/* Subroutines used for MIPS code generation.
Copyright (C) 1989, 1990, 1991, 1993, 1994, 1995, 1996, 1997, 1998,
- 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
+ 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
Free Software Foundation, Inc.
Contributed by A. Lichnewsky, lich@inria.inria.fr.
Changes by Michael Meissner, meissner@osf.org.
ADDRESS_SYMBOLIC
};
+/* Enumerates the setting of the -mr10k-cache-barrier option. */
+enum mips_r10k_cache_barrier_setting {
+ R10K_CACHE_BARRIER_NONE,
+ R10K_CACHE_BARRIER_STORE,
+ R10K_CACHE_BARRIER_LOAD_STORE
+};
+
/* Macros to create an enumeration identifier for a function prototype. */
#define MIPS_FTYPE_NAME1(A, B) MIPS_##A##_FTYPE_##B
#define MIPS_FTYPE_NAME2(A, B, C) MIPS_##A##_FTYPE_##B##_##C
/* The current frame information, calculated by mips_compute_frame_info. */
struct mips_frame_info frame;
- /* The register to use as the function's global pointer. */
+ /* The register to use as the function's global pointer, or INVALID_REGNUM
+ if the function doesn't need one. */
unsigned int global_pointer;
/* True if mips_adjust_insn_length should ignore an instruction's
bool mips_base_mips16;
/* The ambient values of other global variables. */
-static int mips_base_delayed_branch; /* flag_delayed_branch */
static int mips_base_schedule_insns; /* flag_schedule_insns */
static int mips_base_reorder_blocks_and_partition; /* flag_reorder... */
static int mips_base_move_loop_invariants; /* flag_move_loop_invariants */
/* The -mcode-readable setting. */
enum mips_code_readable_setting mips_code_readable = CODE_READABLE_YES;
+/* The -mr10k-cache-barrier setting. */
+static enum mips_r10k_cache_barrier_setting mips_r10k_cache_barrier;
+
/* Index [M][R] is true if register R is allowed to hold a value of mode M. */
bool mips_hard_regno_mode_ok[(int) MAX_MACHINE_MODE][FIRST_PSEUDO_REGISTER];
MD0_REG, MD1_REG, NO_REGS, ST_REGS,
ST_REGS, ST_REGS, ST_REGS, ST_REGS,
ST_REGS, ST_REGS, ST_REGS, NO_REGS,
- NO_REGS, ALL_REGS, ALL_REGS, NO_REGS,
+ NO_REGS, FRAME_REGS, FRAME_REGS, NO_REGS,
COP0_REGS, COP0_REGS, COP0_REGS, COP0_REGS,
COP0_REGS, COP0_REGS, COP0_REGS, COP0_REGS,
COP0_REGS, COP0_REGS, COP0_REGS, COP0_REGS,
/* MIPS IV processors. */
{ "r8000", PROCESSOR_R8000, 4, 0 },
+ { "r10000", PROCESSOR_R10000, 4, 0 },
+ { "r12000", PROCESSOR_R10000, 4, 0 },
+ { "r14000", PROCESSOR_R10000, 4, 0 },
+ { "r16000", PROCESSOR_R10000, 4, 0 },
{ "vr5000", PROCESSOR_R5000, 4, 0 },
{ "vr5400", PROCESSOR_R5400, 4, 0 },
{ "vr5500", PROCESSOR_R5500, 4, PTF_AVOID_BRANCHLIKELY },
{ "sb1", PROCESSOR_SB1, 64, PTF_AVOID_BRANCHLIKELY },
{ "sb1a", PROCESSOR_SB1A, 64, PTF_AVOID_BRANCHLIKELY },
{ "sr71000", PROCESSOR_SR71000, 64, PTF_AVOID_BRANCHLIKELY },
- { "xlr", PROCESSOR_XLR, 64, 0 }
+ { "xlr", PROCESSOR_XLR, 64, 0 },
+
+ /* MIPS64 Release 2 processors. */
+ { "octeon", PROCESSOR_OCTEON, 65, PTF_AVOID_BRANCHLIKELY }
};
/* Default costs. If these are used for a processor we should look
{ /* M4k */
DEFAULT_COSTS
},
+ /* Octeon */
+ {
+ SOFT_FP_COSTS,
+ COSTS_N_INSNS (5), /* int_mult_si */
+ COSTS_N_INSNS (5), /* int_mult_di */
+ COSTS_N_INSNS (72), /* int_div_si */
+ COSTS_N_INSNS (72), /* int_div_di */
+ 1, /* branch_cost */
+ 4 /* memory_latency */
+ },
{ /* R3900 */
COSTS_N_INSNS (2), /* fp_add */
COSTS_N_INSNS (4), /* fp_mult_sf */
1, /* branch_cost */
4 /* memory_latency */
},
+ { /* R1x000 */
+ COSTS_N_INSNS (2), /* fp_add */
+ COSTS_N_INSNS (2), /* fp_mult_sf */
+ COSTS_N_INSNS (2), /* fp_mult_df */
+ COSTS_N_INSNS (12), /* fp_div_sf */
+ COSTS_N_INSNS (19), /* fp_div_df */
+ COSTS_N_INSNS (5), /* int_mult_si */
+ COSTS_N_INSNS (9), /* int_mult_di */
+ COSTS_N_INSNS (34), /* int_div_si */
+ COSTS_N_INSNS (66), /* int_div_di */
+ 1, /* branch_cost */
+ 4 /* memory_latency */
+ },
{ /* SB1 */
/* These costs are the same as the SB-1A below. */
COSTS_N_INSNS (4), /* fp_add */
if (mips16_stub_function_p (x))
return false;
- /* When TARGET_ABSOLUTE_ABICALLS is true, locally-defined functions
- use absolute accesses to set up the global pointer. */
- if (TARGET_ABSOLUTE_ABICALLS
- && GET_CODE (x) == SYMBOL_REF
- && mips_symbol_binds_local_p (x)
- && !SYMBOL_REF_EXTERNAL_P (x))
- return false;
+ if (GET_CODE (x) == SYMBOL_REF)
+ {
+ /* If PLTs and copy relocations are available, the static linker
+ will make sure that $25 is valid on entry to the target function. */
+ if (TARGET_ABICALLS_PIC0)
+ return false;
+
+ /* Locally-defined functions use absolute accesses to set up
+ the global pointer. */
+ if (TARGET_ABSOLUTE_ABICALLS
+ && mips_symbol_binds_local_p (x)
+ && !SYMBOL_REF_EXTERNAL_P (x))
+ return false;
+ }
return true;
}
/* Don't use GOT accesses for locally-binding symbols when -mno-shared
is in effect. */
- if (TARGET_ABICALLS
+ if (TARGET_ABICALLS_PIC2
&& !(TARGET_ABSOLUTE_ABICALLS && mips_symbol_binds_local_p (x)))
{
/* There are three cases to consider:
else
cost = single_cost;
return (cost
- + rtx_cost (XEXP (x, 0), 0)
- + rtx_cost (XEXP (x, 1), GET_CODE (x)));
+ + rtx_cost (XEXP (x, 0), 0, !optimize_size)
+ + rtx_cost (XEXP (x, 1), GET_CODE (x), !optimize_size));
}
/* Return the cost of floating-point multiplications of mode MODE. */
/* Implement TARGET_RTX_COSTS. */
static bool
-mips_rtx_costs (rtx x, int code, int outer_code, int *total)
+mips_rtx_costs (rtx x, int code, int outer_code, int *total,
+ bool speed)
{
enum machine_mode mode = GET_MODE (x);
bool float_mode_p = FLOAT_MODE_P (mode);
&& UINTVAL (XEXP (x, 1)) == 0xffffffff)
{
*total = (mips_zero_extend_cost (mode, XEXP (x, 0))
- + rtx_cost (XEXP (x, 0), 0));
+ + rtx_cost (XEXP (x, 0), 0, speed));
return true;
}
/* Fall through. */
case LO_SUM:
/* Low-part immediates need an extended MIPS16 instruction. */
*total = (COSTS_N_INSNS (TARGET_MIPS16 ? 2 : 1)
- + rtx_cost (XEXP (x, 0), 0));
+ + rtx_cost (XEXP (x, 0), 0, speed));
return true;
case LT:
if (GET_CODE (op0) == MULT && GET_CODE (XEXP (op0, 0)) == NEG)
{
*total = (mips_fp_mult_cost (mode)
- + rtx_cost (XEXP (XEXP (op0, 0), 0), 0)
- + rtx_cost (XEXP (op0, 1), 0)
- + rtx_cost (op1, 0));
+ + rtx_cost (XEXP (XEXP (op0, 0), 0), 0, speed)
+ + rtx_cost (XEXP (op0, 1), 0, speed)
+ + rtx_cost (op1, 0, speed));
return true;
}
if (GET_CODE (op1) == MULT)
{
*total = (mips_fp_mult_cost (mode)
- + rtx_cost (op0, 0)
- + rtx_cost (XEXP (op1, 0), 0)
- + rtx_cost (XEXP (op1, 1), 0));
+ + rtx_cost (op0, 0, speed)
+ + rtx_cost (XEXP (op1, 0), 0, speed)
+ + rtx_cost (XEXP (op1, 1), 0, speed));
return true;
}
}
&& GET_CODE (XEXP (op, 0)) == MULT)
{
*total = (mips_fp_mult_cost (mode)
- + rtx_cost (XEXP (XEXP (op, 0), 0), 0)
- + rtx_cost (XEXP (XEXP (op, 0), 1), 0)
- + rtx_cost (XEXP (op, 1), 0));
+ + rtx_cost (XEXP (XEXP (op, 0), 0), 0, speed)
+ + rtx_cost (XEXP (XEXP (op, 0), 1), 0, speed)
+ + rtx_cost (XEXP (op, 1), 0, speed));
return true;
}
}
if (outer_code == SQRT || GET_CODE (XEXP (x, 1)) == SQRT)
/* An rsqrt<mode>a or rsqrt<mode>b pattern. Count the
division as being free. */
- *total = rtx_cost (XEXP (x, 1), 0);
+ *total = rtx_cost (XEXP (x, 1), 0, speed);
else
- *total = mips_fp_div_cost (mode) + rtx_cost (XEXP (x, 1), 0);
+ *total = mips_fp_div_cost (mode) + rtx_cost (XEXP (x, 1), 0, speed);
return true;
}
/* Fall through. */
&& CONST_INT_P (XEXP (x, 1))
&& exact_log2 (INTVAL (XEXP (x, 1))) >= 0)
{
- *total = COSTS_N_INSNS (2) + rtx_cost (XEXP (x, 0), 0);
+ *total = COSTS_N_INSNS (2) + rtx_cost (XEXP (x, 0), 0, speed);
return true;
}
*total = COSTS_N_INSNS (mips_idiv_insns ());
/* Implement TARGET_ADDRESS_COST. */
static int
-mips_address_cost (rtx addr)
+mips_address_cost (rtx addr, bool speed ATTRIBUTE_UNUSED)
{
return mips_address_insns (addr, SImode, false);
}
if (code == EQ || code == NE)
{
- rtx zie = mips_zero_if_equal (cmp_operands[0], cmp_operands[1]);
- mips_emit_binary (code, target, zie, const0_rtx);
+ if (ISA_HAS_SEQ_SNE
+ && reg_imm10_operand (cmp_operands[1], GET_MODE (cmp_operands[1])))
+ mips_emit_binary (code, target, cmp_operands[0], cmp_operands[1]);
+ else
+ {
+ rtx zie = mips_zero_if_equal (cmp_operands[0], cmp_operands[1]);
+ mips_emit_binary (code, target, zie, const0_rtx);
+ }
}
else
mips_emit_int_order_test (code, 0, target,
if (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT
&& GET_MODE_SIZE (TYPE_MODE (type)) <= UNITS_PER_FPVALUE)
{
- top = build3 (COMPONENT_REF, TREE_TYPE (f_ftop), valist, f_ftop,
- NULL_TREE);
- off = build3 (COMPONENT_REF, TREE_TYPE (f_foff), valist, f_foff,
- NULL_TREE);
+ top = build3 (COMPONENT_REF, TREE_TYPE (f_ftop),
+ unshare_expr (valist), f_ftop, NULL_TREE);
+ off = build3 (COMPONENT_REF, TREE_TYPE (f_foff),
+ unshare_expr (valist), f_foff, NULL_TREE);
/* When va_start saves FPR arguments to the stack, each slot
takes up UNITS_PER_HWFPVALUE bytes, regardless of the
}
else
{
- top = build3 (COMPONENT_REF, TREE_TYPE (f_gtop), valist, f_gtop,
- NULL_TREE);
- off = build3 (COMPONENT_REF, TREE_TYPE (f_goff), valist, f_goff,
- NULL_TREE);
+ top = build3 (COMPONENT_REF, TREE_TYPE (f_gtop),
+ unshare_expr (valist), f_gtop, NULL_TREE);
+ off = build3 (COMPONENT_REF, TREE_TYPE (f_goff),
+ unshare_expr (valist), f_goff, NULL_TREE);
rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD;
if (rsize > UNITS_PER_WORD)
{
/* [1] Emit code for: off &= -rsize. */
- t = build2 (BIT_AND_EXPR, TREE_TYPE (off), off,
- build_int_cst (NULL_TREE, -rsize));
- gimplify_assign (off, t, pre_p);
+ t = build2 (BIT_AND_EXPR, TREE_TYPE (off), unshare_expr (off),
+ build_int_cst (TREE_TYPE (off), -rsize));
+ gimplify_assign (unshare_expr (off), t, pre_p);
}
osize = rsize;
}
{
/* [9] Emit: ovfl = ((intptr_t) ovfl + osize - 1) & -osize. */
u = size_int (osize - 1);
- t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (ovfl), ovfl, u);
+ t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (ovfl),
+ unshare_expr (ovfl), u);
t = fold_convert (sizetype, t);
u = size_int (-osize);
t = build2 (BIT_AND_EXPR, sizetype, t, u);
t = fold_convert (TREE_TYPE (ovfl), t);
- align = build2 (MODIFY_EXPR, TREE_TYPE (ovfl), ovfl, t);
+ align = build2 (MODIFY_EXPR, TREE_TYPE (ovfl),
+ unshare_expr (ovfl), t);
}
else
align = NULL;
assemble_start_function (stubdecl, stubname);
mips_start_function_definition (stubname, false);
- /* If generating abicalls code, either set up the global pointer or
- switch to absolute mode. */
- if (TARGET_ABSOLUTE_ABICALLS)
- fprintf (asm_out_file, "\t.option\tpic0\n");
- else if (TARGET_ABICALLS)
+ /* If generating pic2 code, either set up the global pointer or
+ switch to pic0. */
+ if (TARGET_ABICALLS_PIC2)
{
- output_asm_insn ("%(.cpload\t%^%)", NULL);
- /* Emit an R_MIPS_NONE relocation to tell the linker what the
- target function is. Use a local GOT access when loading the
- symbol, to cut down on the number of unnecessary GOT entries
- for stubs that aren't needed. */
- output_asm_insn (".reloc\t0,R_MIPS_NONE,%0", &symbol);
- symbol = alias;
+ if (TARGET_ABSOLUTE_ABICALLS)
+ fprintf (asm_out_file, "\t.option\tpic0\n");
+ else
+ {
+ output_asm_insn ("%(.cpload\t%^%)", NULL);
+ /* Emit an R_MIPS_NONE relocation to tell the linker what the
+ target function is. Use a local GOT access when loading the
+ symbol, to cut down on the number of unnecessary GOT entries
+ for stubs that aren't needed. */
+ output_asm_insn (".reloc\t0,R_MIPS_NONE,%0", &symbol);
+ symbol = alias;
+ }
}
- /* Load the address of the MIPS16 function into $at. Do this first so
+ /* Load the address of the MIPS16 function into $25. Do this first so
that targets with coprocessor interlocks can use an MFC1 to fill the
delay slot. */
- fprintf (asm_out_file, "\t.set\tnoat\n");
- output_asm_insn ("la\t%@,%0", &symbol);
+ output_asm_insn ("la\t%^,%0", &symbol);
/* Move the arguments from floating-point registers to general registers. */
mips_output_args_xfer (crtl->args.info.fp_code, 'f');
/* Jump to the MIPS16 function. */
- output_asm_insn ("jr\t%@", NULL);
- fprintf (asm_out_file, "\t.set\tat\n");
+ output_asm_insn ("jr\t%^", NULL);
- if (TARGET_ABSOLUTE_ABICALLS)
+ if (TARGET_ABICALLS_PIC2 && TARGET_ABSOLUTE_ABICALLS)
fprintf (asm_out_file, "\t.option\tpic2\n");
mips_end_function_definition (stubname);
if (!fp_ret_p)
{
- /* Load the address of the MIPS16 function into $at. Do this
+ /* Load the address of the MIPS16 function into $25. Do this
first so that targets with coprocessor interlocks can use
an MFC1 to fill the delay slot. */
- fprintf (asm_out_file, "\t.set\tnoat\n");
if (TARGET_EXPLICIT_RELOCS)
{
output_asm_insn ("lui\t%^,%%hi(%0)", &fn);
{
/* Jump to the previously-loaded address. */
output_asm_insn ("jr\t%^", NULL);
- fprintf (asm_out_file, "\t.set\tat\n");
}
else
{
because there is no direct "jx" instruction equivalent to "jalx" to
switch the ISA mode. We only care about cases where the sibling
and normal calls would both be direct. */
- if (mips_use_mips16_mode_p (decl)
+ if (decl
+ && mips_use_mips16_mode_p (decl)
&& const_call_insn_operand (XEXP (DECL_RTL (decl), 0), VOIDmode))
return false;
rtx inc, label, cmp, cmp_result;
/* Load INC with the cache line size (rdhwr INC,$1). */
- inc = gen_reg_rtx (SImode);
- emit_insn (gen_rdhwr (inc, const1_rtx));
+ inc = gen_reg_rtx (Pmode);
+ emit_insn (Pmode == SImode
+ ? gen_rdhwr_synci_step_si (inc)
+ : gen_rdhwr_synci_step_di (inc));
/* Loop back to here. */
label = gen_label_rtx ();
return true;
}
+
+/* Check if MASK and SHIFT are valid in mask-low-and-shift-left
+ operation if MAXLEN is the maxium length of consecutive bits that
+ can make up MASK. MODE is the mode of the operation. See
+ mask_low_and_shift_len for the actual definition. */
+
+bool
+mask_low_and_shift_p (enum machine_mode mode, rtx mask, rtx shift, int maxlen)
+{
+ return IN_RANGE (mask_low_and_shift_len (mode, mask, shift), 1, maxlen);
+}
+
+/* The canonical form of a mask-low-and-shift-left operation is
+ (and (ashift X SHIFT) MASK) where MASK has the lower SHIFT number of bits
+ cleared. Thus we need to shift MASK to the right before checking if it
+ is a valid mask value. MODE is the mode of the operation. If true
+ return the length of the mask, otherwise return -1. */
+
+int
+mask_low_and_shift_len (enum machine_mode mode, rtx mask, rtx shift)
+{
+ HOST_WIDE_INT shval;
+
+ shval = INTVAL (shift) & (GET_MODE_BITSIZE (mode) - 1);
+ return exact_log2 ((UINTVAL (mask) >> shval) + 1);
+}
\f
/* Return true if -msplit-addresses is selected and should be honored.
'#' Print a nop if in a ".set noreorder" block.
'/' Like '#', but do nothing within a delayed-branch sequence.
'?' Print "l" if mips_branch_likely is true
+ '~' Print a nop if mips_branch_likely is true
'.' Print the name of the register with a hard-wired zero (zero or $0).
'@' Print the name of the assembler temporary register (at or $1).
'^' Print the name of the pic call-through register (t9 or $25).
putc ('l', file);
break;
+ case '~':
+ if (mips_branch_likely)
+ fputs ("\n\tnop", file);
+ break;
+
case '.':
fputs (reg_names[GP_REG_FIRST + 0], file);
break;
{
const char *p;
- for (p = "()[]<>*#/?.@^+$|-"; *p; p++)
+ for (p = "()[]<>*#/?~.@^+$|-"; *p; p++)
mips_print_operand_punct[(unsigned char) *p] = true;
}
'X' Print CONST_INT OP in hexadecimal format.
'x' Print the low 16 bits of CONST_INT OP in hexadecimal format.
'd' Print CONST_INT OP in decimal.
+ 'm' Print one less than CONST_INT OP in decimal.
'h' Print the high-part relocation associated with OP, after stripping
any outermost HIGH.
'R' Print the low-part relocation associated with OP.
output_operand_lossage ("invalid use of '%%%c'", letter);
break;
+ case 'm':
+ if (GET_CODE (op) == CONST_INT)
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (op) - 1);
+ else
+ output_operand_lossage ("invalid use of '%%%c'", letter);
+ break;
+
case 'h':
if (code == HIGH)
op = XEXP (op, 0);
/* Implement TARGET_ASM_FUNCTION_RODATA_SECTION.
The complication here is that, with the combination TARGET_ABICALLS
- && !TARGET_GPWORD, jump tables will use absolute addresses, and should
- therefore not be included in the read-only part of a DSO. Handle such
- cases by selecting a normal data section instead of a read-only one.
- The logic apes that in default_function_rodata_section. */
+ && !TARGET_ABSOLUTE_ABICALLS && !TARGET_GPWORD, jump tables will use
+ absolute addresses, and should therefore not be included in the
+ read-only part of a DSO. Handle such cases by selecting a normal
+ data section instead of a read-only one. The logic apes that in
+ default_function_rodata_section. */
static section *
mips_function_rodata_section (tree decl)
{
- if (!TARGET_ABICALLS || TARGET_GPWORD)
+ if (!TARGET_ABICALLS || TARGET_ABSOLUTE_ABICALLS || TARGET_GPWORD)
return default_function_rodata_section (decl);
if (decl && DECL_SECTION_NAME (decl))
case ABI_N32:
return "abiN32";
case ABI_64:
- return "abiN64";
+ return "abi64";
case ABI_EABI:
return TARGET_64BIT ? "eabi64" : "eabi32";
default:
/* If TARGET_ABICALLS, tell GAS to generate -KPIC code. */
if (TARGET_ABICALLS)
- fprintf (asm_out_file, "\t.abicalls\n");
+ {
+ fprintf (asm_out_file, "\t.abicalls\n");
+ if (TARGET_ABICALLS_PIC0)
+ fprintf (asm_out_file, "\t.option\tpic0\n");
+ }
if (flag_verbose_asm)
fprintf (asm_out_file, "\n%s -G value = %d, Arch = %s, ISA = %d\n",
}
/* Return the register that should be used as the global pointer
- within this function. Return 0 if the function doesn't need
- a global pointer. */
+ within this function. Return INVALID_REGNUM if the function
+ doesn't need a global pointer. */
static unsigned int
mips_global_pointer (void)
if (crtl->has_nonlocal_goto)
return GLOBAL_POINTER_REGNUM;
- /* If the gp is never referenced, there's no need to initialize it.
- Note that reload can sometimes introduce constant pool references
- into a function that otherwise didn't need them. For example,
- suppose we have an instruction like:
+ /* There's no need to initialize $gp if it isn't referenced now,
+ and if we can be sure that no new references will be added during
+ or after reload. */
+ if (!df_regs_ever_live_p (GLOBAL_POINTER_REGNUM)
+ && !mips_function_has_gp_insn ())
+ {
+ /* The function doesn't use $gp at the moment. If we're generating
+ -call_nonpic code, no new uses will be introduced during or after
+ reload. */
+ if (TARGET_ABICALLS_PIC0)
+ return INVALID_REGNUM;
- (set (reg:DF R1) (float:DF (reg:SI R2)))
+ /* We need to handle the following implicit gp references:
- If R2 turns out to be constant such as 1, the instruction may have a
- REG_EQUAL note saying that R1 == 1.0. Reload then has the option of
- using this constant if R2 doesn't get allocated to a register.
+ - Reload can sometimes introduce constant pool references
+ into a function that otherwise didn't need them. For example,
+ suppose we have an instruction like:
- In cases like these, reload will have added the constant to the pool
- but no instruction will yet refer to it. */
- if (!df_regs_ever_live_p (GLOBAL_POINTER_REGNUM)
- && !crtl->uses_const_pool
- && !mips16_cfun_returns_in_fpr_p ()
- && !mips_function_has_gp_insn ())
- return 0;
+ (set (reg:DF R1) (float:DF (reg:SI R2)))
+
+ If R2 turns out to be constant such as 1, the instruction may
+ have a REG_EQUAL note saying that R1 == 1.0. Reload then has
+ the option of using this constant if R2 doesn't get allocated
+ to a register.
+
+ In cases like these, reload will have added the constant to the
+ pool but no instruction will yet refer to it.
+
+ - MIPS16 functions that return in FPRs need to call an
+ external libgcc routine. */
+ if (!crtl->uses_const_pool
+ && !mips16_cfun_returns_in_fpr_p ())
+ return INVALID_REGNUM;
+ }
/* We need a global pointer, but perhaps we can use a call-clobbered
register instead of $gp. */
return GLOBAL_POINTER_REGNUM;
}
-/* Return true if the current function must save register REGNO. */
+/* Return true if the current function should treat register REGNO
+ as call-saved. */
static bool
-mips_save_reg_p (unsigned int regno)
+mips_cfun_call_saved_reg_p (unsigned int regno)
{
- /* We need to save $gp if TARGET_CALL_SAVED_GP and if we have not
- chosen a call-clobbered substitute. */
- if (TARGET_CALL_SAVED_GP
- && regno == GLOBAL_POINTER_REGNUM
- && cfun->machine->global_pointer == regno)
+ /* call_insns preserve $28 unless they explicitly say otherwise,
+ so call_really_used_regs[] treats $28 as call-saved. However,
+ we want the ABI property rather than the default call_insn
+ property here. */
+ return (regno == GLOBAL_POINTER_REGNUM
+ ? TARGET_CALL_SAVED_GP
+ : !call_really_used_regs[regno]);
+}
+
+/* Return true if the function body might clobber register REGNO.
+ We know that REGNO is call-saved. */
+
+static bool
+mips_cfun_might_clobber_call_saved_reg_p (unsigned int regno)
+{
+ /* Some functions should be treated as clobbering all call-saved
+ registers. */
+ if (crtl->saves_all_registers)
return true;
- /* Check call-saved registers. */
- if ((crtl->saves_all_registers || df_regs_ever_live_p (regno))
- && !call_really_used_regs[regno])
+ /* DF handles cases where a register is explicitly referenced in
+ the rtl. Incoming values are passed in call-clobbered registers,
+ so we can assume that any live call-saved register is set within
+ the function. */
+ if (df_regs_ever_live_p (regno))
return true;
- /* Save both registers in an FPR pair if either one is used. This is
- needed for the case when MIN_FPRS_PER_FMT == 1, which allows the odd
- register to be used without the even register. */
- if (FP_REG_P (regno)
- && MAX_FPRS_PER_FMT == 2
- && df_regs_ever_live_p (regno + 1)
- && !call_really_used_regs[regno + 1])
+ /* Check for registers that are clobbered by FUNCTION_PROFILER.
+ These clobbers are not explicit in the rtl. */
+ if (crtl->profile && MIPS_SAVE_REG_FOR_PROFILING_P (regno))
return true;
- /* We need to save the old frame pointer before setting up a new one. */
+ /* If we're using a call-saved global pointer, the function's
+ prologue will need to set it up. */
+ if (cfun->machine->global_pointer == regno)
+ return true;
+
+ /* The function's prologue will need to set the frame pointer if
+ frame_pointer_needed. */
if (regno == HARD_FRAME_POINTER_REGNUM && frame_pointer_needed)
return true;
- /* Check for registers that must be saved for FUNCTION_PROFILER. */
- if (crtl->profile && MIPS_SAVE_REG_FOR_PROFILING_P (regno))
+ /* If a MIPS16 function returns a value in FPRs, its epilogue
+ will need to call an external libgcc routine. This yet-to-be
+ generated call_insn will clobber $31. */
+ if (regno == GP_REG_FIRST + 31 && mips16_cfun_returns_in_fpr_p ())
return true;
- /* We need to save the incoming return address if it is ever clobbered
- within the function, if __builtin_eh_return is being used to set a
- different return address, or if a stub is being used to return a
- value in FPRs. */
- if (regno == GP_REG_FIRST + 31
- && (df_regs_ever_live_p (regno)
- || crtl->calls_eh_return
- || mips16_cfun_returns_in_fpr_p ()))
+ return false;
+}
+
+/* Return true if the current function must save register REGNO. */
+
+static bool
+mips_save_reg_p (unsigned int regno)
+{
+ if (mips_cfun_call_saved_reg_p (regno))
+ {
+ if (mips_cfun_might_clobber_call_saved_reg_p (regno))
+ return true;
+
+ /* Save both registers in an FPR pair if either one is used. This is
+ needed for the case when MIN_FPRS_PER_FMT == 1, which allows the odd
+ register to be used without the even register. */
+ if (FP_REG_P (regno)
+ && MAX_FPRS_PER_FMT == 2
+ && mips_cfun_might_clobber_call_saved_reg_p (regno + 1))
+ return true;
+ }
+
+ /* We need to save the incoming return address if __builtin_eh_return
+ is being used to set a different return address. */
+ if (regno == GP_REG_FIRST + 31 && crtl->calls_eh_return)
return true;
return false;
enum mips_loadgp_style
mips_current_loadgp_style (void)
{
- if (!TARGET_USE_GOT || cfun->machine->global_pointer == 0)
+ if (!TARGET_USE_GOT || cfun->machine->global_pointer == INVALID_REGNUM)
return LOADGP_NONE;
if (TARGET_RTP_PIC)
{
gcc_assert (TARGET_ABICALLS && TARGET_OLDABI);
+ if (cfun->machine->global_pointer == INVALID_REGNUM)
+ return;
+
if (TARGET_MIPS16)
{
mips_emit_move (temp, mips_cprestore_slot (temp));
mips_output_cplocal (void)
{
if (!TARGET_EXPLICIT_RELOCS
- && cfun->machine->global_pointer > 0
+ && cfun->machine->global_pointer != INVALID_REGNUM
&& cfun->machine->global_pointer != GLOBAL_POINTER_REGNUM)
output_asm_insn (".cplocal %+", 0);
}
unsigned int nargs;
rtx insn;
- if (cfun->machine->global_pointer > 0)
+ if (cfun->machine->global_pointer != INVALID_REGNUM)
SET_REGNO (pic_offset_table_rtx, cfun->machine->global_pointer);
frame = &cfun->machine->frame;
mips_emit_loadgp ();
/* Initialize the $gp save slot. */
- if (frame->cprestore_size > 0)
+ if (frame->cprestore_size > 0
+ && cfun->machine->global_pointer != INVALID_REGNUM)
{
if (TARGET_MIPS16)
mips_emit_move (mips_cprestore_slot (MIPS_PROLOGUE_TEMP (Pmode)),
MIPS16_PIC_TEMP);
- else
+ else if (TARGET_ABICALLS_PIC2)
emit_insn (gen_cprestore (GEN_INT (frame->args_size)));
+ else
+ emit_move_insn (mips_cprestore_slot (MIPS_PROLOGUE_TEMP (Pmode)),
+ pic_offset_table_rtx);
}
/* If we are profiling, make sure no instructions are scheduled before
return rclass;
}
-/* Implement REGISTER_MOVE_COST. */
+/* RCLASS is a class involved in a REGISTER_MOVE_COST calculation.
+ Return a "canonical" class to represent it in later calculations. */
-int
-mips_register_move_cost (enum machine_mode mode,
- enum reg_class to, enum reg_class from)
+static enum reg_class
+mips_canonicalize_move_class (enum reg_class rclass)
{
- if (TARGET_MIPS16)
+ /* All moves involving accumulator registers have the same cost. */
+ if (reg_class_subset_p (rclass, ACC_REGS))
+ rclass = ACC_REGS;
+
+ /* Likewise promote subclasses of general registers to the most
+ interesting containing class. */
+ if (TARGET_MIPS16 && reg_class_subset_p (rclass, M16_REGS))
+ rclass = M16_REGS;
+ else if (reg_class_subset_p (rclass, GENERAL_REGS))
+ rclass = GENERAL_REGS;
+
+ return rclass;
+}
+
+/* Return the cost of moving a value of mode MODE from a register of
+ class FROM to a GPR. Return 0 for classes that are unions of other
+ classes handled by this function. */
+
+static int
+mips_move_to_gpr_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
+ enum reg_class from)
+{
+ switch (from)
{
- /* ??? We cannot move general registers into HI and LO because
- MIPS16 has no MTHI and MTLO instructions. Make the cost of
- moves in the opposite direction just as high, which stops the
- register allocators from using HI and LO for pseudos. */
- if (reg_class_subset_p (from, GENERAL_REGS)
- && reg_class_subset_p (to, GENERAL_REGS))
- {
- if (reg_class_subset_p (from, M16_REGS)
- || reg_class_subset_p (to, M16_REGS))
- return 2;
- /* Two MOVEs. */
- return 4;
- }
+ case GENERAL_REGS:
+ /* A MIPS16 MOVE instruction, or a non-MIPS16 MOVE macro. */
+ return 2;
+
+ case ACC_REGS:
+ /* MFLO and MFHI. */
+ return 6;
+
+ case FP_REGS:
+ /* MFC1, etc. */
+ return 4;
+
+ case ST_REGS:
+ /* LUI followed by MOVF. */
+ return 4;
+
+ case COP0_REGS:
+ case COP2_REGS:
+ case COP3_REGS:
+ /* This choice of value is historical. */
+ return 5;
+
+ default:
+ return 0;
}
- else if (reg_class_subset_p (from, GENERAL_REGS))
+}
+
+/* Return the cost of moving a value of mode MODE from a GPR to a
+ register of class TO. Return 0 for classes that are unions of
+ other classes handled by this function. */
+
+static int
+mips_move_from_gpr_cost (enum machine_mode mode, enum reg_class to)
+{
+ switch (to)
{
- if (reg_class_subset_p (to, GENERAL_REGS))
- return 2;
- if (reg_class_subset_p (to, FP_REGS))
- return 4;
- if (reg_class_subset_p (to, ALL_COP_AND_GR_REGS))
- return 5;
- if (reg_class_subset_p (to, ACC_REGS))
- return 6;
+ case GENERAL_REGS:
+ /* A MIPS16 MOVE instruction, or a non-MIPS16 MOVE macro. */
+ return 2;
+
+ case ACC_REGS:
+ /* MTLO and MTHI. */
+ return 6;
+
+ case FP_REGS:
+ /* MTC1, etc. */
+ return 4;
+
+ case ST_REGS:
+ /* A secondary reload through an FPR scratch. */
+ return (mips_register_move_cost (mode, GENERAL_REGS, FP_REGS)
+ + mips_register_move_cost (mode, FP_REGS, ST_REGS));
+
+ case COP0_REGS:
+ case COP2_REGS:
+ case COP3_REGS:
+ /* This choice of value is historical. */
+ return 5;
+
+ default:
+ return 0;
}
- else if (reg_class_subset_p (to, GENERAL_REGS))
+}
+
+/* Implement REGISTER_MOVE_COST. Return 0 for classes that are the
+ maximum of the move costs for subclasses; regclass will work out
+ the maximum for us. */
+
+int
+mips_register_move_cost (enum machine_mode mode,
+ enum reg_class from, enum reg_class to)
+{
+ enum reg_class dregs;
+ int cost1, cost2;
+
+ from = mips_canonicalize_move_class (from);
+ to = mips_canonicalize_move_class (to);
+
+ /* Handle moves that can be done without using general-purpose registers. */
+ if (from == FP_REGS)
{
- if (reg_class_subset_p (from, FP_REGS))
- return 4;
- if (reg_class_subset_p (from, ST_REGS))
- /* LUI followed by MOVF. */
+ if (to == FP_REGS && mips_mode_ok_for_mov_fmt_p (mode))
+ /* MOV.FMT. */
return 4;
- if (reg_class_subset_p (from, ALL_COP_AND_GR_REGS))
- return 5;
- if (reg_class_subset_p (from, ACC_REGS))
- return 6;
+ if (to == ST_REGS)
+ /* The sequence generated by mips_expand_fcc_reload. */
+ return 8;
}
- else if (reg_class_subset_p (from, FP_REGS))
+
+ /* Handle cases in which only one class deviates from the ideal. */
+ dregs = TARGET_MIPS16 ? M16_REGS : GENERAL_REGS;
+ if (from == dregs)
+ return mips_move_from_gpr_cost (mode, to);
+ if (to == dregs)
+ return mips_move_to_gpr_cost (mode, from);
+
+ /* Handles cases that require a GPR temporary. */
+ cost1 = mips_move_to_gpr_cost (mode, from);
+ if (cost1 != 0)
{
- if (reg_class_subset_p (to, FP_REGS)
- && mips_mode_ok_for_mov_fmt_p (mode))
- return 4;
- if (reg_class_subset_p (to, ST_REGS))
- /* An expensive sequence. */
- return 8;
+ cost2 = mips_move_from_gpr_cost (mode, to);
+ if (cost2 != 0)
+ return cost1 + cost2;
}
- return 12;
+ return 0;
+}
+
+/* Implement TARGET_IRA_COVER_CLASSES. */
+
+static const enum reg_class *
+mips_ira_cover_classes (void)
+{
+ static const enum reg_class acc_classes[] = {
+ GR_AND_ACC_REGS, FP_REGS, COP0_REGS, COP2_REGS, COP3_REGS,
+ ST_REGS, LIM_REG_CLASSES
+ };
+ static const enum reg_class no_acc_classes[] = {
+ GR_REGS, FP_REGS, COP0_REGS, COP2_REGS, COP3_REGS,
+ ST_REGS, LIM_REG_CLASSES
+ };
+
+ /* Don't allow the register allocators to use LO and HI in MIPS16 mode,
+ which has no MTLO or MTHI instructions. Also, using GR_AND_ACC_REGS
+ as a cover class only works well when we keep per-register costs.
+ Using it when not optimizing can cause us to think accumulators
+ have the same cost as GPRs in cases where GPRs are actually much
+ cheaper. */
+ return TARGET_MIPS16 || !optimize ? no_acc_classes : acc_classes;
}
/* Return the register class required for a secondary register when
if (!reg_class_subset_p (rclass, M16_REGS) && !M16_REG_P (regno))
return M16_REGS;
- /* We can't really copy to HI or LO at all in MIPS16 mode. */
- if (in_p ? reg_classes_intersect_p (rclass, ACC_REGS) : ACC_REG_P (regno))
- return M16_REGS;
-
return NO_REGS;
}
unsigned int length;
rtx taken, not_taken;
+ gcc_assert (LABEL_P (operands[1]));
+
length = get_attr_length (insn);
if (length <= 8)
{
return mips_output_conditional_branch (insn, operands, branch[1], branch[0]);
}
\f
+/* Return the assembly code for __sync_*() loop LOOP. The loop should support
+ both normal and likely branches, using %? and %~ where appropriate. */
+
+const char *
+mips_output_sync_loop (const char *loop)
+{
+ /* Use branch-likely instructions to work around the LL/SC R10000 errata. */
+ mips_branch_likely = TARGET_FIX_R10000;
+ return loop;
+}
+\f
/* Return the assembly code for DIV or DDIV instruction DIVISION, which has
the operands given by OPERANDS. Add in a divide-by-zero check if needed.
but in reality only a maximum of 3 insns can be issued as
floating-point loads and stores also require a slot in the
AGEN pipe. */
- return 4;
+ case PROCESSOR_R10000:
+ /* All R10K Processors are quad-issue (being the first MIPS
+ processors to support this feature). */
+ return 4;
case PROCESSOR_20KC:
case PROCESSOR_R4130:
case PROCESSOR_R5500:
case PROCESSOR_R7000:
case PROCESSOR_R9000:
+ case PROCESSOR_OCTEON:
return 2;
case PROCESSOR_SB1:
if (TUNE_LOONGSON_2EF)
return 4;
+ if (TUNE_OCTEON)
+ return 2;
+
return 0;
}
\f
AVAIL_NON_MIPS16 (dsp_32, !TARGET_64BIT && TARGET_DSP)
AVAIL_NON_MIPS16 (dspr2_32, !TARGET_64BIT && TARGET_DSPR2)
AVAIL_NON_MIPS16 (loongson, TARGET_LOONGSON_VECTORS)
+AVAIL_NON_MIPS16 (cache, TARGET_CACHE_BUILTIN)
/* Construct a mips_builtin_description from the given arguments.
#define CODE_FOR_loongson_pminub CODE_FOR_uminv8qi3
#define CODE_FOR_loongson_pmulhuh CODE_FOR_umulv4hi3_highpart
#define CODE_FOR_loongson_pmulhh CODE_FOR_smulv4hi3_highpart
-#define CODE_FOR_loongson_biadd CODE_FOR_reduc_uplus_v8qi
#define CODE_FOR_loongson_psubw CODE_FOR_subv2si3
#define CODE_FOR_loongson_psubh CODE_FOR_subv4hi3
#define CODE_FOR_loongson_psubb CODE_FOR_subv8qi3
LOONGSON_BUILTIN_SUFFIX (punpcklwd, u, MIPS_UV2SI_FTYPE_UV2SI_UV2SI),
LOONGSON_BUILTIN_SUFFIX (punpcklbh, s, MIPS_V8QI_FTYPE_V8QI_V8QI),
LOONGSON_BUILTIN_SUFFIX (punpcklhw, s, MIPS_V4HI_FTYPE_V4HI_V4HI),
- LOONGSON_BUILTIN_SUFFIX (punpcklwd, s, MIPS_V2SI_FTYPE_V2SI_V2SI)
+ LOONGSON_BUILTIN_SUFFIX (punpcklwd, s, MIPS_V2SI_FTYPE_V2SI_V2SI),
+
+ /* Sundry other built-in functions. */
+ DIRECT_NO_TARGET_BUILTIN (cache, MIPS_VOID_FTYPE_SI_CVPOINTER, cache)
};
/* MODE is a vector mode whose elements have type TYPE. Return the type
return types[mode_index];
}
+/* Return a type for 'const volatile void *'. */
+
+static tree
+mips_build_cvpointer_type (void)
+{
+ static tree cache;
+
+ if (cache == NULL_TREE)
+ cache = build_pointer_type (build_qualified_type
+ (void_type_node,
+ TYPE_QUAL_CONST | TYPE_QUAL_VOLATILE));
+ return cache;
+}
+
/* Source-level argument types. */
#define MIPS_ATYPE_VOID void_type_node
#define MIPS_ATYPE_INT integer_type_node
#define MIPS_ATYPE_POINTER ptr_type_node
+#define MIPS_ATYPE_CVPOINTER mips_build_cvpointer_type ()
/* Standard mode-based argument types. */
#define MIPS_ATYPE_UQI unsigned_intQI_type_node
mips_prepare_builtin_arg (enum insn_code icode,
unsigned int opno, tree exp, unsigned int argno)
{
+ tree arg;
rtx value;
enum machine_mode mode;
- value = expand_normal (CALL_EXPR_ARG (exp, argno));
+ arg = CALL_EXPR_ARG (exp, argno);
+ value = expand_normal (arg);
mode = insn_data[icode].operand[opno].mode;
if (!insn_data[icode].operand[opno].predicate (value, mode))
{
- value = copy_to_mode_reg (mode, value);
+ /* We need to get the mode from ARG for two reasons:
+
+ - to cope with address operands, where MODE is the mode of the
+ memory, rather than of VALUE itself.
+
+ - to cope with special predicates like pmode_register_operand,
+ where MODE is VOIDmode. */
+ value = copy_to_mode_reg (TYPE_MODE (TREE_TYPE (arg)), value);
+
/* Check the predicate again. */
if (!insn_data[icode].operand[opno].predicate (value, mode))
{
opno = 0;
if (has_target_p)
{
- ops[opno] = mips_prepare_builtin_target (icode, opno, target);
+ target = mips_prepare_builtin_target (icode, opno, target);
+ ops[opno] = target;
opno++;
}
static rtx
mips_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
- enum machine_mode mode ATTRIBUTE_UNUSED,
- int ignore ATTRIBUTE_UNUSED)
+ enum machine_mode mode, int ignore)
{
tree fndecl;
unsigned int fcode, avail;
{
error ("built-in function %qs not supported for MIPS16",
IDENTIFIER_POINTER (DECL_NAME (fndecl)));
- return const0_rtx;
+ return ignore ? const0_rtx : CONST0_RTX (mode);
}
switch (d->builtin_type)
{
mips16_emit_constants (pool.first, get_last_insn ());
}
\f
+/* Return true if it is worth r10k_simplify_address's while replacing
+ an address with X. We are looking for constants, and for addresses
+ at a known offset from the incoming stack pointer. */
+
+static bool
+r10k_simplified_address_p (rtx x)
+{
+ if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1)))
+ x = XEXP (x, 0);
+ return x == virtual_incoming_args_rtx || CONSTANT_P (x);
+}
+
+/* X is an expression that appears in INSN. Try to use the UD chains
+ to simplify it, returning the simplified form on success and the
+ original form otherwise. Replace the incoming value of $sp with
+ virtual_incoming_args_rtx (which should never occur in X otherwise). */
+
+static rtx
+r10k_simplify_address (rtx x, rtx insn)
+{
+ rtx newx, op0, op1, set, def_insn, note;
+ df_ref use, def;
+ struct df_link *defs;
+
+ newx = NULL_RTX;
+ if (UNARY_P (x))
+ {
+ op0 = r10k_simplify_address (XEXP (x, 0), insn);
+ if (op0 != XEXP (x, 0))
+ newx = simplify_gen_unary (GET_CODE (x), GET_MODE (x),
+ op0, GET_MODE (XEXP (x, 0)));
+ }
+ else if (BINARY_P (x))
+ {
+ op0 = r10k_simplify_address (XEXP (x, 0), insn);
+ op1 = r10k_simplify_address (XEXP (x, 1), insn);
+ if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
+ newx = simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1);
+ }
+ else if (GET_CODE (x) == LO_SUM)
+ {
+ /* LO_SUMs can be offset from HIGHs, if we know they won't
+ overflow. See mips_classify_address for the rationale behind
+ the lax check. */
+ op0 = r10k_simplify_address (XEXP (x, 0), insn);
+ if (GET_CODE (op0) == HIGH)
+ newx = XEXP (x, 1);
+ }
+ else if (REG_P (x))
+ {
+ /* Uses are recorded by regno_reg_rtx, not X itself. */
+ use = df_find_use (insn, regno_reg_rtx[REGNO (x)]);
+ gcc_assert (use);
+ defs = DF_REF_CHAIN (use);
+
+ /* Require a single definition. */
+ if (defs && defs->next == NULL)
+ {
+ def = defs->ref;
+ if (DF_REF_IS_ARTIFICIAL (def))
+ {
+ /* Replace the incoming value of $sp with
+ virtual_incoming_args_rtx. */
+ if (x == stack_pointer_rtx
+ && DF_REF_BB (def) == ENTRY_BLOCK_PTR)
+ newx = virtual_incoming_args_rtx;
+ }
+ else if (dominated_by_p (CDI_DOMINATORS, DF_REF_BB (use),
+ DF_REF_BB (def)))
+ {
+ /* Make sure that DEF_INSN is a single set of REG. */
+ def_insn = DF_REF_INSN (def);
+ if (NONJUMP_INSN_P (def_insn))
+ {
+ set = single_set (def_insn);
+ if (set && rtx_equal_p (SET_DEST (set), x))
+ {
+ /* Prefer to use notes, since the def-use chains
+ are often shorter. */
+ note = find_reg_equal_equiv_note (def_insn);
+ if (note)
+ newx = XEXP (note, 0);
+ else
+ newx = SET_SRC (set);
+ newx = r10k_simplify_address (newx, def_insn);
+ }
+ }
+ }
+ }
+ }
+ if (newx && r10k_simplified_address_p (newx))
+ return newx;
+ return x;
+}
+
+/* Return true if ADDRESS is known to be an uncached address
+ on R10K systems. */
+
+static bool
+r10k_uncached_address_p (unsigned HOST_WIDE_INT address)
+{
+ unsigned HOST_WIDE_INT upper;
+
+ /* Check for KSEG1. */
+ if (address + 0x60000000 < 0x20000000)
+ return true;
+
+ /* Check for uncached XKPHYS addresses. */
+ if (Pmode == DImode)
+ {
+ upper = (address >> 40) & 0xf9ffff;
+ if (upper == 0x900000 || upper == 0xb80000)
+ return true;
+ }
+ return false;
+}
+
+/* Return true if we can prove that an access to address X in instruction
+ INSN would be safe from R10K speculation. This X is a general
+ expression; it might not be a legitimate address. */
+
+static bool
+r10k_safe_address_p (rtx x, rtx insn)
+{
+ rtx base, offset;
+ HOST_WIDE_INT offset_val;
+
+ x = r10k_simplify_address (x, insn);
+
+ /* Check for references to the stack frame. It doesn't really matter
+ how much of the frame has been allocated at INSN; -mr10k-cache-barrier
+ allows us to assume that accesses to any part of the eventual frame
+ is safe from speculation at any point in the function. */
+ mips_split_plus (x, &base, &offset_val);
+ if (base == virtual_incoming_args_rtx
+ && offset_val >= -cfun->machine->frame.total_size
+ && offset_val < cfun->machine->frame.args_size)
+ return true;
+
+ /* Check for uncached addresses. */
+ if (CONST_INT_P (x))
+ return r10k_uncached_address_p (INTVAL (x));
+
+ /* Check for accesses to a static object. */
+ split_const (x, &base, &offset);
+ return offset_within_block_p (base, INTVAL (offset));
+}
+
+/* Return true if a MEM with MEM_EXPR EXPR and MEM_OFFSET OFFSET is
+ an in-range access to an automatic variable, or to an object with
+ a link-time-constant address. */
+
+static bool
+r10k_safe_mem_expr_p (tree expr, rtx offset)
+{
+ if (expr == NULL_TREE
+ || offset == NULL_RTX
+ || !CONST_INT_P (offset)
+ || INTVAL (offset) < 0
+ || INTVAL (offset) >= int_size_in_bytes (TREE_TYPE (expr)))
+ return false;
+
+ while (TREE_CODE (expr) == COMPONENT_REF)
+ {
+ expr = TREE_OPERAND (expr, 0);
+ if (expr == NULL_TREE)
+ return false;
+ }
+
+ return DECL_P (expr);
+}
+
+/* A for_each_rtx callback for which DATA points to the instruction
+ containing *X. Stop the search if we find a MEM that is not safe
+ from R10K speculation. */
+
+static int
+r10k_needs_protection_p_1 (rtx *loc, void *data)
+{
+ rtx mem;
+
+ mem = *loc;
+ if (!MEM_P (mem))
+ return 0;
+
+ if (r10k_safe_mem_expr_p (MEM_EXPR (mem), MEM_OFFSET (mem)))
+ return -1;
+
+ if (r10k_safe_address_p (XEXP (mem, 0), (rtx) data))
+ return -1;
+
+ return 1;
+}
+
+/* A note_stores callback for which DATA points to an instruction pointer.
+ If *DATA is nonnull, make it null if it X contains a MEM that is not
+ safe from R10K speculation. */
+
+static void
+r10k_needs_protection_p_store (rtx x, const_rtx pat ATTRIBUTE_UNUSED,
+ void *data)
+{
+ rtx *insn_ptr;
+
+ insn_ptr = (rtx *) data;
+ if (*insn_ptr && for_each_rtx (&x, r10k_needs_protection_p_1, *insn_ptr))
+ *insn_ptr = NULL_RTX;
+}
+
+/* A for_each_rtx callback that iterates over the pattern of a CALL_INSN.
+ Return nonzero if the call is not to a declared function. */
+
+static int
+r10k_needs_protection_p_call (rtx *loc, void *data ATTRIBUTE_UNUSED)
+{
+ rtx x;
+
+ x = *loc;
+ if (!MEM_P (x))
+ return 0;
+
+ x = XEXP (x, 0);
+ if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_DECL (x))
+ return -1;
+
+ return 1;
+}
+
+/* Return true if instruction INSN needs to be protected by an R10K
+ cache barrier. */
+
+static bool
+r10k_needs_protection_p (rtx insn)
+{
+ if (CALL_P (insn))
+ return for_each_rtx (&PATTERN (insn), r10k_needs_protection_p_call, NULL);
+
+ if (mips_r10k_cache_barrier == R10K_CACHE_BARRIER_STORE)
+ {
+ note_stores (PATTERN (insn), r10k_needs_protection_p_store, &insn);
+ return insn == NULL_RTX;
+ }
+
+ return for_each_rtx (&PATTERN (insn), r10k_needs_protection_p_1, insn);
+}
+
+/* Return true if BB is only reached by blocks in PROTECTED_BBS and if every
+ edge is unconditional. */
+
+static bool
+r10k_protected_bb_p (basic_block bb, sbitmap protected_bbs)
+{
+ edge_iterator ei;
+ edge e;
+
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ if (!single_succ_p (e->src)
+ || !TEST_BIT (protected_bbs, e->src->index)
+ || (e->flags & EDGE_COMPLEX) != 0)
+ return false;
+ return true;
+}
+
+/* Implement -mr10k-cache-barrier= for the current function. */
+
+static void
+r10k_insert_cache_barriers (void)
+{
+ int *rev_post_order;
+ unsigned int i, n;
+ basic_block bb;
+ sbitmap protected_bbs;
+ rtx insn, end, unprotected_region;
+
+ if (TARGET_MIPS16)
+ {
+ sorry ("%qs does not support MIPS16 code", "-mr10k-cache-barrier");
+ return;
+ }
+
+ /* Restore the BLOCK_FOR_INSN pointers, which are needed by DF. */
+ compute_bb_for_insn ();
+
+ /* Create def-use chains. */
+ df_set_flags (DF_EQ_NOTES);
+ df_chain_add_problem (DF_UD_CHAIN);
+ df_analyze ();
+
+ /* Calculate dominators. */
+ calculate_dominance_info (CDI_DOMINATORS);
+
+ /* Bit X of PROTECTED_BBS is set if the last operation in basic block
+ X is protected by a cache barrier. */
+ protected_bbs = sbitmap_alloc (last_basic_block);
+ sbitmap_zero (protected_bbs);
+
+ /* Iterate over the basic blocks in reverse post-order. */
+ rev_post_order = XNEWVEC (int, last_basic_block);
+ n = pre_and_rev_post_order_compute (NULL, rev_post_order, false);
+ for (i = 0; i < n; i++)
+ {
+ bb = BASIC_BLOCK (rev_post_order[i]);
+
+ /* If this block is only reached by unconditional edges, and if the
+ source of every edge is protected, the beginning of the block is
+ also protected. */
+ if (r10k_protected_bb_p (bb, protected_bbs))
+ unprotected_region = NULL_RTX;
+ else
+ unprotected_region = pc_rtx;
+ end = NEXT_INSN (BB_END (bb));
+
+ /* UNPROTECTED_REGION is:
+
+ - null if we are processing a protected region,
+ - pc_rtx if we are processing an unprotected region but have
+ not yet found the first instruction in it
+ - the first instruction in an unprotected region otherwise. */
+ for (insn = BB_HEAD (bb); insn != end; insn = NEXT_INSN (insn))
+ {
+ if (unprotected_region && INSN_P (insn))
+ {
+ if (recog_memoized (insn) == CODE_FOR_mips_cache)
+ /* This CACHE instruction protects the following code. */
+ unprotected_region = NULL_RTX;
+ else
+ {
+ /* See if INSN is the first instruction in this
+ unprotected region. */
+ if (unprotected_region == pc_rtx)
+ unprotected_region = insn;
+
+ /* See if INSN needs to be protected. If so,
+ we must insert a cache barrier somewhere between
+ PREV_INSN (UNPROTECTED_REGION) and INSN. It isn't
+ clear which position is better performance-wise,
+ but as a tie-breaker, we assume that it is better
+ to allow delay slots to be back-filled where
+ possible, and that it is better not to insert
+ barriers in the middle of already-scheduled code.
+ We therefore insert the barrier at the beginning
+ of the region. */
+ if (r10k_needs_protection_p (insn))
+ {
+ emit_insn_before (gen_r10k_cache_barrier (),
+ unprotected_region);
+ unprotected_region = NULL_RTX;
+ }
+ }
+ }
+
+ if (CALL_P (insn))
+ /* The called function is not required to protect the exit path.
+ The code that follows a call is therefore unprotected. */
+ unprotected_region = pc_rtx;
+ }
+
+ /* Record whether the end of this block is protected. */
+ if (unprotected_region == NULL_RTX)
+ SET_BIT (protected_bbs, bb->index);
+ }
+ XDELETEVEC (rev_post_order);
+
+ sbitmap_free (protected_bbs);
+
+ free_dominance_info (CDI_DOMINATORS);
+
+ df_finish_pass (false);
+
+ free_bb_for_insn ();
+}
+\f
/* A temporary variable used by for_each_rtx callbacks, etc. */
static rtx mips_sim_insn;
orphaned high-part relocation. */
if (mips_orphaned_high_part_p (htab, insn))
delete_insn (insn);
+ /* Also delete cache barriers if the last instruction
+ was an annulled branch. INSN will not be speculatively
+ executed. */
+ else if (recog_memoized (insn) == CODE_FOR_r10k_cache_barrier
+ && last_insn
+ && INSN_ANNULLED_BRANCH_P (SEQ_BEGIN (last_insn)))
+ delete_insn (insn);
else
{
mips_avoid_hazard (last_insn, insn, &hilo_delay,
mips_reorg (void)
{
mips16_lay_out_constants ();
- if (mips_base_delayed_branch)
+ if (mips_r10k_cache_barrier != R10K_CACHE_BARRIER_NONE)
+ r10k_insert_cache_barriers ();
+ if (optimize > 0 && flag_delayed_branch)
dbr_schedule (get_insns ());
mips_reorg_process_insns ();
if (!TARGET_MIPS16
return false;
return true;
+ case OPT_mr10k_cache_barrier_:
+ if (strcmp (arg, "load-store") == 0)
+ mips_r10k_cache_barrier = R10K_CACHE_BARRIER_LOAD_STORE;
+ else if (strcmp (arg, "store") == 0)
+ mips_r10k_cache_barrier = R10K_CACHE_BARRIER_STORE;
+ else if (strcmp (arg, "none") == 0)
+ mips_r10k_cache_barrier = R10K_CACHE_BARRIER_NONE;
+ else
+ return false;
+ return true;
+
default:
return true;
}
target_flags &= ~MASK_ABICALLS;
}
- if (TARGET_ABICALLS)
+ if (TARGET_ABICALLS_PIC2)
/* We need to set flag_pic for executables as well as DSOs
because we may reference symbols that are not defined in
the final executable. (MIPS does not use things like
copy relocs, for example.)
- Also, there is a body of code that uses __PIC__ to distinguish
- between -mabicalls and -mno-abicalls code. */
+ There is a body of code that uses __PIC__ to distinguish
+ between -mabicalls and -mno-abicalls code. The non-__PIC__
+ variant is usually appropriate for TARGET_ABICALLS_PIC0, as
+ long as any indirect jumps use $25. */
flag_pic = 1;
/* -mvr4130-align is a "speed over size" optimization: it usually produces
warning (0, "the %qs architecture does not support paired-single"
" instructions", mips_arch_info->name);
+ if (mips_r10k_cache_barrier != R10K_CACHE_BARRIER_NONE
+ && !TARGET_CACHE_BUILTIN)
+ {
+ error ("%qs requires a target that provides the %qs instruction",
+ "-mr10k-cache-barrier", "cache");
+ mips_r10k_cache_barrier = R10K_CACHE_BARRIER_NONE;
+ }
+
/* If TARGET_DSPR2, enable MASK_DSP. */
if (TARGET_DSPR2)
target_flags |= MASK_DSP;
+ /* .eh_frame addresses should be the same width as a C pointer.
+ Most MIPS ABIs support only one pointer size, so the assembler
+ will usually know exactly how big an .eh_frame address is.
+
+ Unfortunately, this is not true of the 64-bit EABI. The ABI was
+ originally defined to use 64-bit pointers (i.e. it is LP64), and
+ this is still the default mode. However, we also support an n32-like
+ ILP32 mode, which is selected by -mlong32. The problem is that the
+ assembler has traditionally not had an -mlong option, so it has
+ traditionally not known whether we're using the ILP32 or LP64 form.
+
+ As it happens, gas versions up to and including 2.19 use _32-bit_
+ addresses for EABI64 .cfi_* directives. This is wrong for the
+ default LP64 mode, so we can't use the directives by default.
+ Moreover, since gas's current behavior is at odds with gcc's
+ default behavior, it seems unwise to rely on future versions
+ of gas behaving the same way. We therefore avoid using .cfi
+ directives for -mlong32 as well. */
+ if (mips_abi == ABI_EABI && TARGET_64BIT)
+ flag_dwarf2_cfi_asm = 0;
+
mips_init_print_operand_punct ();
/* Set up array to map GCC register number to debug register number.
&& mips_matching_cpu_name_p (mips_arch_info->name, "r4400"))
target_flags |= MASK_FIX_R4400;
+ /* Default to working around R10000 errata only if the processor
+ was selected explicitly. */
+ if ((target_flags_explicit & MASK_FIX_R10000) == 0
+ && mips_matching_cpu_name_p (mips_arch_info->name, "r10000"))
+ target_flags |= MASK_FIX_R10000;
+
+ /* Make sure that branch-likely instructions available when using
+ -mfix-r10000. The instructions are not available if either:
+
+ 1. -mno-branch-likely was passed.
+ 2. The selected ISA does not support branch-likely and
+ the command line does not include -mbranch-likely. */
+ if (TARGET_FIX_R10000
+ && ((target_flags_explicit & MASK_BRANCHLIKELY) == 0
+ ? !ISA_HAS_BRANCHLIKELY
+ : !TARGET_BRANCHLIKELY))
+ sorry ("%qs requires branch-likely instructions", "-mfix-r10000");
+
/* Save base state of options. */
mips_base_target_flags = target_flags;
- mips_base_delayed_branch = flag_delayed_branch;
mips_base_schedule_insns = flag_schedule_insns;
mips_base_reorder_blocks_and_partition = flag_reorder_blocks_and_partition;
mips_base_move_loop_invariants = flag_move_loop_invariants;
Do all CPP-sensitive stuff in non-MIPS16 mode; we'll switch to
MIPS16 mode afterwards if need be. */
mips_set_mips16_mode (false);
-
- /* We call dbr_schedule from within mips_reorg. */
- flag_delayed_branch = 0;
}
/* Swap the register information for registers I and I + 1, which
void
mips_conditional_register_usage (void)
{
- if (!ISA_HAS_DSP)
+
+ if (ISA_HAS_DSP)
+ {
+ /* These DSP control register fields are global. */
+ global_regs[CCDSP_PO_REGNUM] = 1;
+ global_regs[CCDSP_SC_REGNUM] = 1;
+ }
+ else
{
int regno;
#undef TARGET_DWARF_REGISTER_SPAN
#define TARGET_DWARF_REGISTER_SPAN mips_dwarf_register_span
+#undef TARGET_IRA_COVER_CLASSES
+#define TARGET_IRA_COVER_CLASSES mips_ira_cover_classes
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
#include "gt-mips.h"