 0;
}
/* Check third byte. */
output_asm_insn (AS2 (test%L1,%16,%1), xops);
output_asm_insn (AS1 (je,%l10), xops);
-
+
/* Check fourth byte and increment address. */
output_asm_insn (AS2 (add%L0,%5,%0), xops);
output_asm_insn (AS2 (test%L1,%17,%1), xops);
int which_alternative;
rtx operands[];
{
+ enum rtx_code code = GET_CODE (operands[1]);
+
+ /* This should never happen. */
+ if (!(cc_prev_status.flags & CC_IN_80387)
+ && (code == GT || code == LE || code == GE || code == LT))
+ abort ();
+
switch (which_alternative)
{
case 0:
/* r <- cond ? arg : r */
output_asm_insn (AS2 (fcmov%F1,%2,%0), operands);
break;
-
- case 1:
- /* r <- cond ? r : arg */
- output_asm_insn (AS2 (fcmov%f1,%3,%0), operands);
- break;
- case 2:
+ case 1:
/* r <- cond ? r : arg */
- output_asm_insn (AS2 (fcmov%F1,%2,%0), operands);
output_asm_insn (AS2 (fcmov%f1,%3,%0), operands);
break;
int which_alternative;
rtx operands[];
{
- int code = GET_CODE (operands[1]);
- enum machine_mode mode;
- rtx xops[4];
+ enum rtx_code code = GET_CODE (operands[1]);
/* This is very tricky. We have to do it right. For a code segement
like:
&& (cc_prev_status.flags & CC_NO_OVERFLOW))
return NULL_PTR;
- mode = GET_MODE (operands [0]);
- if (mode == DImode)
- {
- xops [0] = gen_rtx_SUBREG (SImode, operands [0], 1);
- xops [1] = operands [1];
- xops [2] = gen_rtx_SUBREG (SImode, operands [2], 1);
- xops [3] = gen_rtx_SUBREG (SImode, operands [3], 1);
- }
-
switch (which_alternative)
{
case 0:
/* r <- cond ? arg : r */
output_asm_insn (AS2 (cmov%C1,%2,%0), operands);
- if (mode == DImode)
- output_asm_insn (AS2 (cmov%C1,%2,%0), xops);
break;
case 1:
/* r <- cond ? r : arg */
output_asm_insn (AS2 (cmov%c1,%3,%0), operands);
- if (mode == DImode)
- output_asm_insn (AS2 (cmov%c1,%3,%0), xops);
break;
- case 2:
- /* rm <- cond ? arg1 : arg2 */
- output_asm_insn (AS2 (cmov%C1,%2,%0), operands);
- output_asm_insn (AS2 (cmov%c1,%3,%0), operands);
- if (mode == DImode)
+ default:
+ abort ();
+ }
+
+ return "";
+}
+
+int
+x86_adjust_cost (insn, link, dep_insn, cost)
+ rtx insn, link, dep_insn;
+ int cost;
+{
+ rtx next_inst;
+
+ if (GET_CODE (dep_insn) == CALL_INSN || GET_CODE (insn) == JUMP_INSN)
+ return 0;
+
+ if (GET_CODE (dep_insn) == INSN
+ && GET_CODE (PATTERN (dep_insn)) == SET
+ && GET_CODE (SET_DEST (PATTERN (dep_insn))) == REG
+ && GET_CODE (insn) == INSN
+ && GET_CODE (PATTERN (insn)) == SET
+ && !reg_overlap_mentioned_p (SET_DEST (PATTERN (dep_insn)),
+ SET_SRC (PATTERN (insn))))
+ return 0; /* ??? */
+
+
+ switch (ix86_cpu)
+ {
+ case PROCESSOR_PENTIUM:
+ if (cost != 0 && is_fp_insn (insn) && is_fp_insn (dep_insn)
+ && !is_fp_dest (dep_insn))
+ return 0;
+
+ if (agi_dependent (insn, dep_insn))
+ return cost ? cost + 1 : 2;
+
+ if (GET_CODE (insn) == INSN
+ && GET_CODE (PATTERN (insn)) == SET
+ && SET_DEST (PATTERN (insn)) == cc0_rtx
+ && (next_inst = next_nonnote_insn (insn))
+ && GET_CODE (next_inst) == JUMP_INSN)
+ /* compare probably paired with jump */
+ return 0;
+ break;
+
+ /* Stores stalls one cycle longer than other insns. */
+ if (is_fp_insn (insn) && cost && is_fp_store (dep_insn))
+ cost++;
+
+ case PROCESSOR_K6:
+ default:
+ if (!is_fp_dest (dep_insn))
{
- output_asm_insn (AS2 (cmov%C1,%2,%0), xops);
- output_asm_insn (AS2 (cmov%c1,%3,%0), xops);
+ if(!agi_dependent (insn, dep_insn))
+ return 0;
+ if (TARGET_486)
+ return 2;
}
+ else
+ if (is_fp_store (insn) && is_fp_insn (dep_insn)
+ && NEXT_INSN (insn) && NEXT_INSN (NEXT_INSN (insn))
+ && NEXT_INSN (NEXT_INSN (NEXT_INSN (insn)))
+ && (GET_CODE (NEXT_INSN (insn)) == INSN)
+ && (GET_CODE (NEXT_INSN (NEXT_INSN (insn))) == JUMP_INSN)
+ && (GET_CODE (NEXT_INSN (NEXT_INSN (NEXT_INSN (insn)))) == NOTE)
+ && (NOTE_LINE_NUMBER (NEXT_INSN (NEXT_INSN (NEXT_INSN (insn))))
+ == NOTE_INSN_LOOP_END))
+ return 3;
break;
+ }
+
+ return cost;
+}
+
+/* Output assembly code for a left shift.
+
+ Always use "sal" when shifting a memory operand or for a non constant
+ shift count.
+
+ When optimizing for size, we know that src == dest, and we should always
+ use "sal". If src != dest, then copy src to dest and use "sal".
+
+ Pentium and PPro (speed):
+
+ When src == dest, use "add" for a shift counts of one, else use
+ "sal". If we modeled Pentium AGI stalls and U/V pipelining better we
+ would want to generate lea for some shifts on the Pentium.
+
+ When src != dest, use "lea" for small shift counts. Otherwise,
+ copy src to dest and use the normal shifting code. Exception for
+ TARGET_DOUBLE_WITH_ADD. */
+
+char *
+output_ashl (insn, operands)
+ rtx insn, *operands;
+{
+ /* Handle case where srcreg != dstreg. */
+ if (REG_P (operands[0]) && REGNO (operands[0]) != REGNO (operands[1]))
+ {
+ if (TARGET_DOUBLE_WITH_ADD && INTVAL (operands[2]) == 1)
+ switch (GET_MODE (operands[0]))
+ {
+ case SImode:
+ output_asm_insn (AS2 (mov%L0,%1,%0), operands);
+ return AS2 (add%L0,%1,%0);
+ case HImode:
+ output_asm_insn (AS2 (mov%L0,%k1,%k0), operands);
+ if (i386_cc_probably_useless_p (insn))
+ {
+ CC_STATUS_INIT;
+ return AS2 (add%L0,%k1,%k0);
+ }
+ return AS2 (add%W0,%k1,%k0);
+ case QImode:
+ output_asm_insn (AS2 (mov%B0,%1,%0), operands);
+ return AS2 (add%B0,%1,%0);
+ default:
+ abort ();
+ }
+ else
+ {
+ CC_STATUS_INIT;
+
+ /* This should be extremely rare (impossible?). We can not encode a
+ shift of the stack pointer using an lea instruction. So copy the
+ stack pointer into the destination register and use an lea. */
+ if (operands[1] == stack_pointer_rtx)
+ {
+ output_asm_insn (AS2 (mov%L0,%k1,%k0), operands);
+ operands[1] = operands[0];
+ }
+
+ /* For shifts up to and including 3 bits, use lea. */
+ operands[1] = gen_rtx_MULT (SImode,
+ gen_rtx_REG (SImode, REGNO (operands[1])),
+ GEN_INT (1 << INTVAL (operands[2])));
+ return AS2 (lea%L0,%a1,%k0);
+ }
+ }
+
+ /* Source and destination match. */
+
+ /* Handle variable shift. */
+ if (REG_P (operands[2]))
+ switch (GET_MODE (operands[0]))
+ {
+ case SImode:
+ return AS2 (sal%L0,%b2,%0);
+ case HImode:
+ if (REG_P (operands[0]) && i386_cc_probably_useless_p (insn))
+ {
+ CC_STATUS_INIT;
+ return AS2 (sal%L0,%b2,%k0);
+ }
+ else
+ return AS2 (sal%W0,%b2,%0);
+ case QImode:
+ return AS2 (sal%B0,%b2,%0);
+ default:
+ abort ();
+ }
+
+ /* Always perform shift by 1 using an add instruction. */
+ if (REG_P (operands[0]) && operands[2] == const1_rtx)
+ switch (GET_MODE (operands[0]))
+ {
+ case SImode:
+ return AS2 (add%L0,%0,%0);
+ case HImode:
+ if (REG_P (operands[0]) && i386_cc_probably_useless_p (insn))
+ {
+ CC_STATUS_INIT;
+ return AS2 (add%L0,%k0,%k0);
+ }
+ else
+ return AS2 (add%W0,%0,%0);
+ case QImode:
+ return AS2 (add%B0,%0,%0);
+ default:
+ abort ();
+ }
+
+#if 0
+ /* ??? Currently disabled. Because our model of Pentium is far from being
+ exact, this change will need some benchmarking. */
+ /* Shift reg by 2 or 3 use an lea instruction for Pentium if this is
+ insn is expected to issue into the V pipe (the insn's mode will be
+ TImode for a U pipe, and !TImode for a V pipe instruction). */
+ if (! optimize_size
+ && REG_P (operands[0])
+ && GET_CODE (operands[2]) == CONST_INT
+ && INTVAL (operands[2]) <= 3
+ && (int)ix86_cpu == (int)PROCESSOR_PENTIUM
+ && GET_MODE (insn) != TImode)
+ {
+ CC_STATUS_INIT;
+ operands[1] = gen_rtx_MULT (SImode, gen_rtx_REG (SImode, REGNO (operands[1])),
+ GEN_INT (1 << INTVAL (operands[2])));
+ return AS2 (lea%L0,%a1,%0);
+ }
+#endif
+ /* Otherwise use a shift instruction. */
+ switch (GET_MODE (operands[0]))
+ {
+ case SImode:
+ return AS2 (sal%L0,%2,%0);
+ case HImode:
+ if (REG_P (operands[0]) && i386_cc_probably_useless_p (insn))
+ {
+ CC_STATUS_INIT;
+ return AS2 (sal%L0,%2,%k0);
+ }
+ else
+ return AS2 (sal%W0,%2,%0);
+ case QImode:
+ return AS2 (sal%B0,%2,%0);
default:
abort ();
}
+}
- return "";
+/* Given the memory address ADDR, calculate the length of the address or
+ the length of just the displacement (controlled by DISP_LENGTH).
+
+ The length returned does not include the one-byte modrm, opcode,
+ or prefix. */
+
+int
+memory_address_info (addr, disp_length)
+ rtx addr;
+ int disp_length;
+{
+ rtx base, index, disp, scale;
+ rtx op0, op1;
+ int len;
+
+ if (GET_CODE (addr) == PRE_DEC
+ || GET_CODE (addr) == POST_INC)
+ return 0;
+
+ /* Register Indirect. */
+ if (register_operand (addr, Pmode))
+ {
+ /* Special cases: ebp and esp need the two-byte modrm form.
+
+ We change [ESI] to [ESI+0] on the K6 when not optimizing
+ for size. */
+ if (addr == stack_pointer_rtx
+ || addr == arg_pointer_rtx
+ || addr == frame_pointer_rtx
+ || (REGNO_REG_CLASS (REGNO (addr)) == SIREG
+ && ix86_cpu == PROCESSOR_K6 && !optimize_size))
+ return 1;
+ else
+ return 0;
+ }
+
+ /* Direct Addressing. */
+ if (CONSTANT_P (addr))
+ return 4;
+
+ index = base = disp = scale = NULL_RTX;
+ op0 = XEXP (addr, 0);
+ op1 = XEXP (addr, 1);
+
+ if (GET_CODE (addr) == PLUS)
+ {
+ if (register_operand (op0, Pmode))
+ {
+ if (register_operand (op1, Pmode))
+ index = op0, base = op1;
+ else
+ base = op0, disp = op1;
+ }
+ else if (GET_CODE (op0) == MULT)
+ {
+ index = XEXP (op0, 0);
+ scale = XEXP (op0, 1);
+ if (register_operand (op1, Pmode))
+ base = op1;
+ else
+ disp = op1;
+ }
+ else if (GET_CODE (op0) == PLUS && GET_CODE (XEXP (op0, 0)) == MULT)
+ {
+ index = XEXP (XEXP (op0, 0), 0);
+ scale = XEXP (XEXP (op0, 0), 1);
+ base = XEXP (op0, 1);
+ disp = op1;
+ }
+ else if (GET_CODE (op0) == PLUS)
+ {
+ index = XEXP (op0, 0);
+ base = XEXP (op0, 1);
+ disp = op1;
+ }
+ else
+ abort ();
+ }
+ else if (GET_CODE (addr) == MULT
+ /* We're called for lea too, which implements ashift on occasion. */
+ || GET_CODE (addr) == ASHIFT)
+ {
+ index = XEXP (addr, 0);
+ scale = XEXP (addr, 1);
+ }
+ else
+ abort ();
+
+ /* Allow arg pointer and stack pointer as index if there is not scaling */
+ if (base && index && !scale
+ && (index == stack_pointer_rtx
+ || index == arg_pointer_rtx
+ || index == frame_pointer_rtx))
+ {
+ rtx tmp = base;
+ base = index;
+ index = tmp;
+ }
+
+ /* Special case: ebp cannot be encoded as a base without a displacement. */
+ if (base == frame_pointer_rtx && !disp)
+ disp = const0_rtx;
+
+ /* Scaling can not be encoded without base or displacement.
+ Except for scale == 1 where we can encode reg + reg instead of reg * 2. */
+ if (!base && index && scale != 1)
+ disp = const0_rtx;
+
+ /* Find the length of the displacement constant. */
+ len = 0;
+ if (disp)
+ {
+ if (GET_CODE (disp) == CONST_INT
+ && CONST_OK_FOR_LETTER_P (INTVAL (disp), 'K'))
+ len = 1;
+ else
+ len = 4;
+ }
+
+ /* An index requires the two-byte modrm form. Not important
+ if we are computing just length of the displacement. */
+ if (index && ! disp_length)
+ len += 1;
+
+ return len;
}