/* Subroutines used for code generation on IBM RS/6000.
Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
- 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
+ 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
Free Software Foundation, Inc.
Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
1, /* streams */
};
+/* Instruction costs on PPC476 processors. */
+static const
+struct processor_costs ppc476_cost = {
+ COSTS_N_INSNS (4), /* mulsi */
+ COSTS_N_INSNS (4), /* mulsi_const */
+ COSTS_N_INSNS (4), /* mulsi_const9 */
+ COSTS_N_INSNS (4), /* muldi */
+ COSTS_N_INSNS (11), /* divsi */
+ COSTS_N_INSNS (11), /* divdi */
+ COSTS_N_INSNS (6), /* fp */
+ COSTS_N_INSNS (6), /* dmul */
+ COSTS_N_INSNS (19), /* sdiv */
+ COSTS_N_INSNS (33), /* ddiv */
+ 32, /* l1 cache line size */
+ 32, /* l1 cache */
+ 512, /* l2 cache */
+ 1, /* streams */
+};
+
/* Instruction costs on PPC601 processors. */
static const
struct processor_costs ppc601_cost = {
1, /* prefetch streams /*/
};
+/* Instruction costs on PPCE500MC64 processors. */
+static const
+struct processor_costs ppce500mc64_cost = {
+ COSTS_N_INSNS (4), /* mulsi */
+ COSTS_N_INSNS (4), /* mulsi_const */
+ COSTS_N_INSNS (4), /* mulsi_const9 */
+ COSTS_N_INSNS (4), /* muldi */
+ COSTS_N_INSNS (14), /* divsi */
+ COSTS_N_INSNS (14), /* divdi */
+ COSTS_N_INSNS (4), /* fp */
+ COSTS_N_INSNS (10), /* dmul */
+ COSTS_N_INSNS (36), /* sdiv */
+ COSTS_N_INSNS (66), /* ddiv */
+ 64, /* cache line size */
+ 32, /* l1 cache */
+ 128, /* l2 cache */
+ 1, /* prefetch streams /*/
+};
+
/* Instruction costs on POWER4 and POWER5 processors. */
static const
struct processor_costs power4_cost = {
12, /* prefetch streams */
};
+/* Instruction costs on POWER A2 processors. */
+static const
+struct processor_costs ppca2_cost = {
+ COSTS_N_INSNS (16), /* mulsi */
+ COSTS_N_INSNS (16), /* mulsi_const */
+ COSTS_N_INSNS (16), /* mulsi_const9 */
+ COSTS_N_INSNS (16), /* muldi */
+ COSTS_N_INSNS (22), /* divsi */
+ COSTS_N_INSNS (28), /* divdi */
+ COSTS_N_INSNS (3), /* fp */
+ COSTS_N_INSNS (3), /* dmul */
+ COSTS_N_INSNS (59), /* sdiv */
+ COSTS_N_INSNS (72), /* ddiv */
+ 64,
+ 16, /* l1 cache */
+ 2048, /* l2 cache */
+ 16, /* prefetch streams */
+};
+
+\f
+/* Table that classifies rs6000 builtin functions (pure, const, etc.). */
+#undef RS6000_BUILTIN
+#undef RS6000_BUILTIN_EQUATE
+#define RS6000_BUILTIN(NAME, TYPE) TYPE,
+#define RS6000_BUILTIN_EQUATE(NAME, VALUE)
+
+static const enum rs6000_btc builtin_classify[(int)RS6000_BUILTIN_COUNT] =
+{
+#include "rs6000-builtin.def"
+};
+
+#undef RS6000_BUILTIN
+#undef RS6000_BUILTIN_EQUATE
+
\f
static bool rs6000_function_ok_for_sibcall (tree, tree);
static const char *rs6000_invalid_within_doloop (const_rtx);
static void rs6000_assemble_visibility (tree, int);
#endif
static int rs6000_ra_ever_killed (void);
+static bool rs6000_attribute_takes_identifier_p (const_tree);
static tree rs6000_handle_longcall_attribute (tree *, tree, tree, int, bool *);
static tree rs6000_handle_altivec_attribute (tree *, tree, tree, int, bool *);
static bool rs6000_ms_bitfield_layout_p (const_tree);
static rtx rs6000_make_savres_rtx (rs6000_stack_t *, rtx, int,
enum machine_mode, bool, bool, bool);
static bool rs6000_reg_live_or_pic_offset_p (int);
-static tree rs6000_builtin_vectorized_function (unsigned int, tree, tree);
+static tree rs6000_builtin_vectorized_function (tree, tree, tree);
static int rs6000_savres_strategy (rs6000_stack_t *, bool, int, int);
static void rs6000_restore_saved_cr (rtx, int);
static void rs6000_output_function_prologue (FILE *, HOST_WIDE_INT);
static tree rs6000_builtin_mask_for_load (void);
static tree rs6000_builtin_mul_widen_even (tree);
static tree rs6000_builtin_mul_widen_odd (tree);
-static tree rs6000_builtin_conversion (unsigned int, tree);
+static tree rs6000_builtin_conversion (unsigned int, tree, tree);
static tree rs6000_builtin_vec_perm (tree, tree *);
static bool rs6000_builtin_support_vector_misalignment (enum
machine_mode,
static void def_builtin (int, const char *, tree, int);
static bool rs6000_vector_alignment_reachable (const_tree, bool);
static void rs6000_init_builtins (void);
+static tree rs6000_builtin_decl (unsigned, bool);
+
static rtx rs6000_expand_unop_builtin (enum insn_code, tree, rtx);
static rtx rs6000_expand_binop_builtin (enum insn_code, tree, rtx);
static rtx rs6000_expand_ternop_builtin (enum insn_code, tree, rtx);
static rtx rs6000_legitimize_address (rtx, rtx, enum machine_mode);
static rtx rs6000_debug_legitimize_address (rtx, rtx, enum machine_mode);
static rtx rs6000_legitimize_tls_address (rtx, enum tls_model);
-static rtx rs6000_delegitimize_address (rtx);
static void rs6000_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
+static rtx rs6000_delegitimize_address (rtx);
static rtx rs6000_tls_get_addr (void);
static rtx rs6000_got_sym (void);
static int rs6000_tls_symbol_ref_1 (rtx *, void *);
#endif
#ifndef TARGET_PROFILE_KERNEL
#define TARGET_PROFILE_KERNEL 0
-#define SET_PROFILE_KERNEL(N)
-#else
-#define SET_PROFILE_KERNEL(N) TARGET_PROFILE_KERNEL = (N)
#endif
/* The VRSAVE bitmask puts bit %v0 as the most significant bit. */
#define TARGET_ATTRIBUTE_TABLE rs6000_attribute_table
#undef TARGET_SET_DEFAULT_TYPE_ATTRIBUTES
#define TARGET_SET_DEFAULT_TYPE_ATTRIBUTES rs6000_set_default_type_attributes
+#undef TARGET_ATTRIBUTE_TAKES_IDENTIFIER_P
+#define TARGET_ATTRIBUTE_TAKES_IDENTIFIER_P rs6000_attribute_takes_identifier_p
#undef TARGET_ASM_ALIGNED_DI_OP
#define TARGET_ASM_ALIGNED_DI_OP DOUBLE_INT_ASM_OP
#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM rs6000_tls_referenced_p
+#undef TARGET_DELEGITIMIZE_ADDRESS
+#define TARGET_DELEGITIMIZE_ADDRESS rs6000_delegitimize_address
+
#undef TARGET_ASM_FUNCTION_PROLOGUE
#define TARGET_ASM_FUNCTION_PROLOGUE rs6000_output_function_prologue
#undef TARGET_ASM_FUNCTION_EPILOGUE
#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS rs6000_init_builtins
+#undef TARGET_BUILTIN_DECL
+#define TARGET_BUILTIN_DECL rs6000_builtin_decl
#undef TARGET_EXPAND_BUILTIN
#define TARGET_EXPAND_BUILTIN rs6000_expand_builtin
#undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
#define TARGET_USE_BLOCKS_FOR_CONSTANT_P rs6000_use_blocks_for_constant_p
-#undef TARGET_DELEGITIMIZE_ADDRESS
-#define TARGET_DELEGITIMIZE_ADDRESS rs6000_delegitimize_address
-
#undef TARGET_BUILTIN_RECIPROCAL
#define TARGET_BUILTIN_RECIPROCAL rs6000_builtin_reciprocal
POWERPC_BASE_MASK | MASK_SOFT_FLOAT | MASK_MULHW | MASK_DLMZB},
{"464fp", PROCESSOR_PPC440,
POWERPC_BASE_MASK | MASK_MULHW | MASK_DLMZB},
+ {"476", PROCESSOR_PPC476,
+ POWERPC_BASE_MASK | MASK_SOFT_FLOAT | MASK_PPC_GFXOPT | MASK_MFCRF
+ | MASK_POPCNTB | MASK_FPRND | MASK_CMPB | MASK_MULHW | MASK_DLMZB},
+ {"476fp", PROCESSOR_PPC476,
+ POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_MFCRF | MASK_POPCNTB
+ | MASK_FPRND | MASK_CMPB | MASK_MULHW | MASK_DLMZB},
{"505", PROCESSOR_MPCCORE, POWERPC_BASE_MASK},
{"601", PROCESSOR_PPC601,
MASK_POWER | POWERPC_BASE_MASK | MASK_MULTIPLE | MASK_STRING},
/* 8548 has a dummy entry for now. */
{"8548", PROCESSOR_PPC8540, POWERPC_BASE_MASK | MASK_STRICT_ALIGN
| MASK_ISEL},
+ {"a2", PROCESSOR_PPCA2,
+ POWERPC_BASE_MASK | MASK_PPC_GFXOPT | MASK_POWERPC64 | MASK_POPCNTB
+ | MASK_CMPB | MASK_NO_UPDATE },
{"e300c2", PROCESSOR_PPCE300C2, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
{"e300c3", PROCESSOR_PPCE300C3, POWERPC_BASE_MASK},
{"e500mc", PROCESSOR_PPCE500MC, POWERPC_BASE_MASK | MASK_PPC_GFXOPT
| MASK_ISEL},
+ {"e500mc64", PROCESSOR_PPCE500MC64, POWERPC_BASE_MASK | MASK_POWERPC64
+ | MASK_PPC_GFXOPT | MASK_ISEL},
{"860", PROCESSOR_MPCCORE, POWERPC_BASE_MASK | MASK_SOFT_FLOAT},
{"970", PROCESSOR_POWER4,
POWERPC_7400_MASK | MASK_PPC_GPOPT | MASK_MFCRF | MASK_POWERPC64},
| MASK_PPC_GFXOPT | MASK_POWERPC64 | MASK_ALTIVEC
| MASK_MFCRF | MASK_POPCNTB | MASK_FPRND | MASK_MULHW
| MASK_DLMZB | MASK_CMPB | MASK_MFPGPR | MASK_DFP
- | MASK_POPCNTD | MASK_VSX | MASK_ISEL)
+ | MASK_POPCNTD | MASK_VSX | MASK_ISEL | MASK_NO_UPDATE)
};
+ /* Numerous experiment shows that IRA based loop pressure
+ calculation works better for RTL loop invariant motion on targets
+ with enough (>= 32) registers. It is an expensive optimization.
+ So it is on only for peak performance. */
+ if (optimize >= 3)
+ flag_ira_loop_pressure = 1;
+
/* Set the pointer size. */
if (TARGET_64BIT)
{
}
if (rs6000_cpu == PROCESSOR_PPCE300C2 || rs6000_cpu == PROCESSOR_PPCE300C3
- || rs6000_cpu == PROCESSOR_PPCE500MC)
+ || rs6000_cpu == PROCESSOR_PPCE500MC || rs6000_cpu == PROCESSOR_PPCE500MC64)
{
if (TARGET_ALTIVEC)
error ("AltiVec not supported in this target");
rs6000_gen_cell_microcode = !(rs6000_cpu == PROCESSOR_CELL
&& !optimize_size);
- /* If we are optimizing big endian systems for space, use the load/store
- multiple and string instructions unless we are not generating
- Cell microcode. */
- if (BYTES_BIG_ENDIAN && optimize_size && !rs6000_gen_cell_microcode)
+ /* If we are optimizing big endian systems for space and it's OK to
+ use instructions that would be microcoded on the Cell, use the
+ load/store multiple and string instructions. */
+ if (BYTES_BIG_ENDIAN && optimize_size && rs6000_gen_cell_microcode)
target_flags |= ~target_flags_explicit & (MASK_MULTIPLE | MASK_STRING);
/* Don't allow -mmultiple or -mstring on little endian systems
}
}
- /* Add some warnings for VSX. Enable -maltivec unless the user explicitly
- used -mno-altivec */
+ /* Add some warnings for VSX. */
if (TARGET_VSX)
{
const char *msg = NULL;
msg = N_("-mvsx used with little endian code");
else if (TARGET_AVOID_XFORM > 0)
msg = N_("-mvsx needs indexed addressing");
+ else if (!TARGET_ALTIVEC && (target_flags_explicit & MASK_ALTIVEC))
+ {
+ if (target_flags_explicit & MASK_VSX)
+ msg = N_("-mvsx and -mno-altivec are incompatible");
+ else
+ msg = N_("-mno-altivec disables vsx");
+ }
if (msg)
{
warning (0, msg);
target_flags &= ~ MASK_VSX;
}
- else if (TARGET_VSX && !TARGET_ALTIVEC
- && (target_flags_explicit & MASK_ALTIVEC) == 0)
+ else if (TARGET_VSX && !TARGET_ALTIVEC)
target_flags |= MASK_ALTIVEC;
}
SUB3TARGET_OVERRIDE_OPTIONS;
#endif
- if (TARGET_E500 || rs6000_cpu == PROCESSOR_PPCE500MC)
+ if (TARGET_E500 || rs6000_cpu == PROCESSOR_PPCE500MC
+ || rs6000_cpu == PROCESSOR_PPCE500MC64)
{
/* The e500 and e500mc do not have string instructions, and we set
MASK_STRING above when optimizing for size. */
&& rs6000_cpu != PROCESSOR_POWER5
&& rs6000_cpu != PROCESSOR_POWER6
&& rs6000_cpu != PROCESSOR_POWER7
+ && rs6000_cpu != PROCESSOR_PPCA2
&& rs6000_cpu != PROCESSOR_CELL);
rs6000_sched_groups = (rs6000_cpu == PROCESSOR_POWER4
|| rs6000_cpu == PROCESSOR_POWER5
rs6000_align_branch_targets = (rs6000_cpu == PROCESSOR_POWER4
|| rs6000_cpu == PROCESSOR_POWER5
|| rs6000_cpu == PROCESSOR_POWER6
- || rs6000_cpu == PROCESSOR_POWER7);
+ || rs6000_cpu == PROCESSOR_POWER7
+ || rs6000_cpu == PROCESSOR_PPCE500MC
+ || rs6000_cpu == PROCESSOR_PPCE500MC64);
/* Allow debug switches to override the above settings. */
if (TARGET_ALWAYS_HINT > 0)
rs6000_cost = &ppc440_cost;
break;
+ case PROCESSOR_PPC476:
+ rs6000_cost = &ppc476_cost;
+ break;
+
case PROCESSOR_PPC601:
rs6000_cost = &ppc601_cost;
break;
rs6000_cost = &ppce500mc_cost;
break;
+ case PROCESSOR_PPCE500MC64:
+ rs6000_cost = &ppce500mc64_cost;
+ break;
+
case PROCESSOR_POWER4:
case PROCESSOR_POWER5:
rs6000_cost = &power4_cost;
rs6000_cost = &power7_cost;
break;
+ case PROCESSOR_PPCA2:
+ rs6000_cost = &ppca2_cost;
+ break;
+
default:
gcc_unreachable ();
}
/* Implement targetm.vectorize.builtin_conversion.
Returns a decl of a function that implements conversion of an integer vector
- into a floating-point vector, or vice-versa. TYPE is the type of the integer
- side of the conversion.
+ into a floating-point vector, or vice-versa. DEST_TYPE is the
+ destination type and SRC_TYPE the source type of the conversion.
Return NULL_TREE if it is not available. */
static tree
-rs6000_builtin_conversion (unsigned int tcode, tree type)
+rs6000_builtin_conversion (unsigned int tcode, tree dest_type, tree src_type)
{
enum tree_code code = (enum tree_code) tcode;
switch (code)
{
case FIX_TRUNC_EXPR:
- switch (TYPE_MODE (type))
+ switch (TYPE_MODE (dest_type))
{
case V2DImode:
if (!VECTOR_UNIT_VSX_P (V2DFmode))
return NULL_TREE;
- return TYPE_UNSIGNED (type)
+ return TYPE_UNSIGNED (dest_type)
? rs6000_builtin_decls[VSX_BUILTIN_XVCVDPUXDS_UNS]
: rs6000_builtin_decls[VSX_BUILTIN_XVCVDPSXDS];
if (VECTOR_UNIT_NONE_P (V4SImode) || VECTOR_UNIT_NONE_P (V4SFmode))
return NULL_TREE;
- return TYPE_UNSIGNED (type)
+ return TYPE_UNSIGNED (dest_type)
? rs6000_builtin_decls[VECTOR_BUILTIN_FIXUNS_V4SF_V4SI]
: rs6000_builtin_decls[VECTOR_BUILTIN_FIX_V4SF_V4SI];
}
case FLOAT_EXPR:
- switch (TYPE_MODE (type))
+ switch (TYPE_MODE (src_type))
{
case V2DImode:
if (!VECTOR_UNIT_VSX_P (V2DFmode))
return NULL_TREE;
- return TYPE_UNSIGNED (type)
+ return TYPE_UNSIGNED (src_type)
? rs6000_builtin_decls[VSX_BUILTIN_XVCVUXDDP]
: rs6000_builtin_decls[VSX_BUILTIN_XVCVSXDDP];
if (VECTOR_UNIT_NONE_P (V4SImode) || VECTOR_UNIT_NONE_P (V4SFmode))
return NULL_TREE;
- return TYPE_UNSIGNED (type)
+ return TYPE_UNSIGNED (src_type)
? rs6000_builtin_decls[VECTOR_BUILTIN_UNSFLOAT_V4SI_V4SF]
: rs6000_builtin_decls[VECTOR_BUILTIN_FLOAT_V4SI_V4SF];
if it is not available. */
static tree
-rs6000_builtin_vectorized_function (unsigned int fn, tree type_out,
+rs6000_builtin_vectorized_function (tree fndecl, tree type_out,
tree type_in)
{
enum machine_mode in_mode, out_mode;
int in_n, out_n;
+ enum built_in_function fn = DECL_FUNCTION_CODE (fndecl);
if (TREE_CODE (type_out) != VECTOR_TYPE
|| TREE_CODE (type_in) != VECTOR_TYPE
- || !TARGET_VECTORIZE_BUILTINS)
+ || !TARGET_VECTORIZE_BUILTINS
+ || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL)
return NULL_TREE;
out_mode = TYPE_MODE (TREE_TYPE (type_out));
if (low == 0)
return num_insns_constant_wide (high) + 1;
+ else if (high == 0)
+ return num_insns_constant_wide (low) + 1;
else
return (num_insns_constant_wide (high)
+ num_insns_constant_wide (low) + 1);
rtx cc_op0, rtx cc_op1)
{
rtx tmp = gen_reg_rtx (V2SFmode);
- rtx tmp1, max, min, equal_zero;
+ rtx tmp1, max, min;
gcc_assert (TARGET_PAIRED_FLOAT);
gcc_assert (GET_MODE (op0) == GET_MODE (op1));
tmp1 = gen_reg_rtx (V2SFmode);
max = gen_reg_rtx (V2SFmode);
min = gen_reg_rtx (V2SFmode);
- equal_zero = gen_reg_rtx (V2SFmode);
-
+ gen_reg_rtx (V2SFmode);
+
emit_insn (gen_subv2sf3 (tmp, cc_op0, cc_op1));
emit_insn (gen_selv2sf4
(max, tmp, cc_op0, cc_op1, CONST0_RTX (SFmode)));
field = TREE_CHAIN (field);
if (! field)
break;
+ /* A packed field does not contribute any extra alignment. */
+ if (DECL_PACKED (field))
+ return align;
type = TREE_TYPE (field);
while (TREE_CODE (type) == ARRAY_TYPE)
type = TREE_TYPE (type);
return ret;
}
-/* If ORIG_X is a constant pool reference, return its known value,
- otherwise ORIG_X. */
-
-static rtx
-rs6000_delegitimize_address (rtx x)
-{
- rtx orig_x = delegitimize_mem_from_attrs (x);
-
- x = orig_x;
-
- if (!MEM_P (x))
- return orig_x;
-
- x = XEXP (x, 0);
-
- if (legitimate_constant_pool_address_p (x)
- && GET_CODE (XEXP (x, 1)) == CONST
- && GET_CODE (XEXP (XEXP (x, 1), 0)) == MINUS
- && GET_CODE (XEXP (XEXP (XEXP (x, 1), 0), 0)) == SYMBOL_REF
- && constant_pool_expr_p (XEXP (XEXP (XEXP (x, 1), 0), 0))
- && GET_CODE (XEXP (XEXP (XEXP (x, 1), 0), 1)) == SYMBOL_REF
- && toc_relative_expr_p (XEXP (XEXP (XEXP (x, 1), 0), 1)))
- return get_pool_constant (XEXP (XEXP (XEXP (x, 1), 0), 0));
-
- return orig_x;
-}
-
/* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
We need to emit DTP-relative relocations. */
fputs ("@dtprel+0x8000", file);
}
+/* In the name of slightly smaller debug output, and to cater to
+ general assembler lossage, recognize various UNSPEC sequences
+ and turn them back into a direct symbol reference. */
+
+static rtx
+rs6000_delegitimize_address (rtx orig_x)
+{
+ rtx x, y;
+
+ orig_x = delegitimize_mem_from_attrs (orig_x);
+ x = orig_x;
+ if (MEM_P (x))
+ x = XEXP (x, 0);
+
+ if (GET_CODE (x) == PLUS
+ && GET_CODE (XEXP (x, 1)) == CONST
+ && GET_CODE (XEXP (x, 0)) == REG
+ && REGNO (XEXP (x, 0)) == TOC_REGISTER)
+ {
+ y = XEXP (XEXP (x, 1), 0);
+ if (GET_CODE (y) == UNSPEC
+ && XINT (y, 1) == UNSPEC_TOCREL)
+ {
+ y = XVECEXP (y, 0, 0);
+ if (!MEM_P (orig_x))
+ return y;
+ else
+ return replace_equiv_address_nv (orig_x, y);
+ }
+ return orig_x;
+ }
+
+ if (TARGET_MACHO
+ && GET_CODE (orig_x) == LO_SUM
+ && GET_CODE (XEXP (x, 1)) == CONST)
+ {
+ y = XEXP (XEXP (x, 1), 0);
+ if (GET_CODE (y) == UNSPEC
+ && XINT (y, 1) == UNSPEC_MACHOPIC_OFFSET)
+ return XVECEXP (y, 0, 0);
+ }
+
+ return orig_x;
+}
+
/* Construct the SYMBOL_REF for the tls_get_addr function. */
static GTY(()) rtx rs6000_tls_symbol;
else
{
rtx tmp3, mem;
- rtx first, last;
+ rtx last;
tmp1 = gen_reg_rtx (Pmode);
tmp2 = gen_reg_rtx (Pmode);
tmp3 = gen_reg_rtx (Pmode);
mem = gen_const_mem (Pmode, tmp1);
- first = emit_insn (gen_load_toc_v4_PIC_1b (gsym));
+ emit_insn (gen_load_toc_v4_PIC_1b (gsym));
emit_move_insn (tmp1,
gen_rtx_REG (Pmode, LR_REGNO));
emit_move_insn (tmp2, mem);
&& legitimate_indexed_address_p (x, reg_ok_strict))
return 1;
if (GET_CODE (x) == PRE_MODIFY
- && VECTOR_MEM_VSX_P (mode)
- && TARGET_UPDATE
- && legitimate_indexed_address_p (XEXP (x, 1), reg_ok_strict)
- && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
- return 1;
- if (GET_CODE (x) == PRE_MODIFY
&& mode != TImode
&& mode != TFmode
&& mode != TDmode
|| TARGET_POWERPC64
|| ((mode != DFmode && mode != DDmode) || TARGET_E500_DOUBLE))
&& (TARGET_POWERPC64 || mode != DImode)
- && !VECTOR_MEM_ALTIVEC_P (mode)
+ && !VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)
&& !SPE_VECTOR_MODE (mode)
/* Restrict addressing for DI because of our SUBREG hackery. */
&& !(TARGET_E500_DOUBLE
gen_rtx_IOR (DImode, copy_rtx (dest),
GEN_INT (ud1)));
}
+ else if (ud3 == 0 && ud4 == 0)
+ {
+ gcc_assert (ud2 & 0x8000);
+ emit_move_insn (dest, GEN_INT (((ud2 << 16) ^ 0x80000000)
+ - 0x80000000));
+ if (ud1 != 0)
+ emit_move_insn (copy_rtx (dest),
+ gen_rtx_IOR (DImode, copy_rtx (dest),
+ GEN_INT (ud1)));
+ emit_move_insn (copy_rtx (dest),
+ gen_rtx_ZERO_EXTEND (DImode,
+ gen_lowpart (SImode,
+ copy_rtx (dest))));
+ }
else if ((ud4 == 0xffff && (ud3 & 0x8000))
|| (ud4 == 0 && ! (ud3 & 0x8000)))
{
return;
}
- /* Fix up invalid (const (plus (symbol_ref) (reg))) that seems to be created
- in the secondary_reload phase, which evidently overwrites the CONST_INT
- with a register. */
- if (GET_CODE (source) == CONST && GET_CODE (XEXP (source, 0)) == PLUS
- && mode == Pmode)
- {
- rtx add_op0 = XEXP (XEXP (source, 0), 0);
- rtx add_op1 = XEXP (XEXP (source, 0), 1);
-
- if (GET_CODE (add_op0) == SYMBOL_REF && GET_CODE (add_op1) == REG)
- {
- rtx tmp = (can_create_pseudo_p ()) ? gen_reg_rtx (Pmode) : dest;
-
- if (TARGET_DEBUG_ADDR)
- {
- fprintf (stderr, "\nrs6000_emit_move: bad source\n");
- debug_rtx (source);
- }
-
- rs6000_emit_move (tmp, add_op0, Pmode);
- emit_insn (gen_rtx_SET (VOIDmode, dest,
- gen_rtx_PLUS (Pmode, tmp, add_op1)));
- return;
- }
- }
-
if (can_create_pseudo_p () && GET_CODE (operands[0]) == MEM
&& !gpc_reg_operand (operands[1], mode))
operands[1] = force_reg (mode, operands[1]);
{
if ((mask & target_flags) || TARGET_PAIRED_FLOAT)
{
+ tree t;
if (rs6000_builtin_decls[code])
fatal_error ("internal error: builtin function to %s already processed.",
name);
- rs6000_builtin_decls[code] =
+ rs6000_builtin_decls[code] = t =
add_builtin_function (name, type, code, BUILT_IN_MD,
NULL, NULL_TREE);
+
+ gcc_assert (code >= 0 && code < (int)RS6000_BUILTIN_COUNT);
+ switch (builtin_classify[code])
+ {
+ default:
+ gcc_unreachable ();
+
+ /* assume builtin can do anything. */
+ case RS6000_BTC_MISC:
+ break;
+
+ /* const function, function only depends on the inputs. */
+ case RS6000_BTC_CONST:
+ TREE_READONLY (t) = 1;
+ TREE_NOTHROW (t) = 1;
+ break;
+
+ /* pure function, function can read global memory. */
+ case RS6000_BTC_PURE:
+ DECL_PURE_P (t) = 1;
+ TREE_NOTHROW (t) = 1;
+ break;
+
+ /* Function is a math function. If rounding mode is on, then treat
+ the function as not reading global memory, but it can have
+ arbitrary side effects. If it is off, then assume the function is
+ a const function. This mimics the ATTR_MATHFN_FPROUNDING
+ attribute in builtin-attribute.def that is used for the math
+ functions. */
+ case RS6000_BTC_FP_PURE:
+ TREE_NOTHROW (t) = 1;
+ if (flag_rounding_math)
+ {
+ DECL_PURE_P (t) = 1;
+ DECL_IS_NOVOPS (t) = 1;
+ }
+ else
+ TREE_READONLY (t) = 1;
+ break;
+ }
}
}
{ MASK_VSX, CODE_FOR_vsx_xxmrghw_v4si, "__builtin_vsx_xxmrghw_4si", VSX_BUILTIN_XXMRGHW_4SI },
{ MASK_VSX, CODE_FOR_vsx_xxmrglw_v4sf, "__builtin_vsx_xxmrglw", VSX_BUILTIN_XXMRGLW_4SF },
{ MASK_VSX, CODE_FOR_vsx_xxmrglw_v4si, "__builtin_vsx_xxmrglw_4si", VSX_BUILTIN_XXMRGLW_4SI },
+ { MASK_VSX, CODE_FOR_vec_interleave_lowv2df, "__builtin_vsx_mergel_2df", VSX_BUILTIN_VEC_MERGEL_V2DF },
+ { MASK_VSX, CODE_FOR_vec_interleave_lowv2di, "__builtin_vsx_mergel_2di", VSX_BUILTIN_VEC_MERGEL_V2DI },
+ { MASK_VSX, CODE_FOR_vec_interleave_highv2df, "__builtin_vsx_mergeh_2df", VSX_BUILTIN_VEC_MERGEH_V2DF },
+ { MASK_VSX, CODE_FOR_vec_interleave_highv2di, "__builtin_vsx_mergeh_2di", VSX_BUILTIN_VEC_MERGEH_V2DI },
{ MASK_ALTIVEC|MASK_VSX, CODE_FOR_nothing, "__builtin_vec_add", ALTIVEC_BUILTIN_VEC_ADD },
{ MASK_ALTIVEC|MASK_VSX, CODE_FOR_nothing, "__builtin_vec_vaddfp", ALTIVEC_BUILTIN_VEC_VADDFP },
{ MASK_VSX, CODE_FOR_nothing, "__builtin_vec_mul", VSX_BUILTIN_VEC_MUL },
{ MASK_VSX, CODE_FOR_nothing, "__builtin_vec_div", VSX_BUILTIN_VEC_DIV },
- { 0, CODE_FOR_divv2sf3, "__builtin_paired_divv2sf3", PAIRED_BUILTIN_DIVV2SF3 },
- { 0, CODE_FOR_addv2sf3, "__builtin_paired_addv2sf3", PAIRED_BUILTIN_ADDV2SF3 },
- { 0, CODE_FOR_subv2sf3, "__builtin_paired_subv2sf3", PAIRED_BUILTIN_SUBV2SF3 },
- { 0, CODE_FOR_mulv2sf3, "__builtin_paired_mulv2sf3", PAIRED_BUILTIN_MULV2SF3 },
+ { 0, CODE_FOR_paired_divv2sf3, "__builtin_paired_divv2sf3", PAIRED_BUILTIN_DIVV2SF3 },
+ { 0, CODE_FOR_paired_addv2sf3, "__builtin_paired_addv2sf3", PAIRED_BUILTIN_ADDV2SF3 },
+ { 0, CODE_FOR_paired_subv2sf3, "__builtin_paired_subv2sf3", PAIRED_BUILTIN_SUBV2SF3 },
+ { 0, CODE_FOR_paired_mulv2sf3, "__builtin_paired_mulv2sf3", PAIRED_BUILTIN_MULV2SF3 },
{ 0, CODE_FOR_paired_muls0, "__builtin_paired_muls0", PAIRED_BUILTIN_MULS0 },
{ 0, CODE_FOR_paired_muls1, "__builtin_paired_muls1", PAIRED_BUILTIN_MULS1 },
{ 0, CODE_FOR_paired_merge00, "__builtin_paired_merge00", PAIRED_BUILTIN_MERGE00 },
{ 0, CODE_FOR_paired_merge11, "__builtin_paired_merge11", PAIRED_BUILTIN_MERGE11 },
/* Place holder, leave as first spe builtin. */
- { 0, CODE_FOR_spe_evaddw, "__builtin_spe_evaddw", SPE_BUILTIN_EVADDW },
- { 0, CODE_FOR_spe_evand, "__builtin_spe_evand", SPE_BUILTIN_EVAND },
+ { 0, CODE_FOR_addv2si3, "__builtin_spe_evaddw", SPE_BUILTIN_EVADDW },
+ { 0, CODE_FOR_andv2si3, "__builtin_spe_evand", SPE_BUILTIN_EVAND },
{ 0, CODE_FOR_spe_evandc, "__builtin_spe_evandc", SPE_BUILTIN_EVANDC },
- { 0, CODE_FOR_spe_evdivws, "__builtin_spe_evdivws", SPE_BUILTIN_EVDIVWS },
+ { 0, CODE_FOR_divv2si3, "__builtin_spe_evdivws", SPE_BUILTIN_EVDIVWS },
{ 0, CODE_FOR_spe_evdivwu, "__builtin_spe_evdivwu", SPE_BUILTIN_EVDIVWU },
{ 0, CODE_FOR_spe_eveqv, "__builtin_spe_eveqv", SPE_BUILTIN_EVEQV },
{ 0, CODE_FOR_spe_evfsadd, "__builtin_spe_evfsadd", SPE_BUILTIN_EVFSADD },
{ 0, CODE_FOR_spe_evslw, "__builtin_spe_evslw", SPE_BUILTIN_EVSLW },
{ 0, CODE_FOR_spe_evsrws, "__builtin_spe_evsrws", SPE_BUILTIN_EVSRWS },
{ 0, CODE_FOR_spe_evsrwu, "__builtin_spe_evsrwu", SPE_BUILTIN_EVSRWU },
- { 0, CODE_FOR_spe_evsubfw, "__builtin_spe_evsubfw", SPE_BUILTIN_EVSUBFW },
+ { 0, CODE_FOR_subv2si3, "__builtin_spe_evsubfw", SPE_BUILTIN_EVSUBFW },
/* SPE binary operations expecting a 5-bit unsigned literal. */
{ 0, CODE_FOR_spe_evaddiw, "__builtin_spe_evaddiw", SPE_BUILTIN_EVADDIW },
/* The SPE unary builtins must start with SPE_BUILTIN_EVABS and
end with SPE_BUILTIN_EVSUBFUSIAAW. */
- { 0, CODE_FOR_spe_evabs, "__builtin_spe_evabs", SPE_BUILTIN_EVABS },
+ { 0, CODE_FOR_absv2si2, "__builtin_spe_evabs", SPE_BUILTIN_EVABS },
{ 0, CODE_FOR_spe_evaddsmiaaw, "__builtin_spe_evaddsmiaaw", SPE_BUILTIN_EVADDSMIAAW },
{ 0, CODE_FOR_spe_evaddssiaaw, "__builtin_spe_evaddssiaaw", SPE_BUILTIN_EVADDSSIAAW },
{ 0, CODE_FOR_spe_evaddumiaaw, "__builtin_spe_evaddumiaaw", SPE_BUILTIN_EVADDUMIAAW },
/* Place-holder. Leave as last unary SPE builtin. */
{ 0, CODE_FOR_spe_evsubfusiaaw, "__builtin_spe_evsubfusiaaw", SPE_BUILTIN_EVSUBFUSIAAW },
- { 0, CODE_FOR_absv2sf2, "__builtin_paired_absv2sf2", PAIRED_BUILTIN_ABSV2SF2 },
+ { 0, CODE_FOR_paired_absv2sf2, "__builtin_paired_absv2sf2", PAIRED_BUILTIN_ABSV2SF2 },
{ 0, CODE_FOR_nabsv2sf2, "__builtin_paired_nabsv2sf2", PAIRED_BUILTIN_NABSV2SF2 },
- { 0, CODE_FOR_negv2sf2, "__builtin_paired_negv2sf2", PAIRED_BUILTIN_NEGV2SF2 },
+ { 0, CODE_FOR_paired_negv2sf2, "__builtin_paired_negv2sf2", PAIRED_BUILTIN_NEGV2SF2 },
{ 0, CODE_FOR_sqrtv2sf2, "__builtin_paired_sqrtv2sf2", PAIRED_BUILTIN_SQRTV2SF2 },
{ 0, CODE_FOR_resv2sf2, "__builtin_paired_resv2sf2", PAIRED_BUILTIN_RESV2SF2 }
};
tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
tree arg0, arg1, arg2;
- enum machine_mode mode0, mode1, mode2;
+ enum machine_mode mode0, mode1;
rtx pat, op0, op1, op2;
const struct builtin_description *d;
size_t i;
op2 = expand_normal (arg2);
mode0 = insn_data[d->icode].operand[0].mode;
mode1 = insn_data[d->icode].operand[1].mode;
- mode2 = insn_data[d->icode].operand[2].mode;
/* Invalid arguments, bail out before generating bad rtl. */
if (arg0 == error_mark_node
rs6000_init_builtins (void)
{
tree tdecl;
-
+
V2SI_type_node = build_vector_type (intSI_type_node, 2);
V2SF_type_node = build_vector_type (float_type_node, 2);
V2DI_type_node = build_vector_type (intDI_type_node, 2);
#endif
}
+/* Returns the rs6000 builtin decl for CODE. */
+
+static tree
+rs6000_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED)
+{
+ if (code >= RS6000_BUILTIN_COUNT)
+ return error_mark_node;
+
+ return rs6000_builtin_decls[code];
+}
+
/* Search through a set of builtins and enable the mask bits.
DESC is an array of builtins.
SIZE is the total number of builtins.
/* Emit the RTL for an sCOND pattern. */
void
+rs6000_emit_sISEL (enum machine_mode mode, rtx operands[])
+{
+ rtx condition_rtx;
+ enum machine_mode op_mode;
+ enum rtx_code cond_code;
+ rtx result = operands[0];
+
+ condition_rtx = rs6000_generate_compare (operands[1], mode);
+ cond_code = GET_CODE (condition_rtx);
+
+ op_mode = GET_MODE (XEXP (operands[1], 0));
+ if (op_mode == VOIDmode)
+ op_mode = GET_MODE (XEXP (operands[1], 1));
+
+ if (TARGET_POWERPC64 && GET_MODE (result) == DImode)
+ {
+ PUT_MODE (condition_rtx, DImode);
+ if (cond_code == GEU || cond_code == GTU || cond_code == LEU
+ || cond_code == LTU)
+ emit_insn (gen_isel_unsigned_di (result, condition_rtx,
+ force_reg (DImode, const1_rtx),
+ force_reg (DImode, const0_rtx),
+ XEXP (condition_rtx, 0)));
+ else
+ emit_insn (gen_isel_signed_di (result, condition_rtx,
+ force_reg (DImode, const1_rtx),
+ force_reg (DImode, const0_rtx),
+ XEXP (condition_rtx, 0)));
+ }
+ else
+ {
+ PUT_MODE (condition_rtx, SImode);
+ if (cond_code == GEU || cond_code == GTU || cond_code == LEU
+ || cond_code == LTU)
+ emit_insn (gen_isel_unsigned_si (result, condition_rtx,
+ force_reg (SImode, const1_rtx),
+ force_reg (SImode, const0_rtx),
+ XEXP (condition_rtx, 0)));
+ else
+ emit_insn (gen_isel_signed_si (result, condition_rtx,
+ force_reg (SImode, const1_rtx),
+ force_reg (SImode, const0_rtx),
+ XEXP (condition_rtx, 0)));
+ }
+}
+
+void
rs6000_emit_sCOND (enum machine_mode mode, rtx operands[])
{
rtx condition_rtx;
enum rtx_code cond_code;
rtx result = operands[0];
+ if (TARGET_ISEL && (mode == SImode || mode == DImode))
+ {
+ rs6000_emit_sISEL (mode, operands);
+ return;
+ }
+
condition_rtx = rs6000_generate_compare (operands[1], mode);
cond_code = GET_CODE (condition_rtx);
rs6000_emit_int_cmove (rtx dest, rtx op, rtx true_cond, rtx false_cond)
{
rtx condition_rtx, cr;
- enum machine_mode mode = GET_MODE (XEXP (op, 0));
+ enum machine_mode mode = GET_MODE (dest);
if (mode != SImode && (!TARGET_POWERPC64 || mode != DImode))
return 0;
/* We still have to do the compare, because isel doesn't do a
compare, it just looks at the CRx bits set by a previous compare
instruction. */
- condition_rtx = rs6000_generate_compare (op, SImode);
+ condition_rtx = rs6000_generate_compare (op, mode);
cr = XEXP (condition_rtx, 0);
if (mode == SImode)
int i;
int j = -1;
bool used_update = false;
+ rtx restore_basereg = NULL_RTX;
if (MEM_P (src) && INT_REGNO_P (reg))
{
}
else if (! rs6000_offsettable_memref_p (src))
{
- rtx basereg;
- basereg = gen_rtx_REG (Pmode, reg);
- emit_insn (gen_rtx_SET (VOIDmode, basereg, XEXP (src, 0)));
- src = replace_equiv_address (src, basereg);
+ if (GET_CODE (XEXP (src, 0)) == PRE_MODIFY)
+ {
+ rtx basereg = XEXP (XEXP (src, 0), 0);
+ if (TARGET_UPDATE)
+ {
+ rtx ndst = simplify_gen_subreg (reg_mode, dst, mode, 0);
+ emit_insn (gen_rtx_SET (VOIDmode, ndst,
+ gen_rtx_MEM (reg_mode, XEXP (src, 0))));
+ used_update = true;
+ }
+ else
+ emit_insn (gen_rtx_SET (VOIDmode, basereg,
+ XEXP (XEXP (src, 0), 1)));
+ src = replace_equiv_address (src, basereg);
+ }
+ else
+ {
+ rtx basereg = gen_rtx_REG (Pmode, reg);
+ emit_insn (gen_rtx_SET (VOIDmode, basereg, XEXP (src, 0)));
+ src = replace_equiv_address (src, basereg);
+ }
}
breg = XEXP (src, 0);
&& REGNO (breg) < REGNO (dst) + nregs)
j = REGNO (breg) - REGNO (dst);
}
-
- if (GET_CODE (dst) == MEM && INT_REGNO_P (reg))
+ else if (MEM_P (dst) && INT_REGNO_P (reg))
{
rtx breg;
emit_insn (gen_add3_insn (breg, breg, delta_rtx));
dst = replace_equiv_address (dst, breg);
}
- else
+ else if (!rs6000_offsettable_memref_p (dst)
+ && GET_CODE (XEXP (dst, 0)) != LO_SUM)
+ {
+ if (GET_CODE (XEXP (dst, 0)) == PRE_MODIFY)
+ {
+ rtx basereg = XEXP (XEXP (dst, 0), 0);
+ if (TARGET_UPDATE)
+ {
+ rtx nsrc = simplify_gen_subreg (reg_mode, src, mode, 0);
+ emit_insn (gen_rtx_SET (VOIDmode,
+ gen_rtx_MEM (reg_mode, XEXP (dst, 0)), nsrc));
+ used_update = true;
+ }
+ else
+ emit_insn (gen_rtx_SET (VOIDmode, basereg,
+ XEXP (XEXP (dst, 0), 1)));
+ dst = replace_equiv_address (dst, basereg);
+ }
+ else
+ {
+ rtx basereg = XEXP (XEXP (dst, 0), 0);
+ rtx offsetreg = XEXP (XEXP (dst, 0), 1);
+ gcc_assert (GET_CODE (XEXP (dst, 0)) == PLUS
+ && REG_P (basereg)
+ && REG_P (offsetreg)
+ && REGNO (basereg) != REGNO (offsetreg));
+ if (REGNO (basereg) == 0)
+ {
+ rtx tmp = offsetreg;
+ offsetreg = basereg;
+ basereg = tmp;
+ }
+ emit_insn (gen_add3_insn (basereg, basereg, offsetreg));
+ restore_basereg = gen_sub3_insn (basereg, basereg, offsetreg);
+ dst = replace_equiv_address (dst, basereg);
+ }
+ }
+ else if (GET_CODE (XEXP (dst, 0)) != LO_SUM)
gcc_assert (rs6000_offsettable_memref_p (dst));
}
simplify_gen_subreg (reg_mode, src, mode,
j * reg_mode_size)));
}
+ if (restore_basereg != NULL_RTX)
+ emit_insn (restore_basereg);
}
}
do_compare_rtx_and_jump (opcode, tocompare, EQ, 1,
SImode, NULL_RTX, NULL_RTX,
- no_toc_save_needed);
+ no_toc_save_needed, -1);
mem = gen_frame_mem (Pmode,
gen_rtx_PLUS (Pmode, stack_top,
use language_string.
C is 0. Fortran is 1. Pascal is 2. Ada is 3. C++ is 9.
Java is 13. Objective-C is 14. Objective-C++ isn't assigned
- a number, so for now use 9. */
- if (! strcmp (language_string, "GNU C"))
+ a number, so for now use 9. LTO isn't assigned a number either,
+ so for now use 0. */
+ if (! strcmp (language_string, "GNU C")
+ || ! strcmp (language_string, "GNU GIMPLE"))
i = 0;
else if (! strcmp (language_string, "GNU F77")
|| ! strcmp (language_string, "GNU Fortran"))
instructions to issue in this cycle. */
static int
-rs6000_variable_issue (FILE *stream ATTRIBUTE_UNUSED,
- int verbose ATTRIBUTE_UNUSED,
- rtx insn, int more)
+rs6000_variable_issue_1 (rtx insn, int more)
{
last_scheduled_insn = insn;
if (GET_CODE (PATTERN (insn)) == USE
return cached_can_issue_more;
}
+static int
+rs6000_variable_issue (FILE *stream, int verbose, rtx insn, int more)
+{
+ int r = rs6000_variable_issue_1 (insn, more);
+ if (verbose)
+ fprintf (stream, "// rs6000_variable_issue (more = %d) = %d\n", more, r);
+ return r;
+}
+
/* Adjust the cost of a scheduling dependency. Return the new cost of
a dependency LINK or INSN on DEP_INSN. COST is the current cost. */
static int
rs6000_issue_rate (void)
{
- /* Use issue rate of 1 for first scheduling pass to decrease degradation. */
- if (!reload_completed)
+ /* Unless scheduling for register pressure, use issue rate of 1 for
+ first scheduling pass to decrease degradation. */
+ if (!reload_completed && !flag_sched_pressure)
return 1;
switch (rs6000_cpu_attr) {
case CPU_PPCE300C2:
case CPU_PPCE300C3:
case CPU_PPCE500MC:
+ case CPU_PPCE500MC64:
return 2;
case CPU_RIOS2:
+ case CPU_PPC476:
case CPU_PPC604:
case CPU_PPC604E:
case CPU_PPC620:
}
\f
+/* Returns TRUE iff the target attribute indicated by ATTR_ID takes a plain
+ identifier as an argument, so the front end shouldn't look it up. */
+
+static bool
+rs6000_attribute_takes_identifier_p (const_tree attr_id)
+{
+ return is_attribute_p ("altivec", attr_id);
+}
+
/* Handle the "altivec" attribute. The attribute may have
arguments as follows:
{
if (XEXP (x, 1) == const0_rtx)
{
- *total = COSTS_N_INSNS (2);
+ if (TARGET_ISEL && !TARGET_MFCRF)
+ *total = COSTS_N_INSNS (8);
+ else
+ *total = COSTS_N_INSNS (2);
return true;
}
else if (mode == Pmode)
case UNORDERED:
if (outer_code == SET && (XEXP (x, 1) == const0_rtx))
{
- *total = COSTS_N_INSNS (2);
+ if (TARGET_ISEL && !TARGET_MFCRF)
+ *total = COSTS_N_INSNS (8);
+ else
+ *total = COSTS_N_INSNS (2);
return true;
}
/* CC COMPARE. */
OSDN Git Service
