/* Subroutines used for code generation on the DEC Alpha.
Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
- 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
+ 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
This file is part of GCC.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING. If not, write to
-the Free Software Foundation, 59 Temple Place - Suite 330,
-Boston, MA 02111-1307, USA. */
+the Free Software Foundation, 51 Franklin Street, Fifth Floor,
+Boston, MA 02110-1301, USA. */
#include "config.h"
#include <splay-tree.h>
#include "cfglayout.h"
#include "tree-gimple.h"
+#include "tree-flow.h"
+#include "tree-stdarg.h"
/* Specify which cpu to schedule for. */
+enum processor_type alpha_tune;
+/* Which cpu we're generating code for. */
enum processor_type alpha_cpu;
-static const char * const alpha_cpu_name[] =
+
+static const char * const alpha_cpu_name[] =
{
"ev4", "ev5", "ev6"
};
enum alpha_fp_trap_mode alpha_fptm;
-/* Specify bit size of immediate TLS offsets. */
-
-int alpha_tls_size = 32;
-
-/* Strings decoded into the above options. */
-
-const char *alpha_cpu_string; /* -mcpu= */
-const char *alpha_tune_string; /* -mtune= */
-const char *alpha_tp_string; /* -mtrap-precision=[p|s|i] */
-const char *alpha_fprm_string; /* -mfp-rounding-mode=[n|m|c|d] */
-const char *alpha_fptm_string; /* -mfp-trap-mode=[n|u|su|sui] */
-const char *alpha_mlat_string; /* -mmemory-latency= */
-const char *alpha_tls_size_string; /* -mtls-size=[16|32|64] */
-
/* Save information from a "cmpxx" operation until the branch or scc is
emitted. */
/* Declarations of static functions. */
static struct machine_function *alpha_init_machine_status (void);
-static rtx alpha_emit_xfloating_compare (enum rtx_code, rtx, rtx);
+static rtx alpha_emit_xfloating_compare (enum rtx_code *, rtx, rtx);
#if TARGET_ABI_OPEN_VMS
static void alpha_write_linkage (FILE *, const char *, tree);
static void unicosmk_output_ssib (FILE *, const char *);
static int unicosmk_need_dex (rtx);
\f
+/* Implement TARGET_HANDLE_OPTION. */
+
+static bool
+alpha_handle_option (size_t code, const char *arg, int value)
+{
+ switch (code)
+ {
+ case OPT_mfp_regs:
+ if (value == 0)
+ target_flags |= MASK_SOFT_FP;
+ break;
+
+ case OPT_mieee:
+ case OPT_mieee_with_inexact:
+ target_flags |= MASK_IEEE_CONFORMANT;
+ break;
+
+ case OPT_mtls_size_:
+ if (value != 16 && value != 32 && value != 64)
+ error ("bad value %qs for -mtls-size switch", arg);
+ break;
+ }
+
+ return true;
+}
+
+#ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
+/* Implement TARGET_MANGLE_FUNDAMENTAL_TYPE. */
+
+static const char *
+alpha_mangle_fundamental_type (tree type)
+{
+ if (TYPE_MAIN_VARIANT (type) == long_double_type_node
+ && TARGET_LONG_DOUBLE_128)
+ return "g";
+
+ /* For all other types, use normal C++ mangling. */
+ return NULL;
+}
+#endif
+
/* Parse target option strings. */
void
override_options (void)
{
- int i;
static const struct cpu_table {
const char *const name;
const enum processor_type processor;
const int flags;
} cpu_table[] = {
-#define EV5_MASK (MASK_CPU_EV5)
-#define EV6_MASK (MASK_CPU_EV6|MASK_BWX|MASK_MAX|MASK_FIX)
{ "ev4", PROCESSOR_EV4, 0 },
{ "ev45", PROCESSOR_EV4, 0 },
{ "21064", PROCESSOR_EV4, 0 },
- { "ev5", PROCESSOR_EV5, EV5_MASK },
- { "21164", PROCESSOR_EV5, EV5_MASK },
- { "ev56", PROCESSOR_EV5, EV5_MASK|MASK_BWX },
- { "21164a", PROCESSOR_EV5, EV5_MASK|MASK_BWX },
- { "pca56", PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
- { "21164PC",PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
- { "21164pc",PROCESSOR_EV5, EV5_MASK|MASK_BWX|MASK_MAX },
- { "ev6", PROCESSOR_EV6, EV6_MASK },
- { "21264", PROCESSOR_EV6, EV6_MASK },
- { "ev67", PROCESSOR_EV6, EV6_MASK|MASK_CIX },
- { "21264a", PROCESSOR_EV6, EV6_MASK|MASK_CIX },
+ { "ev5", PROCESSOR_EV5, 0 },
+ { "21164", PROCESSOR_EV5, 0 },
+ { "ev56", PROCESSOR_EV5, MASK_BWX },
+ { "21164a", PROCESSOR_EV5, MASK_BWX },
+ { "pca56", PROCESSOR_EV5, MASK_BWX|MASK_MAX },
+ { "21164PC",PROCESSOR_EV5, MASK_BWX|MASK_MAX },
+ { "21164pc",PROCESSOR_EV5, MASK_BWX|MASK_MAX },
+ { "ev6", PROCESSOR_EV6, MASK_BWX|MASK_MAX|MASK_FIX },
+ { "21264", PROCESSOR_EV6, MASK_BWX|MASK_MAX|MASK_FIX },
+ { "ev67", PROCESSOR_EV6, MASK_BWX|MASK_MAX|MASK_FIX|MASK_CIX },
+ { "21264a", PROCESSOR_EV6, MASK_BWX|MASK_MAX|MASK_FIX|MASK_CIX },
{ 0, 0, 0 }
};
-
+
+ int i;
+
/* Unicos/Mk doesn't have shared libraries. */
if (TARGET_ABI_UNICOSMK && flag_pic)
{
- warning ("-f%s ignored for Unicos/Mk (not supported)",
+ warning (0, "-f%s ignored for Unicos/Mk (not supported)",
(flag_pic > 1) ? "PIC" : "pic");
flag_pic = 0;
}
- /* On Unicos/Mk, the native compiler consistently generates /d suffices for
+ /* On Unicos/Mk, the native compiler consistently generates /d suffices for
floating-point instructions. Make that the default for this target. */
if (TARGET_ABI_UNICOSMK)
alpha_fprm = ALPHA_FPRM_DYN;
alpha_tp = ALPHA_TP_PROG;
alpha_fptm = ALPHA_FPTM_N;
- /* We cannot use su and sui qualifiers for conversion instructions on
+ /* We cannot use su and sui qualifiers for conversion instructions on
Unicos/Mk. I'm not sure if this is due to assembler or hardware
limitations. Right now, we issue a warning if -mieee is specified
and then ignore it; eventually, we should either get it right or
if (TARGET_IEEE)
{
if (TARGET_ABI_UNICOSMK)
- warning ("-mieee not supported on Unicos/Mk");
+ warning (0, "-mieee not supported on Unicos/Mk");
else
{
alpha_tp = ALPHA_TP_INSN;
if (TARGET_IEEE_WITH_INEXACT)
{
if (TARGET_ABI_UNICOSMK)
- warning ("-mieee-with-inexact not supported on Unicos/Mk");
+ warning (0, "-mieee-with-inexact not supported on Unicos/Mk");
else
{
alpha_tp = ALPHA_TP_INSN;
else if (! strcmp (alpha_tp_string, "i"))
alpha_tp = ALPHA_TP_INSN;
else
- error ("bad value `%s' for -mtrap-precision switch", alpha_tp_string);
+ error ("bad value %qs for -mtrap-precision switch", alpha_tp_string);
}
if (alpha_fprm_string)
else if (! strcmp (alpha_fprm_string,"d"))
alpha_fprm = ALPHA_FPRM_DYN;
else
- error ("bad value `%s' for -mfp-rounding-mode switch",
+ error ("bad value %qs for -mfp-rounding-mode switch",
alpha_fprm_string);
}
else if (strcmp (alpha_fptm_string, "sui") == 0)
alpha_fptm = ALPHA_FPTM_SUI;
else
- error ("bad value `%s' for -mfp-trap-mode switch", alpha_fptm_string);
- }
-
- if (alpha_tls_size_string)
- {
- if (strcmp (alpha_tls_size_string, "16") == 0)
- alpha_tls_size = 16;
- else if (strcmp (alpha_tls_size_string, "32") == 0)
- alpha_tls_size = 32;
- else if (strcmp (alpha_tls_size_string, "64") == 0)
- alpha_tls_size = 64;
- else
- error ("bad value `%s' for -mtls-size switch", alpha_tls_size_string);
+ error ("bad value %qs for -mfp-trap-mode switch", alpha_fptm_string);
}
- alpha_cpu
- = TARGET_CPU_DEFAULT & MASK_CPU_EV6 ? PROCESSOR_EV6
- : (TARGET_CPU_DEFAULT & MASK_CPU_EV5 ? PROCESSOR_EV5 : PROCESSOR_EV4);
-
if (alpha_cpu_string)
{
for (i = 0; cpu_table [i].name; i++)
if (! strcmp (alpha_cpu_string, cpu_table [i].name))
{
- alpha_cpu = cpu_table [i].processor;
- target_flags &= ~ (MASK_BWX | MASK_MAX | MASK_FIX | MASK_CIX
- | MASK_CPU_EV5 | MASK_CPU_EV6);
+ alpha_tune = alpha_cpu = cpu_table [i].processor;
+ target_flags &= ~ (MASK_BWX | MASK_MAX | MASK_FIX | MASK_CIX);
target_flags |= cpu_table [i].flags;
break;
}
if (! cpu_table [i].name)
- error ("bad value `%s' for -mcpu switch", alpha_cpu_string);
+ error ("bad value %qs for -mcpu switch", alpha_cpu_string);
}
if (alpha_tune_string)
for (i = 0; cpu_table [i].name; i++)
if (! strcmp (alpha_tune_string, cpu_table [i].name))
{
- alpha_cpu = cpu_table [i].processor;
+ alpha_tune = cpu_table [i].processor;
break;
}
if (! cpu_table [i].name)
- error ("bad value `%s' for -mcpu switch", alpha_tune_string);
+ error ("bad value %qs for -mcpu switch", alpha_tune_string);
}
/* Do some sanity checks on the above options. */
if (TARGET_ABI_UNICOSMK && alpha_fptm != ALPHA_FPTM_N)
{
- warning ("trap mode not supported on Unicos/Mk");
+ warning (0, "trap mode not supported on Unicos/Mk");
alpha_fptm = ALPHA_FPTM_N;
}
if ((alpha_fptm == ALPHA_FPTM_SU || alpha_fptm == ALPHA_FPTM_SUI)
- && alpha_tp != ALPHA_TP_INSN && ! TARGET_CPU_EV6)
+ && alpha_tp != ALPHA_TP_INSN && alpha_cpu != PROCESSOR_EV6)
{
- warning ("fp software completion requires -mtrap-precision=i");
+ warning (0, "fp software completion requires -mtrap-precision=i");
alpha_tp = ALPHA_TP_INSN;
}
- if (TARGET_CPU_EV6)
+ if (alpha_cpu == PROCESSOR_EV6)
{
/* Except for EV6 pass 1 (not released), we always have precise
arithmetic traps. Which means we can do software completion
{
if (alpha_fprm == ALPHA_FPRM_MINF || alpha_fprm == ALPHA_FPRM_DYN)
{
- warning ("rounding mode not supported for VAX floats");
+ warning (0, "rounding mode not supported for VAX floats");
alpha_fprm = ALPHA_FPRM_NORM;
}
if (alpha_fptm == ALPHA_FPTM_SUI)
{
- warning ("trap mode not supported for VAX floats");
+ warning (0, "trap mode not supported for VAX floats");
alpha_fptm = ALPHA_FPTM_SU;
}
if (target_flags_explicit & MASK_LONG_DOUBLE_128)
- warning ("128-bit long double not supported for VAX floats");
+ warning (0, "128-bit long double not supported for VAX floats");
target_flags &= ~MASK_LONG_DOUBLE_128;
}
&& ISDIGIT ((unsigned char)alpha_mlat_string[1])
&& alpha_mlat_string[2] == '\0')
{
- static int const cache_latency[][4] =
+ static int const cache_latency[][4] =
{
{ 3, 30, -1 }, /* ev4 -- Bcache is a guess */
{ 2, 12, 38 }, /* ev5 -- Bcache from PC164 LMbench numbers */
};
lat = alpha_mlat_string[1] - '0';
- if (lat <= 0 || lat > 3 || cache_latency[alpha_cpu][lat-1] == -1)
+ if (lat <= 0 || lat > 3 || cache_latency[alpha_tune][lat-1] == -1)
{
- warning ("L%d cache latency unknown for %s",
- lat, alpha_cpu_name[alpha_cpu]);
+ warning (0, "L%d cache latency unknown for %s",
+ lat, alpha_cpu_name[alpha_tune]);
lat = 3;
}
else
- lat = cache_latency[alpha_cpu][lat-1];
+ lat = cache_latency[alpha_tune][lat-1];
}
else if (! strcmp (alpha_mlat_string, "main"))
{
}
else
{
- warning ("bad value `%s' for -mmemory-latency", alpha_mlat_string);
+ warning (0, "bad value %qs for -mmemory-latency", alpha_mlat_string);
lat = 3;
}
REAL_MODE_FORMAT (DFmode) = &vax_g_format;
REAL_MODE_FORMAT (TFmode) = NULL;
}
+
+#ifdef TARGET_DEFAULT_LONG_DOUBLE_128
+ if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
+ target_flags |= MASK_LONG_DOUBLE_128;
+#endif
}
\f
/* Returns 1 if VALUE is a mask that contains full bytes of zero or ones. */
return 1;
}
-/* Return true if OP is valid for a particular TLS relocation.
+/* Return true if OP is valid for a particular TLS relocation.
We are already guaranteed that OP is a CONST. */
int
if (GET_CODE (op) != SYMBOL_REF)
return 0;
- if (SYMBOL_REF_LOCAL_P (op))
- {
- if (alpha_tls_size > size)
- return 0;
- }
- else
- {
- if (size != 64)
- return 0;
- }
-
switch (SYMBOL_REF_TLS_MODEL (op))
{
case TLS_MODEL_LOCAL_DYNAMIC:
- return unspec == UNSPEC_DTPREL;
+ return unspec == UNSPEC_DTPREL && size == alpha_tls_size;
case TLS_MODEL_INITIAL_EXEC:
return unspec == UNSPEC_TPREL && size == 64;
case TLS_MODEL_LOCAL_EXEC:
- return unspec == UNSPEC_TPREL;
+ return unspec == UNSPEC_TPREL && size == alpha_tls_size;
default:
- abort ();
+ gcc_unreachable ();
}
}
}
}
+/* The scalar modes supported differs from the default check-what-c-supports
+ version in that sometimes TFmode is available even when long double
+ indicates only DFmode. On unicosmk, we have the situation that HImode
+ doesn't map to any C type, but of course we still support that. */
+
+static bool
+alpha_scalar_mode_supported_p (enum machine_mode mode)
+{
+ switch (mode)
+ {
+ case QImode:
+ case HImode:
+ case SImode:
+ case DImode:
+ case TImode: /* via optabs.c */
+ return true;
+
+ case SFmode:
+ case DFmode:
+ return true;
+
+ case TFmode:
+ return TARGET_HAS_XFLOATING_LIBS;
+
+ default:
+ return false;
+ }
+}
+
+/* Alpha implements a couple of integer vector mode operations when
+ TARGET_MAX is enabled. We do not check TARGET_MAX here, however,
+ which allows the vectorizer to operate on e.g. move instructions,
+ or when expand_vector_operations can do something useful. */
+
+static bool
+alpha_vector_mode_supported_p (enum machine_mode mode)
+{
+ return mode == V8QImode || mode == V4HImode || mode == V2SImode;
+}
+
/* Return 1 if this function can directly return via $26. */
int
switch (tls_symbolic_operand_type (x))
{
+ case TLS_MODEL_NONE:
+ break;
+
case TLS_MODEL_GLOBAL_DYNAMIC:
start_sequence ();
tga = get_tls_get_addr ();
dest = gen_reg_rtx (Pmode);
seq = GEN_INT (alpha_next_sequence_number++);
-
+
emit_insn (gen_movdi_er_tlsgd (r16, pic_offset_table_rtx, x, seq));
insn = gen_call_value_osf_tlsgd (r0, tga, seq);
insn = emit_call_insn (insn);
emit_insn (gen_rtx_SET (VOIDmode, tp, insn));
}
return gen_rtx_LO_SUM (Pmode, tp, eqv);
+
+ default:
+ gcc_unreachable ();
}
if (local_symbolic_operand (x, Pmode))
}
}
+/* Primarily this is required for TLS symbols, but given that our move
+ patterns *ought* to be able to handle any symbol at any time, we
+ should never be spilling symbolic operands to the constant pool, ever. */
+
+static bool
+alpha_cannot_force_const_mem (rtx x)
+{
+ enum rtx_code code = GET_CODE (x);
+ return code == SYMBOL_REF || code == LABEL_REF || code == CONST;
+}
+
/* We do not allow indirect calls to be optimized into sibling calls, nor
can we allow a call to a function with a different GP to be optimized
into a sibcall. */
return decl_has_samegp (decl);
}
-/* For TARGET_EXPLICIT_RELOCS, we don't obfuscate a SYMBOL_REF to a
- small symbolic operand until after reload. At which point we need
- to replace (mem (symbol_ref)) with (mem (lo_sum $29 symbol_ref))
- so that sched2 has the proper dependency information. */
-/*
- {"some_small_symbolic_operand", {SET, PARALLEL, PREFETCH, UNSPEC, \
- UNSPEC_VOLATILE}},
-*/
-
-static int
-some_small_symbolic_operand_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
+int
+some_small_symbolic_operand_int (rtx *px, void *data ATTRIBUTE_UNUSED)
{
rtx x = *px;
return small_symbolic_operand (x, Pmode) != 0;
}
-int
-some_small_symbolic_operand (rtx x, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return for_each_rtx (&x, some_small_symbolic_operand_1, NULL);
-}
-
static int
split_small_symbolic_operand_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
{
return false;
}
-
+
/* Try a machine-dependent way of reloading an illegitimate address
operand. If we find one, push the reload and return the new rtx. */
-
+
rtx
alpha_legitimize_reload_address (rtx x,
enum machine_mode mode ATTRIBUTE_UNUSED,
if (optimize_size)
cost_data = &alpha_rtx_cost_size;
else
- cost_data = &alpha_rtx_cost_data[alpha_cpu];
+ cost_data = &alpha_rtx_cost_data[alpha_tune];
switch (code)
{
else
*total = COSTS_N_INSNS (2);
return true;
-
+
case CONST:
case SYMBOL_REF:
case LABEL_REF:
/* Otherwise we do a load from the GOT. */
*total = COSTS_N_INSNS (optimize_size ? 1 : alpha_memory_latency);
return true;
-
+
+ case HIGH:
+ /* This is effectively an add_operand. */
+ *total = 2;
+ return true;
+
case PLUS:
case MINUS:
if (float_mode_p)
case UNSIGNED_FLOAT:
case FIX:
case UNSIGNED_FIX:
- case FLOAT_EXTEND:
case FLOAT_TRUNCATE:
*total = cost_data->fp_add;
return false;
+ case FLOAT_EXTEND:
+ if (GET_CODE (XEXP (x, 0)) == MEM)
+ *total = 0;
+ else
+ *total = cost_data->fp_add;
+ return false;
+
default:
return false;
}
get_aligned_mem (rtx ref, rtx *paligned_mem, rtx *pbitnum)
{
rtx base;
- HOST_WIDE_INT offset = 0;
+ HOST_WIDE_INT disp, offset;
- if (GET_CODE (ref) != MEM)
- abort ();
+ gcc_assert (GET_CODE (ref) == MEM);
if (reload_in_progress
&& ! memory_address_p (GET_MODE (ref), XEXP (ref, 0)))
{
base = find_replacement (&XEXP (ref, 0));
-
- if (! memory_address_p (GET_MODE (ref), base))
- abort ();
+ gcc_assert (memory_address_p (GET_MODE (ref), base));
}
else
- {
- base = XEXP (ref, 0);
- }
+ base = XEXP (ref, 0);
if (GET_CODE (base) == PLUS)
- offset += INTVAL (XEXP (base, 1)), base = XEXP (base, 0);
+ disp = INTVAL (XEXP (base, 1)), base = XEXP (base, 0);
+ else
+ disp = 0;
+
+ /* Find the byte offset within an aligned word. If the memory itself is
+ claimed to be aligned, believe it. Otherwise, aligned_memory_operand
+ will have examined the base register and determined it is aligned, and
+ thus displacements from it are naturally alignable. */
+ if (MEM_ALIGN (ref) >= 32)
+ offset = 0;
+ else
+ offset = disp & 3;
- *paligned_mem
- = widen_memory_access (ref, SImode, (offset & ~3) - offset);
+ /* Access the entire aligned word. */
+ *paligned_mem = widen_memory_access (ref, SImode, -offset);
+ /* Convert the byte offset within the word to a bit offset. */
if (WORDS_BIG_ENDIAN)
- *pbitnum = GEN_INT (32 - (GET_MODE_BITSIZE (GET_MODE (ref))
- + (offset & 3) * 8));
+ offset = 32 - (GET_MODE_BITSIZE (GET_MODE (ref)) + offset * 8);
else
- *pbitnum = GEN_INT ((offset & 3) * 8);
+ offset *= 8;
+ *pbitnum = GEN_INT (offset);
}
-/* Similar, but just get the address. Handle the two reload cases.
+/* Similar, but just get the address. Handle the two reload cases.
Add EXTRA_OFFSET to the address we return. */
rtx
rtx base;
HOST_WIDE_INT offset = 0;
- if (GET_CODE (ref) != MEM)
- abort ();
+ gcc_assert (GET_CODE (ref) == MEM);
if (reload_in_progress
&& ! memory_address_p (GET_MODE (ref), XEXP (ref, 0)))
{
base = find_replacement (&XEXP (ref, 0));
- if (! memory_address_p (GET_MODE (ref), base))
- abort ();
+ gcc_assert (memory_address_p (GET_MODE (ref), base));
}
else
- {
- base = XEXP (ref, 0);
- }
+ base = XEXP (ref, 0);
if (GET_CODE (base) == PLUS)
offset += INTVAL (XEXP (base, 1)), base = XEXP (base, 0);
}
/* On the Alpha, all (non-symbolic) constants except zero go into
- a floating-point register via memory. Note that we cannot
+ a floating-point register via memory. Note that we cannot
return anything that is not a subset of CLASS, and that some
symbolic constants cannot be dropped to memory. */
return class;
/* These sorts of constants we can easily drop to memory. */
- if (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE)
+ if (GET_CODE (x) == CONST_INT
+ || GET_CODE (x) == CONST_DOUBLE
+ || GET_CODE (x) == CONST_VECTOR)
{
if (class == FLOAT_REGS)
return NO_REGS;
/* Loading and storing HImode or QImode values to and from memory
usually requires a scratch register. The exceptions are loading
QImode and HImode from an aligned address to a general register
- unless byte instructions are permitted.
+ unless byte instructions are permitted.
We also cannot load an unaligned address or a paradoxical SUBREG
- into an FP register.
+ into an FP register.
We also cannot do integral arithmetic into FP regs, as might result
from register elimination into a DImode fp register. */
enum reg_class
-secondary_reload_class (enum reg_class class, enum machine_mode mode,
+alpha_secondary_reload_class (enum reg_class class, enum machine_mode mode,
rtx x, int in)
{
if ((mode == QImode || mode == HImode) && ! TARGET_BWX)
/* Subfunction of the following function. Update the flags of any MEM
found in part of X. */
-static void
-alpha_set_memflags_1 (rtx x, int in_struct_p, int volatile_p, int unchanging_p)
+static int
+alpha_set_memflags_1 (rtx *xp, void *data)
{
- int i;
-
- switch (GET_CODE (x))
- {
- case SEQUENCE:
- abort ();
-
- case PARALLEL:
- for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
- alpha_set_memflags_1 (XVECEXP (x, 0, i), in_struct_p, volatile_p,
- unchanging_p);
- break;
+ rtx x = *xp, orig = (rtx) data;
- case INSN:
- alpha_set_memflags_1 (PATTERN (x), in_struct_p, volatile_p,
- unchanging_p);
- break;
+ if (GET_CODE (x) != MEM)
+ return 0;
- case SET:
- alpha_set_memflags_1 (SET_DEST (x), in_struct_p, volatile_p,
- unchanging_p);
- alpha_set_memflags_1 (SET_SRC (x), in_struct_p, volatile_p,
- unchanging_p);
- break;
+ MEM_VOLATILE_P (x) = MEM_VOLATILE_P (orig);
+ MEM_IN_STRUCT_P (x) = MEM_IN_STRUCT_P (orig);
+ MEM_SCALAR_P (x) = MEM_SCALAR_P (orig);
+ MEM_NOTRAP_P (x) = MEM_NOTRAP_P (orig);
+ MEM_READONLY_P (x) = MEM_READONLY_P (orig);
- case MEM:
- MEM_IN_STRUCT_P (x) = in_struct_p;
- MEM_VOLATILE_P (x) = volatile_p;
- RTX_UNCHANGING_P (x) = unchanging_p;
- /* Sadly, we cannot use alias sets because the extra aliasing
- produced by the AND interferes. Given that two-byte quantities
- are the only thing we would be able to differentiate anyway,
- there does not seem to be any point in convoluting the early
- out of the alias check. */
- break;
+ /* Sadly, we cannot use alias sets because the extra aliasing
+ produced by the AND interferes. Given that two-byte quantities
+ are the only thing we would be able to differentiate anyway,
+ there does not seem to be any point in convoluting the early
+ out of the alias check. */
- default:
- break;
- }
+ return -1;
}
/* Given INSN, which is an INSN list or the PATTERN of a single insn
void
alpha_set_memflags (rtx insn, rtx ref)
{
- int in_struct_p, volatile_p, unchanging_p;
+ rtx *base_ptr;
if (GET_CODE (ref) != MEM)
return;
- in_struct_p = MEM_IN_STRUCT_P (ref);
- volatile_p = MEM_VOLATILE_P (ref);
- unchanging_p = RTX_UNCHANGING_P (ref);
-
- /* This is only called from alpha.md, after having had something
+ /* This is only called from alpha.md, after having had something
generated from one of the insn patterns. So if everything is
zero, the pattern is already up-to-date. */
- if (! in_struct_p && ! volatile_p && ! unchanging_p)
+ if (!MEM_VOLATILE_P (ref)
+ && !MEM_IN_STRUCT_P (ref)
+ && !MEM_SCALAR_P (ref)
+ && !MEM_NOTRAP_P (ref)
+ && !MEM_READONLY_P (ref))
return;
- alpha_set_memflags_1 (insn, in_struct_p, volatile_p, unchanging_p);
+ if (INSN_P (insn))
+ base_ptr = &PATTERN (insn);
+ else
+ base_ptr = &insn;
+ for_each_rtx (base_ptr, alpha_set_memflags_1, (void *) ref);
}
\f
-/* Internal routine for alpha_emit_set_const to check for N or below insns. */
+static rtx alpha_emit_set_const (rtx, enum machine_mode, HOST_WIDE_INT,
+ int, bool);
+
+/* Internal routine for alpha_emit_set_const to check for N or below insns.
+ If NO_OUTPUT is true, then we only check to see if N insns are possible,
+ and return pc_rtx if successful. */
static rtx
alpha_emit_set_const_1 (rtx target, enum machine_mode mode,
- HOST_WIDE_INT c, int n)
+ HOST_WIDE_INT c, int n, bool no_output)
{
HOST_WIDE_INT new;
int i, bits;
{
/* We used to use copy_to_suggested_reg (GEN_INT (c), target, mode)
but that meant that we can't handle INT_MIN on 32-bit machines
- (like NT/Alpha), because we recurse indefinitely through
+ (like NT/Alpha), because we recurse indefinitely through
emit_move_insn to gen_movdi. So instead, since we know exactly
what we want, create it explicitly. */
+ if (no_output)
+ return pc_rtx;
if (target == NULL)
target = gen_reg_rtx (mode);
emit_insn (gen_rtx_SET (VOIDmode, target, GEN_INT (c)));
}
else if (n >= 2 + (extra != 0))
{
+ if (no_output)
+ return pc_rtx;
if (no_new_pseudos)
{
emit_insn (gen_rtx_SET (VOIDmode, target, GEN_INT (high << 16)));
high bits. */
new = ((c & 0xffff) ^ 0x8000) - 0x8000;
- if (new != 0
- && (temp = alpha_emit_set_const (subtarget, mode, c - new, i)) != 0)
- return expand_binop (mode, add_optab, temp, GEN_INT (new),
- target, 0, OPTAB_WIDEN);
+ if (new != 0)
+ {
+ temp = alpha_emit_set_const (subtarget, mode, c - new, i, no_output);
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_binop (mode, add_optab, temp, GEN_INT (new),
+ target, 0, OPTAB_WIDEN);
+ }
+ }
/* Next try complementing. */
- if ((temp = alpha_emit_set_const (subtarget, mode, ~ c, i)) != 0)
- return expand_unop (mode, one_cmpl_optab, temp, target, 0);
+ temp = alpha_emit_set_const (subtarget, mode, ~c, i, no_output);
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_unop (mode, one_cmpl_optab, temp, target, 0);
+ }
/* Next try to form a constant and do a left shift. We can do this
if some low-order bits are zero; the exact_log2 call below tells
bits to shift, but try all possibilities in case a ZAPNOT will
be useful. */
- if ((bits = exact_log2 (c & - c)) > 0)
+ bits = exact_log2 (c & -c);
+ if (bits > 0)
for (; bits > 0; bits--)
- if ((temp = (alpha_emit_set_const
- (subtarget, mode, c >> bits, i))) != 0
- || ((temp = (alpha_emit_set_const
- (subtarget, mode,
- ((unsigned HOST_WIDE_INT) c) >> bits, i)))
- != 0))
- return expand_binop (mode, ashl_optab, temp, GEN_INT (bits),
- target, 0, OPTAB_WIDEN);
+ {
+ new = c >> bits;
+ temp = alpha_emit_set_const (subtarget, mode, new, i, no_output);
+ if (!temp && c < 0)
+ {
+ new = (unsigned HOST_WIDE_INT)c >> bits;
+ temp = alpha_emit_set_const (subtarget, mode, new,
+ i, no_output);
+ }
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_binop (mode, ashl_optab, temp, GEN_INT (bits),
+ target, 0, OPTAB_WIDEN);
+ }
+ }
/* Now try high-order zero bits. Here we try the shifted-in bits as
all zero and all ones. Be careful to avoid shifting outside the
/* On narrow hosts, don't shift a 1 into the high bit, since we'll
confuse the recursive call and set all of the high 32 bits. */
- if ((bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
- - floor_log2 (c) - 1 - (HOST_BITS_PER_WIDE_INT < 64))) > 0)
+ bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
+ - floor_log2 (c) - 1 - (HOST_BITS_PER_WIDE_INT < 64));
+ if (bits > 0)
for (; bits > 0; bits--)
- if ((temp = alpha_emit_set_const (subtarget, mode,
- c << bits, i)) != 0
- || ((temp = (alpha_emit_set_const
- (subtarget, mode,
- ((c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1)),
- i)))
- != 0))
- return expand_binop (mode, lshr_optab, temp, GEN_INT (bits),
- target, 1, OPTAB_WIDEN);
+ {
+ new = c << bits;
+ temp = alpha_emit_set_const (subtarget, mode, new, i, no_output);
+ if (!temp)
+ {
+ new = (c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1);
+ temp = alpha_emit_set_const (subtarget, mode, new,
+ i, no_output);
+ }
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_binop (mode, lshr_optab, temp, GEN_INT (bits),
+ target, 1, OPTAB_WIDEN);
+ }
+ }
/* Now try high-order 1 bits. We get that with a sign-extension.
But one bit isn't enough here. Be careful to avoid shifting outside
the mode and to avoid shifting outside the host wide int size. */
- if ((bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
- - floor_log2 (~ c) - 2)) > 0)
+ bits = (MIN (HOST_BITS_PER_WIDE_INT, GET_MODE_SIZE (mode) * 8)
+ - floor_log2 (~ c) - 2);
+ if (bits > 0)
for (; bits > 0; bits--)
- if ((temp = alpha_emit_set_const (subtarget, mode,
- c << bits, i)) != 0
- || ((temp = (alpha_emit_set_const
- (subtarget, mode,
- ((c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1)),
- i)))
- != 0))
- return expand_binop (mode, ashr_optab, temp, GEN_INT (bits),
- target, 0, OPTAB_WIDEN);
+ {
+ new = c << bits;
+ temp = alpha_emit_set_const (subtarget, mode, new, i, no_output);
+ if (!temp)
+ {
+ new = (c << bits) | (((HOST_WIDE_INT) 1 << bits) - 1);
+ temp = alpha_emit_set_const (subtarget, mode, new,
+ i, no_output);
+ }
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_binop (mode, ashr_optab, temp, GEN_INT (bits),
+ target, 0, OPTAB_WIDEN);
+ }
+ }
}
#if HOST_BITS_PER_WIDE_INT == 64
if (mode == SImode)
new = ((new & 0xffffffff) ^ 0x80000000) - 0x80000000;
- if (new != c && new != -1
- && (temp = alpha_emit_set_const (subtarget, mode, new, n - 1)) != 0)
- return expand_binop (mode, and_optab, temp, GEN_INT (c | ~ new),
- target, 0, OPTAB_WIDEN);
+ if (new != c)
+ {
+ temp = alpha_emit_set_const (subtarget, mode, new, n - 1, no_output);
+ if (temp)
+ {
+ if (no_output)
+ return temp;
+ return expand_binop (mode, and_optab, temp, GEN_INT (c | ~ new),
+ target, 0, OPTAB_WIDEN);
+ }
+ }
#endif
return 0;
emitted. If it would take more than N insns, zero is returned and no
insns and emitted. */
-rtx
+static rtx
alpha_emit_set_const (rtx target, enum machine_mode mode,
- HOST_WIDE_INT c, int n)
+ HOST_WIDE_INT c, int n, bool no_output)
{
- rtx result = 0;
+ enum machine_mode orig_mode = mode;
rtx orig_target = target;
+ rtx result = 0;
int i;
/* If we can't make any pseudos, TARGET is an SImode hard register, we
can't load this constant in one insn, do this in DImode. */
if (no_new_pseudos && mode == SImode
- && GET_CODE (target) == REG && REGNO (target) < FIRST_PSEUDO_REGISTER
- && (result = alpha_emit_set_const_1 (target, mode, c, 1)) == 0)
+ && GET_CODE (target) == REG && REGNO (target) < FIRST_PSEUDO_REGISTER)
+ {
+ result = alpha_emit_set_const_1 (target, mode, c, 1, no_output);
+ if (result)
+ return result;
+
+ target = no_output ? NULL : gen_lowpart (DImode, target);
+ mode = DImode;
+ }
+ else if (mode == V8QImode || mode == V4HImode || mode == V2SImode)
{
- target = gen_lowpart (DImode, target);
+ target = no_output ? NULL : gen_lowpart (DImode, target);
mode = DImode;
}
/* Try 1 insn, then 2, then up to N. */
for (i = 1; i <= n; i++)
{
- result = alpha_emit_set_const_1 (target, mode, c, i);
+ result = alpha_emit_set_const_1 (target, mode, c, i, no_output);
if (result)
{
- rtx insn = get_last_insn ();
- rtx set = single_set (insn);
+ rtx insn, set;
+
+ if (no_output)
+ return result;
+
+ insn = get_last_insn ();
+ set = single_set (insn);
if (! CONSTANT_P (SET_SRC (set)))
set_unique_reg_note (get_last_insn (), REG_EQUAL, GEN_INT (c));
break;
}
/* Allow for the case where we changed the mode of TARGET. */
- if (result == target)
- result = orig_target;
+ if (result)
+ {
+ if (result == target)
+ result = orig_target;
+ else if (mode != orig_mode)
+ result = gen_lowpart (orig_mode, result);
+ }
return result;
}
exponential run times encountered when looking for longer sequences
with alpha_emit_set_const. */
-rtx
+static rtx
alpha_emit_set_long_const (rtx target, HOST_WIDE_INT c1, HOST_WIDE_INT c2)
{
HOST_WIDE_INT d1, d2, d3, d4;
/* Decompose the entire word */
#if HOST_BITS_PER_WIDE_INT >= 64
- if (c2 != -(c1 < 0))
- abort ();
+ gcc_assert (c2 == -(c1 < 0));
d1 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
c1 -= d1;
d2 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
d3 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
c1 -= d3;
d4 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
- if (c1 != d4)
- abort ();
+ gcc_assert (c1 == d4);
#else
d1 = ((c1 & 0xffff) ^ 0x8000) - 0x8000;
c1 -= d1;
d2 = ((c1 & 0xffffffff) ^ 0x80000000) - 0x80000000;
- if (c1 != d2)
- abort ();
+ gcc_assert (c1 == d2);
c2 += (d2 < 0);
d3 = ((c2 & 0xffff) ^ 0x8000) - 0x8000;
c2 -= d3;
d4 = ((c2 & 0xffffffff) ^ 0x80000000) - 0x80000000;
- if (c2 != d4)
- abort ();
+ gcc_assert (c2 == d4);
#endif
/* Construct the high word */
return target;
}
+/* Given an integral CONST_INT, CONST_DOUBLE, or CONST_VECTOR, return
+ the low 64 bits. */
+
+static void
+alpha_extract_integer (rtx x, HOST_WIDE_INT *p0, HOST_WIDE_INT *p1)
+{
+ HOST_WIDE_INT i0, i1;
+
+ if (GET_CODE (x) == CONST_VECTOR)
+ x = simplify_subreg (DImode, x, GET_MODE (x), 0);
+
+
+ if (GET_CODE (x) == CONST_INT)
+ {
+ i0 = INTVAL (x);
+ i1 = -(i0 < 0);
+ }
+ else if (HOST_BITS_PER_WIDE_INT >= 64)
+ {
+ i0 = CONST_DOUBLE_LOW (x);
+ i1 = -(i0 < 0);
+ }
+ else
+ {
+ i0 = CONST_DOUBLE_LOW (x);
+ i1 = CONST_DOUBLE_HIGH (x);
+ }
+
+ *p0 = i0;
+ *p1 = i1;
+}
+
+/* Implement LEGITIMATE_CONSTANT_P. This is all constants for which we
+ are willing to load the value into a register via a move pattern.
+ Normally this is all symbolic constants, integral constants that
+ take three or fewer instructions, and floating-point zero. */
+
+bool
+alpha_legitimate_constant_p (rtx x)
+{
+ enum machine_mode mode = GET_MODE (x);
+ HOST_WIDE_INT i0, i1;
+
+ switch (GET_CODE (x))
+ {
+ case CONST:
+ case LABEL_REF:
+ case SYMBOL_REF:
+ case HIGH:
+ return true;
+
+ case CONST_DOUBLE:
+ if (x == CONST0_RTX (mode))
+ return true;
+ if (FLOAT_MODE_P (mode))
+ return false;
+ goto do_integer;
+
+ case CONST_VECTOR:
+ if (x == CONST0_RTX (mode))
+ return true;
+ if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
+ return false;
+ if (GET_MODE_SIZE (mode) != 8)
+ return false;
+ goto do_integer;
+
+ case CONST_INT:
+ do_integer:
+ if (TARGET_BUILD_CONSTANTS)
+ return true;
+ alpha_extract_integer (x, &i0, &i1);
+ if (HOST_BITS_PER_WIDE_INT >= 64 || i1 == (-i0 < 0))
+ return alpha_emit_set_const_1 (x, mode, i0, 3, true) != NULL;
+ return false;
+
+ default:
+ return false;
+ }
+}
+
+/* Operand 1 is known to be a constant, and should require more than one
+ instruction to load. Emit that multi-part load. */
+
+bool
+alpha_split_const_mov (enum machine_mode mode, rtx *operands)
+{
+ HOST_WIDE_INT i0, i1;
+ rtx temp = NULL_RTX;
+
+ alpha_extract_integer (operands[1], &i0, &i1);
+
+ if (HOST_BITS_PER_WIDE_INT >= 64 || i1 == -(i0 < 0))
+ temp = alpha_emit_set_const (operands[0], mode, i0, 3, false);
+
+ if (!temp && TARGET_BUILD_CONSTANTS)
+ temp = alpha_emit_set_long_const (operands[0], i0, i1);
+
+ if (temp)
+ {
+ if (!rtx_equal_p (operands[0], temp))
+ emit_move_insn (operands[0], temp);
+ return true;
+ }
+
+ return false;
+}
+
/* Expand a move instruction; return true if all work is done.
We don't handle non-bwx subword loads here. */
/* Split large integers. */
if (GET_CODE (operands[1]) == CONST_INT
- || GET_CODE (operands[1]) == CONST_DOUBLE)
+ || GET_CODE (operands[1]) == CONST_DOUBLE
+ || GET_CODE (operands[1]) == CONST_VECTOR)
{
- HOST_WIDE_INT i0, i1;
- rtx temp = NULL_RTX;
-
- if (GET_CODE (operands[1]) == CONST_INT)
- {
- i0 = INTVAL (operands[1]);
- i1 = -(i0 < 0);
- }
- else if (HOST_BITS_PER_WIDE_INT >= 64)
- {
- i0 = CONST_DOUBLE_LOW (operands[1]);
- i1 = -(i0 < 0);
- }
- else
- {
- i0 = CONST_DOUBLE_LOW (operands[1]);
- i1 = CONST_DOUBLE_HIGH (operands[1]);
- }
-
- if (HOST_BITS_PER_WIDE_INT >= 64 || i1 == -(i0 < 0))
- temp = alpha_emit_set_const (operands[0], mode, i0, 3);
-
- if (!temp && TARGET_BUILD_CONSTANTS)
- temp = alpha_emit_set_long_const (operands[0], i0, i1);
-
- if (temp)
- {
- if (rtx_equal_p (operands[0], temp))
- return true;
- operands[1] = temp;
- return false;
- }
- }
+ if (alpha_split_const_mov (mode, operands))
+ return true;
+ }
/* Otherwise we've nothing left but to drop the thing to memory. */
operands[1] = force_const_mem (mode, operands[1]);
return false;
}
+/* Implement the movmisalign patterns. One of the operands is a memory
+ that is not naturally aligned. Emit instructions to load it. */
+
+void
+alpha_expand_movmisalign (enum machine_mode mode, rtx *operands)
+{
+ /* Honor misaligned loads, for those we promised to do so. */
+ if (MEM_P (operands[1]))
+ {
+ rtx tmp;
+
+ if (register_operand (operands[0], mode))
+ tmp = operands[0];
+ else
+ tmp = gen_reg_rtx (mode);
+
+ alpha_expand_unaligned_load (tmp, operands[1], 8, 0, 0);
+ if (tmp != operands[0])
+ emit_move_insn (operands[0], tmp);
+ }
+ else if (MEM_P (operands[0]))
+ {
+ if (!reg_or_0_operand (operands[1], mode))
+ operands[1] = force_reg (mode, operands[1]);
+ alpha_expand_unaligned_store (operands[0], operands[1], 8, 0);
+ }
+ else
+ gcc_unreachable ();
+}
+
/* Generate an unsigned DImode to FP conversion. This is the same code
optabs would emit if we didn't have TFmode patterns.
if (alpha_compare.fp_p && GET_MODE (op0) == TFmode)
{
- if (! TARGET_HAS_XFLOATING_LIBS)
- abort ();
-
- /* X_floating library comparison functions return
- -1 unordered
- 0 false
- 1 true
- Convert the compare against the raw return value. */
-
- switch (code)
- {
- case UNORDERED:
- cmp_code = EQ;
- code = LT;
- break;
- case ORDERED:
- cmp_code = EQ;
- code = GE;
- break;
- case NE:
- cmp_code = NE;
- code = NE;
- break;
- default:
- cmp_code = code;
- code = GT;
- break;
- }
-
- op0 = alpha_emit_xfloating_compare (cmp_code, op0, op1);
+ op0 = alpha_emit_xfloating_compare (&code, op0, op1);
op1 = const0_rtx;
alpha_compare.fp_p = 0;
}
break;
default:
- abort ();
+ gcc_unreachable ();
}
if (alpha_compare.fp_p)
/* When we are not as concerned about non-finite values, and we
are comparing against zero, we can branch directly. */
if (op1 == CONST0_RTX (DFmode))
- cmp_code = NIL, branch_code = code;
+ cmp_code = UNKNOWN, branch_code = code;
else if (op0 == CONST0_RTX (DFmode))
{
/* Undo the swap we probably did just above. */
tem = op0, op0 = op1, op1 = tem;
branch_code = swap_condition (cmp_code);
- cmp_code = NIL;
+ cmp_code = UNKNOWN;
}
}
else
{
/* ??? We mark the branch mode to be CCmode to prevent the
- compare and branch from being combined, since the compare
+ compare and branch from being combined, since the compare
insn follows IEEE rules that the branch does not. */
branch_mode = CCmode;
}
{
/* Whee. Compare and branch against 0 directly. */
if (op1 == const0_rtx)
- cmp_code = NIL, branch_code = code;
+ cmp_code = UNKNOWN, branch_code = code;
/* If the constants doesn't fit into an immediate, but can
be generated by lda/ldah, we adjust the argument and
/* Emit an initial compare instruction, if necessary. */
tem = op0;
- if (cmp_code != NIL)
+ if (cmp_code != UNKNOWN)
{
tem = gen_reg_rtx (cmp_mode);
emit_move_insn (tem, gen_rtx_fmt_ee (cmp_code, cmp_mode, op0, op1));
if (fp_p && GET_MODE (op0) == TFmode)
{
- if (! TARGET_HAS_XFLOATING_LIBS)
- abort ();
-
- /* X_floating library comparison functions return
- -1 unordered
- 0 false
- 1 true
- Convert the compare against the raw return value. */
-
- if (code == UNORDERED || code == ORDERED)
- cmp_code = EQ;
- else
- cmp_code = code;
-
- op0 = alpha_emit_xfloating_compare (cmp_code, op0, op1);
+ op0 = alpha_emit_xfloating_compare (&code, op0, op1);
op1 = const0_rtx;
fp_p = 0;
-
- if (code == UNORDERED)
- code = LT;
- else if (code == ORDERED)
- code = GE;
- else
- code = GT;
}
if (fp_p && !TARGET_FIX)
/* The general case: fold the comparison code to the types of compares
that we have, choosing the branch as necessary. */
- cmp_code = NIL;
+ cmp_code = UNKNOWN;
switch (code)
{
case EQ: case LE: case LT: case LEU: case LTU:
break;
default:
- abort ();
+ gcc_unreachable ();
}
if (!fp_p)
}
/* Emit an initial compare instruction, if necessary. */
- if (cmp_code != NIL)
+ if (cmp_code != UNKNOWN)
{
enum machine_mode mode = fp_p ? DFmode : DImode;
break;
default:
- abort ();
+ gcc_unreachable ();
}
tem = gen_reg_rtx (cmp_op_mode);
break;
default:
- abort ();
+ gcc_unreachable ();
}
if (!fp_p)
/* On EV6, we've got enough shifters to make non-arithmetic shifts
viable over a longer latency cmove. On EV5, the E0 slot is a
scarce resource, and on EV4 shift has the same latency as a cmove. */
- && (diff <= 8 || alpha_cpu == PROCESSOR_EV6))
+ && (diff <= 8 || alpha_tune == PROCESSOR_EV6))
{
tmp = gen_rtx_fmt_ee (code, DImode, cond, const0_rtx);
emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (subtarget), tmp));
rtx libcall;
};
-static GTY(()) struct xfloating_op xfloating_ops[] =
+static GTY(()) struct xfloating_op xfloating_ops[] =
{
{ PLUS, "_OtsAddX", "OTS$ADD_X", 0 },
{ MINUS, "_OtsSubX", "OTS$SUB_X", 0 },
long n = ARRAY_SIZE (xfloating_ops);
long i;
+ gcc_assert (TARGET_HAS_XFLOATING_LIBS);
+
/* How irritating. Nothing to key off for the main table. */
if (TARGET_FLOAT_VAX && (code == FLOAT_EXTEND || code == FLOAT_TRUNCATE))
{
return func;
}
- abort();
+ gcc_unreachable ();
}
/* Most X_floating operations take the rounding mode as an argument.
mode = 4;
break;
default:
- abort ();
+ gcc_unreachable ();
/* XXX For reference, round to +inf is mode = 3. */
}
Note that these functions do not follow normal calling conventions:
TFmode arguments are passed in two integer registers (as opposed to
- indirect); TFmode return values appear in R16+R17.
+ indirect); TFmode return values appear in R16+R17.
FUNC is the function to call.
TARGET is where the output belongs.
break;
case VOIDmode:
- if (GET_CODE (operands[i]) != CONST_INT)
- abort ();
+ gcc_assert (GET_CODE (operands[i]) == CONST_INT);
/* FALLTHRU */
case DImode:
reg = gen_rtx_REG (DImode, regno);
break;
default:
- abort ();
+ gcc_unreachable ();
}
emit_move_insn (reg, operands[i]);
reg = gen_rtx_REG (DImode, 0);
break;
default:
- abort ();
+ gcc_unreachable ();
}
tmp = gen_rtx_MEM (QImode, func);
out_operands[0] = operands[1];
out_operands[1] = operands[2];
out_operands[2] = GEN_INT (mode);
- alpha_emit_xfloating_libcall (func, operands[0], out_operands, 3,
+ alpha_emit_xfloating_libcall (func, operands[0], out_operands, 3,
gen_rtx_fmt_ee (code, TFmode, operands[1],
operands[2]));
}
/* Emit an X_floating library function call for a comparison. */
static rtx
-alpha_emit_xfloating_compare (enum rtx_code code, rtx op0, rtx op1)
+alpha_emit_xfloating_compare (enum rtx_code *pcode, rtx op0, rtx op1)
{
- rtx func;
- rtx out, operands[2];
+ enum rtx_code cmp_code, res_code;
+ rtx func, out, operands[2];
- func = alpha_lookup_xfloating_lib_func (code);
+ /* X_floating library comparison functions return
+ -1 unordered
+ 0 false
+ 1 true
+ Convert the compare against the raw return value. */
+
+ cmp_code = *pcode;
+ switch (cmp_code)
+ {
+ case UNORDERED:
+ cmp_code = EQ;
+ res_code = LT;
+ break;
+ case ORDERED:
+ cmp_code = EQ;
+ res_code = GE;
+ break;
+ case NE:
+ res_code = NE;
+ break;
+ case EQ:
+ case LT:
+ case GT:
+ case LE:
+ case GE:
+ res_code = GT;
+ break;
+ default:
+ gcc_unreachable ();
+ }
+ *pcode = res_code;
+
+ func = alpha_lookup_xfloating_lib_func (cmp_code);
operands[0] = op0;
operands[1] = op1;
/* ??? Strange mode for equiv because what's actually returned
is -1,0,1, not a proper boolean value. */
alpha_emit_xfloating_libcall (func, out, operands, 2,
- gen_rtx_fmt_ee (code, CCmode, op0, op1));
+ gen_rtx_fmt_ee (cmp_code, CCmode, op0, op1));
return out;
}
operands[1]));
}
-/* Split a TFmode OP[1] into DImode OP[2,3] and likewise for
- OP[0] into OP[0,1]. Naturally, output operand ordering is
- little-endian. */
-
+/* Split a TImode or TFmode move from OP[1] to OP[0] into a pair of
+ DImode moves from OP[2,3] to OP[0,1]. If FIXUP_OVERLAP is true,
+ guarantee that the sequence
+ set (OP[0] OP[2])
+ set (OP[1] OP[3])
+ is valid. Naturally, output operand ordering is little-endian.
+ This is used by *movtf_internal and *movti_internal. */
+
void
-alpha_split_tfmode_pair (rtx operands[4])
+alpha_split_tmode_pair (rtx operands[4], enum machine_mode mode,
+ bool fixup_overlap)
{
- if (GET_CODE (operands[1]) == REG)
+ switch (GET_CODE (operands[1]))
{
+ case REG:
operands[3] = gen_rtx_REG (DImode, REGNO (operands[1]) + 1);
operands[2] = gen_rtx_REG (DImode, REGNO (operands[1]));
- }
- else if (GET_CODE (operands[1]) == MEM)
- {
+ break;
+
+ case MEM:
operands[3] = adjust_address (operands[1], DImode, 8);
operands[2] = adjust_address (operands[1], DImode, 0);
+ break;
+
+ case CONST_INT:
+ case CONST_DOUBLE:
+ gcc_assert (operands[1] == CONST0_RTX (mode));
+ operands[2] = operands[3] = const0_rtx;
+ break;
+
+ default:
+ gcc_unreachable ();
}
- else if (operands[1] == CONST0_RTX (TFmode))
- operands[2] = operands[3] = const0_rtx;
- else
- abort ();
- if (GET_CODE (operands[0]) == REG)
+ switch (GET_CODE (operands[0]))
{
+ case REG:
operands[1] = gen_rtx_REG (DImode, REGNO (operands[0]) + 1);
operands[0] = gen_rtx_REG (DImode, REGNO (operands[0]));
- }
- else if (GET_CODE (operands[0]) == MEM)
- {
+ break;
+
+ case MEM:
operands[1] = adjust_address (operands[0], DImode, 8);
operands[0] = adjust_address (operands[0], DImode, 0);
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ if (fixup_overlap && reg_overlap_mentioned_p (operands[0], operands[3]))
+ {
+ rtx tmp;
+ tmp = operands[0], operands[0] = operands[1], operands[1] = tmp;
+ tmp = operands[2], operands[2] = operands[3], operands[3] = tmp;
}
- else
- abort ();
}
-/* Implement negtf2 or abstf2. Op0 is destination, op1 is source,
- op2 is a register containing the sign bit, operation is the
+/* Implement negtf2 or abstf2. Op0 is destination, op1 is source,
+ op2 is a register containing the sign bit, operation is the
logical operation to be performed. */
void
rtx scratch;
int move;
- alpha_split_tfmode_pair (operands);
+ alpha_split_tmode_pair (operands, TFmode, false);
/* Detect three flavors of operand overlap. */
move = 1;
rtx meml, memh, addr, extl, exth, tmp, mema;
enum machine_mode mode;
+ if (TARGET_BWX && size == 2)
+ {
+ meml = adjust_address (mem, QImode, ofs);
+ memh = adjust_address (mem, QImode, ofs+1);
+ if (BYTES_BIG_ENDIAN)
+ tmp = meml, meml = memh, memh = tmp;
+ extl = gen_reg_rtx (DImode);
+ exth = gen_reg_rtx (DImode);
+ emit_insn (gen_zero_extendqidi2 (extl, meml));
+ emit_insn (gen_zero_extendqidi2 (exth, memh));
+ exth = expand_simple_binop (DImode, ASHIFT, exth, GEN_INT (8),
+ NULL, 1, OPTAB_LIB_WIDEN);
+ addr = expand_simple_binop (DImode, IOR, extl, exth,
+ NULL, 1, OPTAB_LIB_WIDEN);
+
+ if (sign && GET_MODE (tgt) != HImode)
+ {
+ addr = gen_lowpart (HImode, addr);
+ emit_insn (gen_extend_insn (tgt, addr, GET_MODE (tgt), HImode, 0));
+ }
+ else
+ {
+ if (GET_MODE (tgt) != DImode)
+ addr = gen_lowpart (GET_MODE (tgt), addr);
+ emit_move_insn (tgt, addr);
+ }
+ return;
+ }
+
meml = gen_reg_rtx (DImode);
memh = gen_reg_rtx (DImode);
addr = gen_reg_rtx (DImode);
mema = force_reg (Pmode, mema);
/* AND addresses cannot be in any alias set, since they may implicitly
- alias surrounding code. Ideally we'd have some alias set that
+ alias surrounding code. Ideally we'd have some alias set that
covered all types except those with alignment 8 or higher. */
tmp = change_address (mem, DImode,
- gen_rtx_AND (DImode,
+ gen_rtx_AND (DImode,
plus_constant (mema, ofs),
GEN_INT (-8)));
set_mem_alias_set (tmp, 0);
emit_move_insn (meml, tmp);
tmp = change_address (mem, DImode,
- gen_rtx_AND (DImode,
+ gen_rtx_AND (DImode,
plus_constant (mema, ofs + size - 1),
GEN_INT (-8)));
set_mem_alias_set (tmp, 0);
addr for the target, because addr is marked as a pointer and combine
knows that pointers are always sign-extended 32 bit values. */
addr = expand_binop (DImode, ior_optab, extl, exth, tgt, 1, OPTAB_WIDEN);
- addr = expand_binop (DImode, ashr_optab, addr, GEN_INT (48),
+ addr = expand_binop (DImode, ashr_optab, addr, GEN_INT (48),
addr, 1, OPTAB_WIDEN);
}
else
break;
default:
- abort ();
+ gcc_unreachable ();
}
emit_insn (gen_extxl_be (exth, memh, GEN_INT (size*8), addr));
}
break;
default:
- abort();
+ gcc_unreachable ();
}
}
}
if (addr != tgt)
- emit_move_insn (tgt, gen_lowpart(GET_MODE (tgt), addr));
+ emit_move_insn (tgt, gen_lowpart (GET_MODE (tgt), addr));
}
/* Similarly, use ins and msk instructions to perform unaligned stores. */
HOST_WIDE_INT size, HOST_WIDE_INT ofs)
{
rtx dstl, dsth, addr, insl, insh, meml, memh, dsta;
-
+
+ if (TARGET_BWX && size == 2)
+ {
+ if (src != const0_rtx)
+ {
+ dstl = gen_lowpart (QImode, src);
+ dsth = expand_simple_binop (DImode, LSHIFTRT, src, GEN_INT (8),
+ NULL, 1, OPTAB_LIB_WIDEN);
+ dsth = gen_lowpart (QImode, dsth);
+ }
+ else
+ dstl = dsth = const0_rtx;
+
+ meml = adjust_address (dst, QImode, ofs);
+ memh = adjust_address (dst, QImode, ofs+1);
+ if (BYTES_BIG_ENDIAN)
+ addr = meml, meml = memh, memh = addr;
+
+ emit_move_insn (meml, dstl);
+ emit_move_insn (memh, dsth);
+ return;
+ }
+
dstl = gen_reg_rtx (DImode);
dsth = gen_reg_rtx (DImode);
insl = gen_reg_rtx (DImode);
dsta = force_reg (Pmode, dsta);
/* AND addresses cannot be in any alias set, since they may implicitly
- alias surrounding code. Ideally we'd have some alias set that
+ alias surrounding code. Ideally we'd have some alias set that
covered all types except those with alignment 8 or higher. */
meml = change_address (dst, DImode,
- gen_rtx_AND (DImode,
+ gen_rtx_AND (DImode,
plus_constant (dsta, ofs),
GEN_INT (-8)));
set_mem_alias_set (meml, 0);
memh = change_address (dst, DImode,
- gen_rtx_AND (DImode,
+ gen_rtx_AND (DImode,
plus_constant (dsta, ofs + size - 1),
GEN_INT (-8)));
set_mem_alias_set (memh, 0);
{
addr = copy_addr_to_reg (plus_constant (dsta, ofs));
- if (src != const0_rtx)
+ if (src != CONST0_RTX (GET_MODE (src)))
{
emit_insn (gen_insxh (insh, gen_lowpart (DImode, src),
GEN_INT (size*8), addr));
}
}
- if (src != const0_rtx)
+ if (src != CONST0_RTX (GET_MODE (src)))
{
dsth = expand_binop (DImode, ior_optab, insh, dsth, dsth, 0, OPTAB_WIDEN);
dstl = expand_binop (DImode, ior_optab, insl, dstl, dstl, 0, OPTAB_WIDEN);
}
-
+
if (WORDS_BIG_ENDIAN)
{
emit_move_insn (meml, dstl);
if (ofs != 0)
smem = adjust_address (smem, GET_MODE (smem), ofs);
-
+
/* Load up all of the source data. */
for (i = 0; i < words; ++i)
{
emit_move_insn (data_regs[words], tmp);
/* Extract the half-word fragments. Unfortunately DEC decided to make
- extxh with offset zero a noop instead of zeroing the register, so
+ extxh with offset zero a noop instead of zeroing the register, so
we must take care of that edge condition ourselves with cmov. */
sreg = copy_addr_to_reg (smema);
- areg = expand_binop (DImode, and_optab, sreg, GEN_INT (7), NULL,
+ areg = expand_binop (DImode, and_optab, sreg, GEN_INT (7), NULL,
1, OPTAB_WIDEN);
if (WORDS_BIG_ENDIAN)
emit_move_insn (sreg, plus_constant (sreg, 7));
ins_tmps[i] = gen_reg_rtx(DImode);
st_tmp_1 = gen_reg_rtx(DImode);
st_tmp_2 = gen_reg_rtx(DImode);
-
+
if (ofs != 0)
dmem = adjust_address (dmem, GET_MODE (dmem), ofs);
rtx data_regs[2 * MAX_MOVE_WORDS + 16];
rtx tmp;
unsigned int i, words, ofs, nregs = 0;
-
+
if (orig_bytes <= 0)
return 1;
else if (orig_bytes > MAX_MOVE_WORDS * UNITS_PER_WORD)
src_align = 16;
}
}
-
+
tmp = XEXP (orig_dst, 0);
if (GET_CODE (tmp) == REG)
dst_align = MAX (dst_align, REGNO_POINTER_ALIGN (REGNO (tmp)));
ofs += 1;
}
- if (nregs > ARRAY_SIZE (data_regs))
- abort ();
+ gcc_assert (nregs <= ARRAY_SIZE (data_regs));
/* Now save it back out again. */
else
alpha_expand_unaligned_store_words (data_regs + i, orig_dst,
words, ofs);
-
+
i += words;
ofs += words * 8;
}
ofs += 2;
}
- while (i < nregs && GET_MODE (data_regs[i]) == QImode)
+ /* The remainder must be byte copies. */
+ while (i < nregs)
{
+ gcc_assert (GET_MODE (data_regs[i]) == QImode);
emit_move_insn (adjust_address (orig_dst, QImode, ofs), data_regs[i]);
i++;
ofs += 1;
}
- if (i != nregs)
- abort ();
-
return 1;
}
alpha_expand_block_clear (rtx operands[])
{
rtx bytes_rtx = operands[1];
- rtx align_rtx = operands[2];
+ rtx align_rtx = operands[3];
HOST_WIDE_INT orig_bytes = INTVAL (bytes_rtx);
HOST_WIDE_INT bytes = orig_bytes;
HOST_WIDE_INT align = INTVAL (align_rtx) * BITS_PER_UNIT;
rtx orig_dst = operands[0];
rtx tmp;
int i, words, ofs = 0;
-
+
if (orig_bytes <= 0)
return 1;
if (orig_bytes > MAX_MOVE_WORDS * UNITS_PER_WORD)
result = gen_int_mode (mask, DImode);
}
- else if (HOST_BITS_PER_WIDE_INT == 32)
+ else
{
HOST_WIDE_INT mask_lo = 0, mask_hi = 0;
+ gcc_assert (HOST_BITS_PER_WIDE_INT == 32);
+
for (i = 7; i >= 4; --i)
{
mask_hi <<= 8;
result = immed_double_const (mask_lo, mask_hi, DImode);
}
- else
- abort ();
return result;
}
emit_insn ((*gen) (op0, op1, op2));
}
-\f
-/* 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
-alpha_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
-{
- enum attr_type insn_type, dep_insn_type;
-
- /* If the dependence is an anti-dependence, there is no cost. For an
- output dependence, there is sometimes a cost, but it doesn't seem
- worth handling those few cases. */
- if (REG_NOTE_KIND (link) != 0)
- return cost;
-
- /* If we can't recognize the insns, we can't really do anything. */
- if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
- return cost;
-
- insn_type = get_attr_type (insn);
- dep_insn_type = get_attr_type (dep_insn);
- /* Bring in the user-defined memory latency. */
- if (dep_insn_type == TYPE_ILD
- || dep_insn_type == TYPE_FLD
- || dep_insn_type == TYPE_LDSYM)
- cost += alpha_memory_latency-1;
+/* A subroutine of the atomic operation splitters. Jump to LABEL if
+ COND is true. Mark the jump as unlikely to be taken. */
- /* Everything else handled in DFA bypasses now. */
+static void
+emit_unlikely_jump (rtx cond, rtx label)
+{
+ rtx very_unlikely = GEN_INT (REG_BR_PROB_BASE / 100 - 1);
+ rtx x;
- return cost;
+ x = gen_rtx_IF_THEN_ELSE (VOIDmode, cond, label, pc_rtx);
+ x = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, x));
+ REG_NOTES (x) = gen_rtx_EXPR_LIST (REG_BR_PROB, very_unlikely, NULL_RTX);
}
-/* The number of instructions that can be issued per cycle. */
+/* A subroutine of the atomic operation splitters. Emit a load-locked
+ instruction in MODE. */
-static int
-alpha_issue_rate (void)
+static void
+emit_load_locked (enum machine_mode mode, rtx reg, rtx mem)
{
- return (alpha_cpu == PROCESSOR_EV4 ? 2 : 4);
+ rtx (*fn) (rtx, rtx) = NULL;
+ if (mode == SImode)
+ fn = gen_load_locked_si;
+ else if (mode == DImode)
+ fn = gen_load_locked_di;
+ emit_insn (fn (reg, mem));
}
-/* How many alternative schedules to try. This should be as wide as the
- scheduling freedom in the DFA, but no wider. Making this value too
- large results extra work for the scheduler.
-
- For EV4, loads can be issued to either IB0 or IB1, thus we have 2
- alternative schedules. For EV5, we can choose between E0/E1 and
- FA/FM. For EV6, an arithmetic insn can be issued to U0/U1/L0/L1. */
+/* A subroutine of the atomic operation splitters. Emit a store-conditional
+ instruction in MODE. */
-static int
-alpha_multipass_dfa_lookahead (void)
+static void
+emit_store_conditional (enum machine_mode mode, rtx res, rtx mem, rtx val)
{
- return (alpha_cpu == PROCESSOR_EV6 ? 4 : 2);
+ rtx (*fn) (rtx, rtx, rtx) = NULL;
+ if (mode == SImode)
+ fn = gen_store_conditional_si;
+ else if (mode == DImode)
+ fn = gen_store_conditional_di;
+ emit_insn (fn (res, mem, val));
}
-\f
-/* Machine-specific function data. */
-struct machine_function GTY(())
-{
- /* For unicosmk. */
- /* List of call information words for calls from this function. */
- struct rtx_def *first_ciw;
- struct rtx_def *last_ciw;
- int ciw_count;
+/* A subroutine of the atomic operation splitters. Emit an insxl
+ instruction in MODE. */
- /* List of deferred case vectors. */
- struct rtx_def *addr_list;
+static rtx
+emit_insxl (enum machine_mode mode, rtx op1, rtx op2)
+{
+ rtx ret = gen_reg_rtx (DImode);
+ rtx (*fn) (rtx, rtx, rtx);
- /* For OSF. */
- const char *some_ld_name;
+ if (WORDS_BIG_ENDIAN)
+ {
+ if (mode == QImode)
+ fn = gen_insbl_be;
+ else
+ fn = gen_inswl_be;
+ }
+ else
+ {
+ if (mode == QImode)
+ fn = gen_insbl_le;
+ else
+ fn = gen_inswl_le;
+ }
+ emit_insn (fn (ret, op1, op2));
- /* For TARGET_LD_BUGGY_LDGP. */
- struct rtx_def *gp_save_rtx;
-};
+ return ret;
+}
-/* How to allocate a 'struct machine_function'. */
+/* Expand an an atomic fetch-and-operate pattern. CODE is the binary operation
+ to perform. MEM is the memory on which to operate. VAL is the second
+ operand of the binary operator. BEFORE and AFTER are optional locations to
+ return the value of MEM either before of after the operation. SCRATCH is
+ a scratch register. */
-static struct machine_function *
-alpha_init_machine_status (void)
+void
+alpha_split_atomic_op (enum rtx_code code, rtx mem, rtx val,
+ rtx before, rtx after, rtx scratch)
{
- return ((struct machine_function *)
- ggc_alloc_cleared (sizeof (struct machine_function)));
-}
+ enum machine_mode mode = GET_MODE (mem);
+ rtx label, x, cond = gen_rtx_REG (DImode, REGNO (scratch));
-/* Functions to save and restore alpha_return_addr_rtx. */
+ emit_insn (gen_memory_barrier ());
-/* Start the ball rolling with RETURN_ADDR_RTX. */
+ label = gen_label_rtx ();
+ emit_label (label);
+ label = gen_rtx_LABEL_REF (DImode, label);
-rtx
-alpha_return_addr (int count, rtx frame ATTRIBUTE_UNUSED)
-{
- if (count != 0)
- return const0_rtx;
+ if (before == NULL)
+ before = scratch;
+ emit_load_locked (mode, before, mem);
- return get_hard_reg_initial_val (Pmode, REG_RA);
+ if (code == NOT)
+ x = gen_rtx_AND (mode, gen_rtx_NOT (mode, before), val);
+ else
+ x = gen_rtx_fmt_ee (code, mode, before, val);
+ if (after)
+ emit_insn (gen_rtx_SET (VOIDmode, after, copy_rtx (x)));
+ emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
+
+ emit_store_conditional (mode, cond, mem, scratch);
+
+ x = gen_rtx_EQ (DImode, cond, const0_rtx);
+ emit_unlikely_jump (x, label);
+
+ emit_insn (gen_memory_barrier ());
}
-/* Return or create a memory slot containing the gp value for the current
- function. Needed only if TARGET_LD_BUGGY_LDGP. */
+/* Expand a compare and swap operation. */
-rtx
-alpha_gp_save_rtx (void)
+void
+alpha_split_compare_and_swap (rtx retval, rtx mem, rtx oldval, rtx newval,
+ rtx scratch)
{
- rtx seq, m = cfun->machine->gp_save_rtx;
+ enum machine_mode mode = GET_MODE (mem);
+ rtx label1, label2, x, cond = gen_lowpart (DImode, scratch);
- if (m == NULL)
- {
- start_sequence ();
+ emit_insn (gen_memory_barrier ());
- m = assign_stack_local (DImode, UNITS_PER_WORD, BITS_PER_WORD);
- m = validize_mem (m);
- emit_move_insn (m, pic_offset_table_rtx);
+ label1 = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ label2 = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ emit_label (XEXP (label1, 0));
- seq = get_insns ();
- end_sequence ();
- emit_insn_after (seq, entry_of_function ());
+ emit_load_locked (mode, retval, mem);
- cfun->machine->gp_save_rtx = m;
+ x = gen_lowpart (DImode, retval);
+ if (oldval == const0_rtx)
+ x = gen_rtx_NE (DImode, x, const0_rtx);
+ else
+ {
+ x = gen_rtx_EQ (DImode, x, oldval);
+ emit_insn (gen_rtx_SET (VOIDmode, cond, x));
+ x = gen_rtx_EQ (DImode, cond, const0_rtx);
}
+ emit_unlikely_jump (x, label2);
- return m;
+ emit_move_insn (scratch, newval);
+ emit_store_conditional (mode, cond, mem, scratch);
+
+ x = gen_rtx_EQ (DImode, cond, const0_rtx);
+ emit_unlikely_jump (x, label1);
+
+ emit_insn (gen_memory_barrier ());
+ emit_label (XEXP (label2, 0));
}
-static int
-alpha_ra_ever_killed (void)
+void
+alpha_expand_compare_and_swap_12 (rtx dst, rtx mem, rtx oldval, rtx newval)
{
- rtx top;
+ enum machine_mode mode = GET_MODE (mem);
+ rtx addr, align, wdst;
+ rtx (*fn5) (rtx, rtx, rtx, rtx, rtx);
- if (!has_hard_reg_initial_val (Pmode, REG_RA))
- return regs_ever_live[REG_RA];
+ addr = force_reg (DImode, XEXP (mem, 0));
+ align = expand_simple_binop (Pmode, AND, addr, GEN_INT (-8),
+ NULL_RTX, 1, OPTAB_DIRECT);
- push_topmost_sequence ();
- top = get_insns ();
- pop_topmost_sequence ();
+ oldval = convert_modes (DImode, mode, oldval, 1);
+ newval = emit_insxl (mode, newval, addr);
+
+ wdst = gen_reg_rtx (DImode);
+ if (mode == QImode)
+ fn5 = gen_sync_compare_and_swapqi_1;
+ else
+ fn5 = gen_sync_compare_and_swaphi_1;
+ emit_insn (fn5 (wdst, addr, oldval, newval, align));
+
+ emit_move_insn (dst, gen_lowpart (mode, wdst));
+}
+
+void
+alpha_split_compare_and_swap_12 (enum machine_mode mode, rtx dest, rtx addr,
+ rtx oldval, rtx newval, rtx align,
+ rtx scratch, rtx cond)
+{
+ rtx label1, label2, mem, width, mask, x;
+
+ mem = gen_rtx_MEM (DImode, align);
+ MEM_VOLATILE_P (mem) = 1;
+
+ emit_insn (gen_memory_barrier ());
+ label1 = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ label2 = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ emit_label (XEXP (label1, 0));
+
+ emit_load_locked (DImode, scratch, mem);
+
+ width = GEN_INT (GET_MODE_BITSIZE (mode));
+ mask = GEN_INT (mode == QImode ? 0xff : 0xffff);
+ if (WORDS_BIG_ENDIAN)
+ emit_insn (gen_extxl_be (dest, scratch, width, addr));
+ else
+ emit_insn (gen_extxl_le (dest, scratch, width, addr));
+
+ if (oldval == const0_rtx)
+ x = gen_rtx_NE (DImode, dest, const0_rtx);
+ else
+ {
+ x = gen_rtx_EQ (DImode, dest, oldval);
+ emit_insn (gen_rtx_SET (VOIDmode, cond, x));
+ x = gen_rtx_EQ (DImode, cond, const0_rtx);
+ }
+ emit_unlikely_jump (x, label2);
+
+ if (WORDS_BIG_ENDIAN)
+ emit_insn (gen_mskxl_be (scratch, scratch, mask, addr));
+ else
+ emit_insn (gen_mskxl_le (scratch, scratch, mask, addr));
+ emit_insn (gen_iordi3 (scratch, scratch, newval));
+
+ emit_store_conditional (DImode, scratch, mem, scratch);
+
+ x = gen_rtx_EQ (DImode, scratch, const0_rtx);
+ emit_unlikely_jump (x, label1);
+
+ emit_insn (gen_memory_barrier ());
+ emit_label (XEXP (label2, 0));
+}
+
+/* Expand an atomic exchange operation. */
+
+void
+alpha_split_lock_test_and_set (rtx retval, rtx mem, rtx val, rtx scratch)
+{
+ enum machine_mode mode = GET_MODE (mem);
+ rtx label, x, cond = gen_lowpart (DImode, scratch);
+
+ emit_insn (gen_memory_barrier ());
+
+ label = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ emit_label (XEXP (label, 0));
+
+ emit_load_locked (mode, retval, mem);
+ emit_move_insn (scratch, val);
+ emit_store_conditional (mode, cond, mem, scratch);
+
+ x = gen_rtx_EQ (DImode, cond, const0_rtx);
+ emit_unlikely_jump (x, label);
+}
+
+void
+alpha_expand_lock_test_and_set_12 (rtx dst, rtx mem, rtx val)
+{
+ enum machine_mode mode = GET_MODE (mem);
+ rtx addr, align, wdst;
+ rtx (*fn4) (rtx, rtx, rtx, rtx);
+
+ /* Force the address into a register. */
+ addr = force_reg (DImode, XEXP (mem, 0));
+
+ /* Align it to a multiple of 8. */
+ align = expand_simple_binop (Pmode, AND, addr, GEN_INT (-8),
+ NULL_RTX, 1, OPTAB_DIRECT);
+
+ /* Insert val into the correct byte location within the word. */
+ val = emit_insxl (mode, val, addr);
+
+ wdst = gen_reg_rtx (DImode);
+ if (mode == QImode)
+ fn4 = gen_sync_lock_test_and_setqi_1;
+ else
+ fn4 = gen_sync_lock_test_and_sethi_1;
+ emit_insn (fn4 (wdst, addr, val, align));
+
+ emit_move_insn (dst, gen_lowpart (mode, wdst));
+}
+
+void
+alpha_split_lock_test_and_set_12 (enum machine_mode mode, rtx dest, rtx addr,
+ rtx val, rtx align, rtx scratch)
+{
+ rtx label, mem, width, mask, x;
+
+ mem = gen_rtx_MEM (DImode, align);
+ MEM_VOLATILE_P (mem) = 1;
+
+ emit_insn (gen_memory_barrier ());
+ label = gen_rtx_LABEL_REF (DImode, gen_label_rtx ());
+ emit_label (XEXP (label, 0));
+
+ emit_load_locked (DImode, scratch, mem);
+
+ width = GEN_INT (GET_MODE_BITSIZE (mode));
+ mask = GEN_INT (mode == QImode ? 0xff : 0xffff);
+ if (WORDS_BIG_ENDIAN)
+ {
+ emit_insn (gen_extxl_be (dest, scratch, width, addr));
+ emit_insn (gen_mskxl_be (scratch, scratch, mask, addr));
+ }
+ else
+ {
+ emit_insn (gen_extxl_le (dest, scratch, width, addr));
+ emit_insn (gen_mskxl_le (scratch, scratch, mask, addr));
+ }
+ emit_insn (gen_iordi3 (scratch, scratch, val));
+
+ emit_store_conditional (DImode, scratch, mem, scratch);
+
+ x = gen_rtx_EQ (DImode, scratch, const0_rtx);
+ emit_unlikely_jump (x, label);
+}
+\f
+/* 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
+alpha_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
+{
+ enum attr_type insn_type, dep_insn_type;
+
+ /* If the dependence is an anti-dependence, there is no cost. For an
+ output dependence, there is sometimes a cost, but it doesn't seem
+ worth handling those few cases. */
+ if (REG_NOTE_KIND (link) != 0)
+ return cost;
+
+ /* If we can't recognize the insns, we can't really do anything. */
+ if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
+ return cost;
+
+ insn_type = get_attr_type (insn);
+ dep_insn_type = get_attr_type (dep_insn);
+
+ /* Bring in the user-defined memory latency. */
+ if (dep_insn_type == TYPE_ILD
+ || dep_insn_type == TYPE_FLD
+ || dep_insn_type == TYPE_LDSYM)
+ cost += alpha_memory_latency-1;
+
+ /* Everything else handled in DFA bypasses now. */
+
+ return cost;
+}
+
+/* The number of instructions that can be issued per cycle. */
+
+static int
+alpha_issue_rate (void)
+{
+ return (alpha_tune == PROCESSOR_EV4 ? 2 : 4);
+}
+
+/* How many alternative schedules to try. This should be as wide as the
+ scheduling freedom in the DFA, but no wider. Making this value too
+ large results extra work for the scheduler.
+
+ For EV4, loads can be issued to either IB0 or IB1, thus we have 2
+ alternative schedules. For EV5, we can choose between E0/E1 and
+ FA/FM. For EV6, an arithmetic insn can be issued to U0/U1/L0/L1. */
+
+static int
+alpha_multipass_dfa_lookahead (void)
+{
+ return (alpha_tune == PROCESSOR_EV6 ? 4 : 2);
+}
+\f
+/* Machine-specific function data. */
+
+struct machine_function GTY(())
+{
+ /* For unicosmk. */
+ /* List of call information words for calls from this function. */
+ struct rtx_def *first_ciw;
+ struct rtx_def *last_ciw;
+ int ciw_count;
+
+ /* List of deferred case vectors. */
+ struct rtx_def *addr_list;
+
+ /* For OSF. */
+ const char *some_ld_name;
+
+ /* For TARGET_LD_BUGGY_LDGP. */
+ struct rtx_def *gp_save_rtx;
+};
+
+/* How to allocate a 'struct machine_function'. */
+
+static struct machine_function *
+alpha_init_machine_status (void)
+{
+ return ((struct machine_function *)
+ ggc_alloc_cleared (sizeof (struct machine_function)));
+}
+
+/* Functions to save and restore alpha_return_addr_rtx. */
+
+/* Start the ball rolling with RETURN_ADDR_RTX. */
+
+rtx
+alpha_return_addr (int count, rtx frame ATTRIBUTE_UNUSED)
+{
+ if (count != 0)
+ return const0_rtx;
+
+ return get_hard_reg_initial_val (Pmode, REG_RA);
+}
+
+/* Return or create a memory slot containing the gp value for the current
+ function. Needed only if TARGET_LD_BUGGY_LDGP. */
+
+rtx
+alpha_gp_save_rtx (void)
+{
+ rtx seq, m = cfun->machine->gp_save_rtx;
+
+ if (m == NULL)
+ {
+ start_sequence ();
+
+ m = assign_stack_local (DImode, UNITS_PER_WORD, BITS_PER_WORD);
+ m = validize_mem (m);
+ emit_move_insn (m, pic_offset_table_rtx);
+
+ seq = get_insns ();
+ end_sequence ();
+ emit_insn_after (seq, entry_of_function ());
+
+ cfun->machine->gp_save_rtx = m;
+ }
+
+ return m;
+}
+
+static int
+alpha_ra_ever_killed (void)
+{
+ rtx top;
+
+ if (!has_hard_reg_initial_val (Pmode, REG_RA))
+ return regs_ever_live[REG_RA];
+
+ push_topmost_sequence ();
+ top = get_insns ();
+ pop_topmost_sequence ();
return reg_set_between_p (gen_rtx_REG (Pmode, REG_RA), top, NULL_RTX);
}
case ALPHA_FPTM_SU:
case ALPHA_FPTM_SUI:
return "sv";
+ default:
+ gcc_unreachable ();
}
- break;
case TRAP_SUFFIX_V_SV_SVI:
switch (alpha_fptm)
return "sv";
case ALPHA_FPTM_SUI:
return "svi";
+ default:
+ gcc_unreachable ();
}
break;
return "su";
case ALPHA_FPTM_SUI:
return "sui";
+ default:
+ gcc_unreachable ();
}
break;
+
+ default:
+ gcc_unreachable ();
}
- abort ();
+ gcc_unreachable ();
}
/* Return the rounding mode suffix applicable to the current
{
case ALPHA_FPRM_NORM:
return NULL;
- case ALPHA_FPRM_MINF:
+ case ALPHA_FPRM_MINF:
return "m";
case ALPHA_FPRM_CHOP:
return "c";
case ALPHA_FPRM_DYN:
return "d";
+ default:
+ gcc_unreachable ();
}
break;
case ROUND_SUFFIX_C:
return "c";
+
+ default:
+ gcc_unreachable ();
}
- abort ();
+ gcc_unreachable ();
}
/* Locate some local-dynamic symbol still in use by this function
&& for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
return cfun->machine->some_ld_name;
- abort ();
+ gcc_unreachable ();
}
/* Print an operand. Recognize special options, documented below. */
}
break;
+ case 'j':
+ {
+ const char *lituse;
+
+#ifdef HAVE_AS_JSRDIRECT_RELOCS
+ lituse = "lituse_jsrdirect";
+#else
+ lituse = "lituse_jsr";
+#endif
+
+ gcc_assert (INTVAL (x) != 0);
+ fprintf (file, "\t\t!%s!%d", lituse, (int) INTVAL (x));
+ }
+ break;
case 'r':
/* If this operand is the constant zero, write it as "$31". */
if (GET_CODE (x) == REG)
if (GET_CODE (x) != CONST_INT
|| (unsigned HOST_WIDE_INT) INTVAL (x) >= (WORDS_BIG_ENDIAN
? 56
- : 64)
+ : 64)
|| (INTVAL (x) & 7) != 0)
output_operand_lossage ("invalid %%s value");
if (offset)
fprintf (file, "+" HOST_WIDE_INT_PRINT_DEC, offset);
-
+
addr = XEXP (addr, 0);
- if (GET_CODE (addr) == REG)
- basereg = REGNO (addr);
- else if (GET_CODE (addr) == SUBREG
- && GET_CODE (SUBREG_REG (addr)) == REG)
- basereg = subreg_regno (addr);
- else
- abort ();
+ switch (GET_CODE (addr))
+ {
+ case REG:
+ basereg = REGNO (addr);
+ break;
+
+ case SUBREG:
+ basereg = subreg_regno (addr);
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
fprintf (file, "($%d)\t\t!%s", basereg,
(basereg == 29 ? reloc16 : reloclo));
return;
}
- if (GET_CODE (addr) == REG)
- basereg = REGNO (addr);
- else if (GET_CODE (addr) == SUBREG
- && GET_CODE (SUBREG_REG (addr)) == REG)
- basereg = subreg_regno (addr);
- else if (GET_CODE (addr) == CONST_INT)
- offset = INTVAL (addr);
+ switch (GET_CODE (addr))
+ {
+ case REG:
+ basereg = REGNO (addr);
+ break;
+
+ case SUBREG:
+ basereg = subreg_regno (addr);
+ break;
+
+ case CONST_INT:
+ offset = INTVAL (addr);
+ break;
#if TARGET_ABI_OPEN_VMS
- else if (GET_CODE (addr) == SYMBOL_REF)
- {
+ case SYMBOL_REF:
fprintf (file, "%s", XSTR (addr, 0));
return;
- }
- else if (GET_CODE (addr) == CONST
- && GET_CODE (XEXP (addr, 0)) == PLUS
- && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF)
- {
+
+ case CONST:
+ gcc_assert (GET_CODE (XEXP (addr, 0)) == PLUS
+ && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF);
fprintf (file, "%s+" HOST_WIDE_INT_PRINT_DEC,
XSTR (XEXP (XEXP (addr, 0), 0), 0),
INTVAL (XEXP (XEXP (addr, 0), 1)));
return;
- }
+
#endif
-
- else
- abort ();
+ default:
+ gcc_unreachable ();
+ }
fprintf (file, HOST_WIDE_INT_PRINT_DEC "($%d)", offset, basereg);
}
code. CXT is an RTX for the static chain value for the function.
The three offset parameters are for the individual template's
- layout. A JMPOFS < 0 indicates that the trampoline does not
+ layout. A JMPOFS < 0 indicates that the trampoline does not
contain instructions at all.
We assume here that a function will be called many more times than
temp = expand_binop (DImode, sub_optab, fnaddr, temp, temp, 1,
OPTAB_WIDEN);
temp = expand_shift (RSHIFT_EXPR, Pmode, temp,
- build_int_cst (NULL_TREE, 2, 0), NULL_RTX, 1);
+ build_int_cst (NULL_TREE, 2), NULL_RTX, 1);
temp = expand_and (SImode, gen_lowpart (SImode, temp),
GEN_INT (0x3fff), 0);
#ifdef ENABLE_CHECKING
/* With alpha_split_complex_arg, we shouldn't see any raw complex
values here. */
- if (COMPLEX_MODE_P (mode))
- abort ();
+ gcc_assert (!COMPLEX_MODE_P (mode));
#endif
/* Set up defaults for FP operands passed in FP registers, and
return gen_rtx_REG (mode, num_args + basereg);
}
+static int
+alpha_arg_partial_bytes (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
+ enum machine_mode mode ATTRIBUTE_UNUSED,
+ tree type ATTRIBUTE_UNUSED,
+ bool named ATTRIBUTE_UNUSED)
+{
+ int words = 0;
+
+#if TARGET_ABI_OPEN_VMS
+ if (cum->num_args < 6
+ && 6 < cum->num_args + ALPHA_ARG_SIZE (mode, type, named))
+ words = 6 - cum->num_args;
+#elif TARGET_ABI_UNICOSMK
+ /* Never any split arguments. */
+#elif TARGET_ABI_OSF
+ if (*cum < 6 && 6 < *cum + ALPHA_ARG_SIZE (mode, type, named))
+ words = 6 - *cum;
+#else
+#error Unhandled ABI
+#endif
+
+ return words * UNITS_PER_WORD;
+}
+
+
/* Return true if TYPE must be returned in memory, instead of in registers. */
static bool
break;
default:
- /* ??? We get called on all sorts of random stuff from
- aggregate_value_p. We can't abort, but it's not clear
- what's safe to return. Pretend it's a struct I guess. */
+ /* ??? We get called on all sorts of random stuff from
+ aggregate_value_p. We must return something, but it's not
+ clear what's safe to return. Pretend it's a struct I
+ guess. */
return true;
}
unsigned int regnum, dummy;
enum mode_class class;
-#ifdef ENABLE_CHECKING
- if (valtype && alpha_return_in_memory (valtype, func))
- abort ();
-#endif
+ gcc_assert (!valtype || !alpha_return_in_memory (valtype, func));
if (valtype)
mode = TYPE_MODE (valtype);
}
default:
- abort ();
+ gcc_unreachable ();
}
return gen_rtx_REG (mode, regnum);
}
-/* TCmode complex values are passed by invisible reference. We
+/* TCmode complex values are passed by invisible reference. We
should not split these values. */
static bool
TYPE_FIELDS (record) = base;
layout_type (record);
+ va_list_gpr_counter_field = ofs;
return record;
}
-/* Perform any needed actions needed for a function that is receiving a
+#if TARGET_ABI_OSF
+/* Helper function for alpha_stdarg_optimize_hook. Skip over casts
+ and constant additions. */
+
+static tree
+va_list_skip_additions (tree lhs)
+{
+ tree rhs, stmt;
+
+ if (TREE_CODE (lhs) != SSA_NAME)
+ return lhs;
+
+ for (;;)
+ {
+ stmt = SSA_NAME_DEF_STMT (lhs);
+
+ if (TREE_CODE (stmt) == PHI_NODE)
+ return stmt;
+
+ if (TREE_CODE (stmt) != MODIFY_EXPR
+ || TREE_OPERAND (stmt, 0) != lhs)
+ return lhs;
+
+ rhs = TREE_OPERAND (stmt, 1);
+ if (TREE_CODE (rhs) == WITH_SIZE_EXPR)
+ rhs = TREE_OPERAND (rhs, 0);
+
+ if ((TREE_CODE (rhs) != NOP_EXPR
+ && TREE_CODE (rhs) != CONVERT_EXPR
+ && (TREE_CODE (rhs) != PLUS_EXPR
+ || TREE_CODE (TREE_OPERAND (rhs, 1)) != INTEGER_CST
+ || !host_integerp (TREE_OPERAND (rhs, 1), 1)))
+ || TREE_CODE (TREE_OPERAND (rhs, 0)) != SSA_NAME)
+ return rhs;
+
+ lhs = TREE_OPERAND (rhs, 0);
+ }
+}
+
+/* Check if LHS = RHS statement is
+ LHS = *(ap.__base + ap.__offset + cst)
+ or
+ LHS = *(ap.__base
+ + ((ap.__offset + cst <= 47)
+ ? ap.__offset + cst - 48 : ap.__offset + cst) + cst2).
+ If the former, indicate that GPR registers are needed,
+ if the latter, indicate that FPR registers are needed.
+ On alpha, cfun->va_list_gpr_size is used as size of the needed
+ regs and cfun->va_list_fpr_size is a bitmask, bit 0 set if
+ GPR registers are needed and bit 1 set if FPR registers are needed.
+ Return true if va_list references should not be scanned for the current
+ statement. */
+
+static bool
+alpha_stdarg_optimize_hook (struct stdarg_info *si, tree lhs, tree rhs)
+{
+ tree base, offset, arg1, arg2;
+ int offset_arg = 1;
+
+ if (TREE_CODE (rhs) != INDIRECT_REF
+ || TREE_CODE (TREE_OPERAND (rhs, 0)) != SSA_NAME)
+ return false;
+
+ lhs = va_list_skip_additions (TREE_OPERAND (rhs, 0));
+ if (lhs == NULL_TREE
+ || TREE_CODE (lhs) != PLUS_EXPR)
+ return false;
+
+ base = TREE_OPERAND (lhs, 0);
+ if (TREE_CODE (base) == SSA_NAME)
+ base = va_list_skip_additions (base);
+
+ if (TREE_CODE (base) != COMPONENT_REF
+ || TREE_OPERAND (base, 1) != TYPE_FIELDS (va_list_type_node))
+ {
+ base = TREE_OPERAND (lhs, 0);
+ if (TREE_CODE (base) == SSA_NAME)
+ base = va_list_skip_additions (base);
+
+ if (TREE_CODE (base) != COMPONENT_REF
+ || TREE_OPERAND (base, 1) != TYPE_FIELDS (va_list_type_node))
+ return false;
+
+ offset_arg = 0;
+ }
+
+ base = get_base_address (base);
+ if (TREE_CODE (base) != VAR_DECL
+ || !bitmap_bit_p (si->va_list_vars, DECL_UID (base)))
+ return false;
+
+ offset = TREE_OPERAND (lhs, offset_arg);
+ if (TREE_CODE (offset) == SSA_NAME)
+ offset = va_list_skip_additions (offset);
+
+ if (TREE_CODE (offset) == PHI_NODE)
+ {
+ HOST_WIDE_INT sub;
+
+ if (PHI_NUM_ARGS (offset) != 2)
+ goto escapes;
+
+ arg1 = va_list_skip_additions (PHI_ARG_DEF (offset, 0));
+ arg2 = va_list_skip_additions (PHI_ARG_DEF (offset, 1));
+ if (TREE_CODE (arg2) != MINUS_EXPR && TREE_CODE (arg2) != PLUS_EXPR)
+ {
+ tree tem = arg1;
+ arg1 = arg2;
+ arg2 = tem;
+
+ if (TREE_CODE (arg2) != MINUS_EXPR && TREE_CODE (arg2) != PLUS_EXPR)
+ goto escapes;
+ }
+ if (!host_integerp (TREE_OPERAND (arg2, 1), 0))
+ goto escapes;
+
+ sub = tree_low_cst (TREE_OPERAND (arg2, 1), 0);
+ if (TREE_CODE (arg2) == MINUS_EXPR)
+ sub = -sub;
+ if (sub < -48 || sub > -32)
+ goto escapes;
+
+ arg2 = va_list_skip_additions (TREE_OPERAND (arg2, 0));
+ if (arg1 != arg2)
+ goto escapes;
+
+ if (TREE_CODE (arg1) == SSA_NAME)
+ arg1 = va_list_skip_additions (arg1);
+
+ if (TREE_CODE (arg1) != COMPONENT_REF
+ || TREE_OPERAND (arg1, 1) != va_list_gpr_counter_field
+ || get_base_address (arg1) != base)
+ goto escapes;
+
+ /* Need floating point regs. */
+ cfun->va_list_fpr_size |= 2;
+ }
+ else if (TREE_CODE (offset) != COMPONENT_REF
+ || TREE_OPERAND (offset, 1) != va_list_gpr_counter_field
+ || get_base_address (offset) != base)
+ goto escapes;
+ else
+ /* Need general regs. */
+ cfun->va_list_fpr_size |= 1;
+ return false;
+
+escapes:
+ si->va_list_escapes = true;
+ return false;
+}
+#endif
+
+/* Perform any needed actions needed for a function that is receiving a
variable number of arguments. */
static void
-alpha_setup_incoming_varargs (CUMULATIVE_ARGS *pcum,
- enum machine_mode mode ATTRIBUTE_UNUSED,
- tree type ATTRIBUTE_UNUSED,
- int *pretend_size, int no_rtl)
+alpha_setup_incoming_varargs (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ tree type, int *pretend_size, int no_rtl)
{
+ CUMULATIVE_ARGS cum = *pcum;
+
+ /* Skip the current argument. */
+ FUNCTION_ARG_ADVANCE (cum, mode, type, 1);
+
#if TARGET_ABI_UNICOSMK
/* On Unicos/Mk, the standard subroutine __T3E_MISMATCH stores all register
arguments on the stack. Unfortunately, it doesn't always store the first
one (i.e. the one that arrives in $16 or $f16). This is not a problem
with stdargs as we always have at least one named argument there. */
- int num_reg_words = pcum->num_reg_words;
- if (num_reg_words < 6)
+ if (cum.num_reg_words < 6)
{
if (!no_rtl)
{
- emit_insn (gen_umk_mismatch_args (GEN_INT (num_reg_words + 1)));
+ emit_insn (gen_umk_mismatch_args (GEN_INT (cum.num_reg_words)));
emit_insn (gen_arg_home_umk ());
}
*pretend_size = 0;
This is not only because we won't need the space, but because AP
includes the current_pretend_args_size and we don't want to mess up
any ap-relative addresses already made. */
- if (pcum->num_args < 6)
+ if (cum.num_args < 6)
{
if (!no_rtl)
{
not the most efficient way to implement varargs with just one register
class, but it isn't worth doing anything more efficient in this rare
case. */
- CUMULATIVE_ARGS cum = *pcum;
-
if (cum >= 6)
return;
if (!no_rtl)
{
- int set = get_varargs_alias_set ();
+ int count, set = get_varargs_alias_set ();
rtx tmp;
- tmp = gen_rtx_MEM (BLKmode,
- plus_constant (virtual_incoming_args_rtx,
- (cum + 6) * UNITS_PER_WORD));
- set_mem_alias_set (tmp, set);
- move_block_from_reg (16 + cum, tmp, 6 - cum);
-
- tmp = gen_rtx_MEM (BLKmode,
- plus_constant (virtual_incoming_args_rtx,
- cum * UNITS_PER_WORD));
- set_mem_alias_set (tmp, set);
- move_block_from_reg (16 + (TARGET_FPREGS ? 32 : 0) + cum, tmp,
- 6 - cum);
+ count = cfun->va_list_gpr_size / UNITS_PER_WORD;
+ if (count > 6 - cum)
+ count = 6 - cum;
+
+ /* Detect whether integer registers or floating-point registers
+ are needed by the detected va_arg statements. See above for
+ how these values are computed. Note that the "escape" value
+ is VA_LIST_MAX_FPR_SIZE, which is 255, which has both of
+ these bits set. */
+ gcc_assert ((VA_LIST_MAX_FPR_SIZE & 3) == 3);
+
+ if (cfun->va_list_fpr_size & 1)
+ {
+ tmp = gen_rtx_MEM (BLKmode,
+ plus_constant (virtual_incoming_args_rtx,
+ (cum + 6) * UNITS_PER_WORD));
+ MEM_NOTRAP_P (tmp) = 1;
+ set_mem_alias_set (tmp, set);
+ move_block_from_reg (16 + cum, tmp, count);
+ }
+
+ if (cfun->va_list_fpr_size & 2)
+ {
+ tmp = gen_rtx_MEM (BLKmode,
+ plus_constant (virtual_incoming_args_rtx,
+ cum * UNITS_PER_WORD));
+ MEM_NOTRAP_P (tmp) = 1;
+ set_mem_alias_set (tmp, set);
+ move_block_from_reg (16 + cum + TARGET_FPREGS*32, tmp, count);
+ }
}
*pretend_size = 12 * UNITS_PER_WORD;
#endif
in order to account for the integer arg registers which are counted
in argsize above, but which are not actually stored on the stack.
Must further be careful here about structures straddling the last
- integer argument register; that futzes with pretend_args_size,
+ integer argument register; that futzes with pretend_args_size,
which changes the meaning of AP. */
- if (NUM_ARGS <= 6)
+ if (NUM_ARGS < 6)
offset = TARGET_ABI_OPEN_VMS ? UNITS_PER_WORD : 6 * UNITS_PER_WORD;
else
offset = -6 * UNITS_PER_WORD + current_function_pretend_args_size;
{
nextarg = plus_constant (nextarg, offset);
nextarg = plus_constant (nextarg, NUM_ARGS * UNITS_PER_WORD);
- t = build (MODIFY_EXPR, TREE_TYPE (valist), valist,
- make_tree (ptr_type_node, nextarg));
+ t = build2 (MODIFY_EXPR, TREE_TYPE (valist), valist,
+ make_tree (ptr_type_node, nextarg));
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
base_field = TYPE_FIELDS (TREE_TYPE (valist));
offset_field = TREE_CHAIN (base_field);
- base_field = build (COMPONENT_REF, TREE_TYPE (base_field),
- valist, base_field, NULL_TREE);
- offset_field = build (COMPONENT_REF, TREE_TYPE (offset_field),
- valist, offset_field, NULL_TREE);
+ base_field = build3 (COMPONENT_REF, TREE_TYPE (base_field),
+ valist, base_field, NULL_TREE);
+ offset_field = build3 (COMPONENT_REF, TREE_TYPE (offset_field),
+ valist, offset_field, NULL_TREE);
t = make_tree (ptr_type_node, virtual_incoming_args_rtx);
- t = build (PLUS_EXPR, ptr_type_node, t,
- build_int_cst (NULL_TREE, offset, 0));
- t = build (MODIFY_EXPR, TREE_TYPE (base_field), base_field, t);
+ t = build2 (PLUS_EXPR, ptr_type_node, t,
+ build_int_cst (NULL_TREE, offset));
+ t = build2 (MODIFY_EXPR, TREE_TYPE (base_field), base_field, t);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
- t = build_int_cst (NULL_TREE, NUM_ARGS * UNITS_PER_WORD, 0);
- t = build (MODIFY_EXPR, TREE_TYPE (offset_field), offset_field, t);
+ t = build_int_cst (NULL_TREE, NUM_ARGS * UNITS_PER_WORD);
+ t = build2 (MODIFY_EXPR, TREE_TYPE (offset_field), offset_field, t);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
}
reserved for the registers. */
if (targetm.calls.must_pass_in_stack (TYPE_MODE (type), type))
{
- t = build_int_cst (TREE_TYPE (offset), 6*8, 0);
- t = build (MODIFY_EXPR, TREE_TYPE (offset), offset,
- build (MAX_EXPR, TREE_TYPE (offset), offset, t));
+ t = build_int_cst (TREE_TYPE (offset), 6*8);
+ t = build2 (MODIFY_EXPR, TREE_TYPE (offset), offset,
+ build2 (MAX_EXPR, TREE_TYPE (offset), offset, t));
gimplify_and_add (t, pre_p);
}
imag_part = alpha_gimplify_va_arg_1 (TREE_TYPE (type), base,
offset, pre_p);
- return build (COMPLEX_EXPR, type, real_temp, imag_part);
+ return build2 (COMPLEX_EXPR, type, real_temp, imag_part);
}
else if (TREE_CODE (type) == REAL_TYPE)
{
tree fpaddend, cond, fourtyeight;
- fourtyeight = build_int_cst (TREE_TYPE (addend), 6*8, 0);
- fpaddend = fold (build (MINUS_EXPR, TREE_TYPE (addend),
- addend, fourtyeight));
- cond = fold (build (LT_EXPR, boolean_type_node, addend, fourtyeight));
- addend = fold (build (COND_EXPR, TREE_TYPE (addend), cond,
- fpaddend, addend));
+ fourtyeight = build_int_cst (TREE_TYPE (addend), 6*8);
+ fpaddend = fold_build2 (MINUS_EXPR, TREE_TYPE (addend),
+ addend, fourtyeight);
+ cond = fold_build2 (LT_EXPR, boolean_type_node, addend, fourtyeight);
+ addend = fold_build3 (COND_EXPR, TREE_TYPE (addend), cond,
+ fpaddend, addend);
}
/* Build the final address and force that value into a temporary. */
- addr = build (PLUS_EXPR, ptr_type, fold_convert (ptr_type, base),
- fold_convert (ptr_type, addend));
+ addr = build2 (PLUS_EXPR, ptr_type, fold_convert (ptr_type, base),
+ fold_convert (ptr_type, addend));
internal_post = NULL;
gimplify_expr (&addr, pre_p, &internal_post, is_gimple_val, fb_rvalue);
append_to_statement_list (internal_post, pre_p);
t = size_binop (MULT_EXPR, t, size_int (8));
}
t = fold_convert (TREE_TYPE (offset), t);
- t = build (MODIFY_EXPR, void_type_node, offset,
- build (PLUS_EXPR, TREE_TYPE (offset), offset, t));
+ t = build2 (MODIFY_EXPR, void_type_node, offset,
+ build2 (PLUS_EXPR, TREE_TYPE (offset), offset, t));
gimplify_and_add (t, pre_p);
- return build_fold_indirect_ref (addr);
+ return build_va_arg_indirect_ref (addr);
}
static tree
base_field = TYPE_FIELDS (va_list_type_node);
offset_field = TREE_CHAIN (base_field);
- base_field = build (COMPONENT_REF, TREE_TYPE (base_field),
- valist, base_field, NULL_TREE);
- offset_field = build (COMPONENT_REF, TREE_TYPE (offset_field),
- valist, offset_field, NULL_TREE);
+ base_field = build3 (COMPONENT_REF, TREE_TYPE (base_field),
+ valist, base_field, NULL_TREE);
+ offset_field = build3 (COMPONENT_REF, TREE_TYPE (offset_field),
+ valist, offset_field, NULL_TREE);
/* Pull the fields of the structure out into temporaries. Since we never
modify the base field, we can use a formal temporary. Sign-extend the
r = alpha_gimplify_va_arg_1 (type, base, offset, pre_p);
/* Stuff the offset temporary back into its field. */
- t = build (MODIFY_EXPR, void_type_node, offset_field,
- fold_convert (TREE_TYPE (offset_field), offset));
+ t = build2 (MODIFY_EXPR, void_type_node, offset_field,
+ fold_convert (TREE_TYPE (offset_field), offset));
gimplify_and_add (t, pre_p);
if (indirect)
- r = build_fold_indirect_ref (r);
+ r = build_va_arg_indirect_ref (r);
return r;
}
CODE_FOR_builtin_unpkbw,
/* TARGET_CIX */
- CODE_FOR_builtin_cttz,
- CODE_FOR_builtin_ctlz,
- CODE_FOR_builtin_ctpop
+ CODE_FOR_ctzdi2,
+ CODE_FOR_clzdi2,
+ CODE_FOR_popcountdi2
};
struct alpha_builtin_def
const char *name;
enum alpha_builtin code;
unsigned int target_mask;
+ bool is_const;
};
static struct alpha_builtin_def const zero_arg_builtins[] = {
- { "__builtin_alpha_implver", ALPHA_BUILTIN_IMPLVER, 0 },
- { "__builtin_alpha_rpcc", ALPHA_BUILTIN_RPCC, 0 }
+ { "__builtin_alpha_implver", ALPHA_BUILTIN_IMPLVER, 0, true },
+ { "__builtin_alpha_rpcc", ALPHA_BUILTIN_RPCC, 0, false }
};
static struct alpha_builtin_def const one_arg_builtins[] = {
- { "__builtin_alpha_amask", ALPHA_BUILTIN_AMASK, 0 },
- { "__builtin_alpha_pklb", ALPHA_BUILTIN_PKLB, MASK_MAX },
- { "__builtin_alpha_pkwb", ALPHA_BUILTIN_PKWB, MASK_MAX },
- { "__builtin_alpha_unpkbl", ALPHA_BUILTIN_UNPKBL, MASK_MAX },
- { "__builtin_alpha_unpkbw", ALPHA_BUILTIN_UNPKBW, MASK_MAX },
- { "__builtin_alpha_cttz", ALPHA_BUILTIN_CTTZ, MASK_CIX },
- { "__builtin_alpha_ctlz", ALPHA_BUILTIN_CTLZ, MASK_CIX },
- { "__builtin_alpha_ctpop", ALPHA_BUILTIN_CTPOP, MASK_CIX }
+ { "__builtin_alpha_amask", ALPHA_BUILTIN_AMASK, 0, true },
+ { "__builtin_alpha_pklb", ALPHA_BUILTIN_PKLB, MASK_MAX, true },
+ { "__builtin_alpha_pkwb", ALPHA_BUILTIN_PKWB, MASK_MAX, true },
+ { "__builtin_alpha_unpkbl", ALPHA_BUILTIN_UNPKBL, MASK_MAX, true },
+ { "__builtin_alpha_unpkbw", ALPHA_BUILTIN_UNPKBW, MASK_MAX, true },
+ { "__builtin_alpha_cttz", ALPHA_BUILTIN_CTTZ, MASK_CIX, true },
+ { "__builtin_alpha_ctlz", ALPHA_BUILTIN_CTLZ, MASK_CIX, true },
+ { "__builtin_alpha_ctpop", ALPHA_BUILTIN_CTPOP, MASK_CIX, true }
};
static struct alpha_builtin_def const two_arg_builtins[] = {
- { "__builtin_alpha_cmpbge", ALPHA_BUILTIN_CMPBGE, 0 },
- { "__builtin_alpha_extbl", ALPHA_BUILTIN_EXTBL, 0 },
- { "__builtin_alpha_extwl", ALPHA_BUILTIN_EXTWL, 0 },
- { "__builtin_alpha_extll", ALPHA_BUILTIN_EXTLL, 0 },
- { "__builtin_alpha_extql", ALPHA_BUILTIN_EXTQL, 0 },
- { "__builtin_alpha_extwh", ALPHA_BUILTIN_EXTWH, 0 },
- { "__builtin_alpha_extlh", ALPHA_BUILTIN_EXTLH, 0 },
- { "__builtin_alpha_extqh", ALPHA_BUILTIN_EXTQH, 0 },
- { "__builtin_alpha_insbl", ALPHA_BUILTIN_INSBL, 0 },
- { "__builtin_alpha_inswl", ALPHA_BUILTIN_INSWL, 0 },
- { "__builtin_alpha_insll", ALPHA_BUILTIN_INSLL, 0 },
- { "__builtin_alpha_insql", ALPHA_BUILTIN_INSQL, 0 },
- { "__builtin_alpha_inswh", ALPHA_BUILTIN_INSWH, 0 },
- { "__builtin_alpha_inslh", ALPHA_BUILTIN_INSLH, 0 },
- { "__builtin_alpha_insqh", ALPHA_BUILTIN_INSQH, 0 },
- { "__builtin_alpha_mskbl", ALPHA_BUILTIN_MSKBL, 0 },
- { "__builtin_alpha_mskwl", ALPHA_BUILTIN_MSKWL, 0 },
- { "__builtin_alpha_mskll", ALPHA_BUILTIN_MSKLL, 0 },
- { "__builtin_alpha_mskql", ALPHA_BUILTIN_MSKQL, 0 },
- { "__builtin_alpha_mskwh", ALPHA_BUILTIN_MSKWH, 0 },
- { "__builtin_alpha_msklh", ALPHA_BUILTIN_MSKLH, 0 },
- { "__builtin_alpha_mskqh", ALPHA_BUILTIN_MSKQH, 0 },
- { "__builtin_alpha_umulh", ALPHA_BUILTIN_UMULH, 0 },
- { "__builtin_alpha_zap", ALPHA_BUILTIN_ZAP, 0 },
- { "__builtin_alpha_zapnot", ALPHA_BUILTIN_ZAPNOT, 0 },
- { "__builtin_alpha_minub8", ALPHA_BUILTIN_MINUB8, MASK_MAX },
- { "__builtin_alpha_minsb8", ALPHA_BUILTIN_MINSB8, MASK_MAX },
- { "__builtin_alpha_minuw4", ALPHA_BUILTIN_MINUW4, MASK_MAX },
- { "__builtin_alpha_minsw4", ALPHA_BUILTIN_MINSW4, MASK_MAX },
- { "__builtin_alpha_maxub8", ALPHA_BUILTIN_MAXUB8, MASK_MAX },
- { "__builtin_alpha_maxsb8", ALPHA_BUILTIN_MAXSB8, MASK_MAX },
- { "__builtin_alpha_maxuw4", ALPHA_BUILTIN_MAXUW4, MASK_MAX },
- { "__builtin_alpha_maxsw4", ALPHA_BUILTIN_MAXSW4, MASK_MAX },
- { "__builtin_alpha_perr", ALPHA_BUILTIN_PERR, MASK_MAX }
+ { "__builtin_alpha_cmpbge", ALPHA_BUILTIN_CMPBGE, 0, true },
+ { "__builtin_alpha_extbl", ALPHA_BUILTIN_EXTBL, 0, true },
+ { "__builtin_alpha_extwl", ALPHA_BUILTIN_EXTWL, 0, true },
+ { "__builtin_alpha_extll", ALPHA_BUILTIN_EXTLL, 0, true },
+ { "__builtin_alpha_extql", ALPHA_BUILTIN_EXTQL, 0, true },
+ { "__builtin_alpha_extwh", ALPHA_BUILTIN_EXTWH, 0, true },
+ { "__builtin_alpha_extlh", ALPHA_BUILTIN_EXTLH, 0, true },
+ { "__builtin_alpha_extqh", ALPHA_BUILTIN_EXTQH, 0, true },
+ { "__builtin_alpha_insbl", ALPHA_BUILTIN_INSBL, 0, true },
+ { "__builtin_alpha_inswl", ALPHA_BUILTIN_INSWL, 0, true },
+ { "__builtin_alpha_insll", ALPHA_BUILTIN_INSLL, 0, true },
+ { "__builtin_alpha_insql", ALPHA_BUILTIN_INSQL, 0, true },
+ { "__builtin_alpha_inswh", ALPHA_BUILTIN_INSWH, 0, true },
+ { "__builtin_alpha_inslh", ALPHA_BUILTIN_INSLH, 0, true },
+ { "__builtin_alpha_insqh", ALPHA_BUILTIN_INSQH, 0, true },
+ { "__builtin_alpha_mskbl", ALPHA_BUILTIN_MSKBL, 0, true },
+ { "__builtin_alpha_mskwl", ALPHA_BUILTIN_MSKWL, 0, true },
+ { "__builtin_alpha_mskll", ALPHA_BUILTIN_MSKLL, 0, true },
+ { "__builtin_alpha_mskql", ALPHA_BUILTIN_MSKQL, 0, true },
+ { "__builtin_alpha_mskwh", ALPHA_BUILTIN_MSKWH, 0, true },
+ { "__builtin_alpha_msklh", ALPHA_BUILTIN_MSKLH, 0, true },
+ { "__builtin_alpha_mskqh", ALPHA_BUILTIN_MSKQH, 0, true },
+ { "__builtin_alpha_umulh", ALPHA_BUILTIN_UMULH, 0, true },
+ { "__builtin_alpha_zap", ALPHA_BUILTIN_ZAP, 0, true },
+ { "__builtin_alpha_zapnot", ALPHA_BUILTIN_ZAPNOT, 0, true },
+ { "__builtin_alpha_minub8", ALPHA_BUILTIN_MINUB8, MASK_MAX, true },
+ { "__builtin_alpha_minsb8", ALPHA_BUILTIN_MINSB8, MASK_MAX, true },
+ { "__builtin_alpha_minuw4", ALPHA_BUILTIN_MINUW4, MASK_MAX, true },
+ { "__builtin_alpha_minsw4", ALPHA_BUILTIN_MINSW4, MASK_MAX, true },
+ { "__builtin_alpha_maxub8", ALPHA_BUILTIN_MAXUB8, MASK_MAX, true },
+ { "__builtin_alpha_maxsb8", ALPHA_BUILTIN_MAXSB8, MASK_MAX, true },
+ { "__builtin_alpha_maxuw4", ALPHA_BUILTIN_MAXUW4, MASK_MAX, true },
+ { "__builtin_alpha_maxsw4", ALPHA_BUILTIN_MAXSW4, MASK_MAX, true },
+ { "__builtin_alpha_perr", ALPHA_BUILTIN_PERR, MASK_MAX, true }
};
+static GTY(()) tree alpha_v8qi_u;
+static GTY(()) tree alpha_v8qi_s;
+static GTY(()) tree alpha_v4hi_u;
+static GTY(()) tree alpha_v4hi_s;
+
static void
alpha_init_builtins (void)
{
const struct alpha_builtin_def *p;
- tree ftype;
+ tree dimode_integer_type_node;
+ tree ftype, attrs[2];
size_t i;
- ftype = build_function_type (long_integer_type_node, void_list_node);
+ dimode_integer_type_node = lang_hooks.types.type_for_mode (DImode, 0);
+
+ attrs[0] = tree_cons (get_identifier ("nothrow"), NULL, NULL);
+ attrs[1] = tree_cons (get_identifier ("const"), NULL, attrs[0]);
+
+ ftype = build_function_type (dimode_integer_type_node, void_list_node);
p = zero_arg_builtins;
for (i = 0; i < ARRAY_SIZE (zero_arg_builtins); ++i, ++p)
if ((target_flags & p->target_mask) == p->target_mask)
lang_hooks.builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
- NULL, NULL_TREE);
+ NULL, attrs[p->is_const]);
- ftype = build_function_type_list (long_integer_type_node,
- long_integer_type_node, NULL_TREE);
+ ftype = build_function_type_list (dimode_integer_type_node,
+ dimode_integer_type_node, NULL_TREE);
p = one_arg_builtins;
for (i = 0; i < ARRAY_SIZE (one_arg_builtins); ++i, ++p)
if ((target_flags & p->target_mask) == p->target_mask)
lang_hooks.builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
- NULL, NULL_TREE);
+ NULL, attrs[p->is_const]);
+
+ ftype = build_function_type_list (dimode_integer_type_node,
+ dimode_integer_type_node,
+ dimode_integer_type_node, NULL_TREE);
+
+ p = two_arg_builtins;
+ for (i = 0; i < ARRAY_SIZE (two_arg_builtins); ++i, ++p)
+ if ((target_flags & p->target_mask) == p->target_mask)
+ lang_hooks.builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
+ NULL, attrs[p->is_const]);
+
+ ftype = build_function_type (ptr_type_node, void_list_node);
+ lang_hooks.builtin_function ("__builtin_thread_pointer", ftype,
+ ALPHA_BUILTIN_THREAD_POINTER, BUILT_IN_MD,
+ NULL, attrs[0]);
+
+ ftype = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
+ lang_hooks.builtin_function ("__builtin_set_thread_pointer", ftype,
+ ALPHA_BUILTIN_SET_THREAD_POINTER, BUILT_IN_MD,
+ NULL, attrs[0]);
+
+ alpha_v8qi_u = build_vector_type (unsigned_intQI_type_node, 8);
+ alpha_v8qi_s = build_vector_type (intQI_type_node, 8);
+ alpha_v4hi_u = build_vector_type (unsigned_intHI_type_node, 4);
+ alpha_v4hi_s = build_vector_type (intHI_type_node, 4);
+}
+
+/* Expand an expression EXP that calls a built-in function,
+ with result going to TARGET if that's convenient
+ (and in mode MODE if that's convenient).
+ SUBTARGET may be used as the target for computing one of EXP's operands.
+ IGNORE is nonzero if the value is to be ignored. */
+
+static rtx
+alpha_expand_builtin (tree exp, rtx target,
+ rtx subtarget ATTRIBUTE_UNUSED,
+ enum machine_mode mode ATTRIBUTE_UNUSED,
+ int ignore ATTRIBUTE_UNUSED)
+{
+#define MAX_ARGS 2
+
+ tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
+ unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
+ tree arglist = TREE_OPERAND (exp, 1);
+ enum insn_code icode;
+ rtx op[MAX_ARGS], pat;
+ int arity;
+ bool nonvoid;
+
+ if (fcode >= ALPHA_BUILTIN_max)
+ internal_error ("bad builtin fcode");
+ icode = code_for_builtin[fcode];
+ if (icode == 0)
+ internal_error ("bad builtin fcode");
+
+ nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
+
+ for (arglist = TREE_OPERAND (exp, 1), arity = 0;
+ arglist;
+ arglist = TREE_CHAIN (arglist), arity++)
+ {
+ const struct insn_operand_data *insn_op;
+
+ tree arg = TREE_VALUE (arglist);
+ if (arg == error_mark_node)
+ return NULL_RTX;
+ if (arity > MAX_ARGS)
+ return NULL_RTX;
+
+ insn_op = &insn_data[icode].operand[arity + nonvoid];
+
+ op[arity] = expand_expr (arg, NULL_RTX, insn_op->mode, 0);
+
+ if (!(*insn_op->predicate) (op[arity], insn_op->mode))
+ op[arity] = copy_to_mode_reg (insn_op->mode, op[arity]);
+ }
+
+ if (nonvoid)
+ {
+ enum machine_mode tmode = insn_data[icode].operand[0].mode;
+ if (!target
+ || GET_MODE (target) != tmode
+ || !(*insn_data[icode].operand[0].predicate) (target, tmode))
+ target = gen_reg_rtx (tmode);
+ }
+
+ switch (arity)
+ {
+ case 0:
+ pat = GEN_FCN (icode) (target);
+ break;
+ case 1:
+ if (nonvoid)
+ pat = GEN_FCN (icode) (target, op[0]);
+ else
+ pat = GEN_FCN (icode) (op[0]);
+ break;
+ case 2:
+ pat = GEN_FCN (icode) (target, op[0], op[1]);
+ break;
+ default:
+ gcc_unreachable ();
+ }
+ if (!pat)
+ return NULL_RTX;
+ emit_insn (pat);
+
+ if (nonvoid)
+ return target;
+ else
+ return const0_rtx;
+}
+
+
+/* Several bits below assume HWI >= 64 bits. This should be enforced
+ by config.gcc. */
+#if HOST_BITS_PER_WIDE_INT < 64
+# error "HOST_WIDE_INT too small"
+#endif
+
+/* Fold the builtin for the CMPBGE instruction. This is a vector comparison
+ with an 8 bit output vector. OPINT contains the integer operands; bit N
+ of OP_CONST is set if OPINT[N] is valid. */
+
+static tree
+alpha_fold_builtin_cmpbge (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ if (op_const == 3)
+ {
+ int i, val;
+ for (i = 0, val = 0; i < 8; ++i)
+ {
+ unsigned HOST_WIDE_INT c0 = (opint[0] >> (i * 8)) & 0xff;
+ unsigned HOST_WIDE_INT c1 = (opint[1] >> (i * 8)) & 0xff;
+ if (c0 >= c1)
+ val |= 1 << i;
+ }
+ return build_int_cst (long_integer_type_node, val);
+ }
+ else if (op_const == 2 && opint[1] == 0)
+ return build_int_cst (long_integer_type_node, 0xff);
+ return NULL;
+}
+
+/* Fold the builtin for the ZAPNOT instruction. This is essentially a
+ specialized form of an AND operation. Other byte manipulation instructions
+ are defined in terms of this instruction, so this is also used as a
+ subroutine for other builtins.
+
+ OP contains the tree operands; OPINT contains the extracted integer values.
+ Bit N of OP_CONST it set if OPINT[N] is valid. OP may be null if only
+ OPINT may be considered. */
+
+static tree
+alpha_fold_builtin_zapnot (tree *op, unsigned HOST_WIDE_INT opint[],
+ long op_const)
+{
+ if (op_const & 2)
+ {
+ unsigned HOST_WIDE_INT mask = 0;
+ int i;
+
+ for (i = 0; i < 8; ++i)
+ if ((opint[1] >> i) & 1)
+ mask |= (unsigned HOST_WIDE_INT)0xff << (i * 8);
+
+ if (op_const & 1)
+ return build_int_cst (long_integer_type_node, opint[0] & mask);
+
+ if (op)
+ return fold (build2 (BIT_AND_EXPR, long_integer_type_node, op[0],
+ build_int_cst (long_integer_type_node, mask)));
+ }
+ else if ((op_const & 1) && opint[0] == 0)
+ return build_int_cst (long_integer_type_node, 0);
+ return NULL;
+}
+
+/* Fold the builtins for the EXT family of instructions. */
+
+static tree
+alpha_fold_builtin_extxx (tree op[], unsigned HOST_WIDE_INT opint[],
+ long op_const, unsigned HOST_WIDE_INT bytemask,
+ bool is_high)
+{
+ long zap_const = 2;
+ tree *zap_op = NULL;
+
+ if (op_const & 2)
+ {
+ unsigned HOST_WIDE_INT loc;
+
+ loc = opint[1] & 7;
+ if (BYTES_BIG_ENDIAN)
+ loc ^= 7;
+ loc *= 8;
+
+ if (loc != 0)
+ {
+ if (op_const & 1)
+ {
+ unsigned HOST_WIDE_INT temp = opint[0];
+ if (is_high)
+ temp <<= loc;
+ else
+ temp >>= loc;
+ opint[0] = temp;
+ zap_const = 3;
+ }
+ }
+ else
+ zap_op = op;
+ }
+
+ opint[1] = bytemask;
+ return alpha_fold_builtin_zapnot (zap_op, opint, zap_const);
+}
+
+/* Fold the builtins for the INS family of instructions. */
+
+static tree
+alpha_fold_builtin_insxx (tree op[], unsigned HOST_WIDE_INT opint[],
+ long op_const, unsigned HOST_WIDE_INT bytemask,
+ bool is_high)
+{
+ if ((op_const & 1) && opint[0] == 0)
+ return build_int_cst (long_integer_type_node, 0);
+
+ if (op_const & 2)
+ {
+ unsigned HOST_WIDE_INT temp, loc, byteloc;
+ tree *zap_op = NULL;
+
+ loc = opint[1] & 7;
+ if (BYTES_BIG_ENDIAN)
+ loc ^= 7;
+ bytemask <<= loc;
+
+ temp = opint[0];
+ if (is_high)
+ {
+ byteloc = (64 - (loc * 8)) & 0x3f;
+ if (byteloc == 0)
+ zap_op = op;
+ else
+ temp >>= byteloc;
+ bytemask >>= 8;
+ }
+ else
+ {
+ byteloc = loc * 8;
+ if (byteloc == 0)
+ zap_op = op;
+ else
+ temp <<= byteloc;
+ }
+
+ opint[0] = temp;
+ opint[1] = bytemask;
+ return alpha_fold_builtin_zapnot (zap_op, opint, op_const);
+ }
+
+ return NULL;
+}
+
+static tree
+alpha_fold_builtin_mskxx (tree op[], unsigned HOST_WIDE_INT opint[],
+ long op_const, unsigned HOST_WIDE_INT bytemask,
+ bool is_high)
+{
+ if (op_const & 2)
+ {
+ unsigned HOST_WIDE_INT loc;
+
+ loc = opint[1] & 7;
+ if (BYTES_BIG_ENDIAN)
+ loc ^= 7;
+ bytemask <<= loc;
+
+ if (is_high)
+ bytemask >>= 8;
+
+ opint[1] = bytemask ^ 0xff;
+ }
+
+ return alpha_fold_builtin_zapnot (op, opint, op_const);
+}
+
+static tree
+alpha_fold_builtin_umulh (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ switch (op_const)
+ {
+ case 3:
+ {
+ unsigned HOST_WIDE_INT l;
+ HOST_WIDE_INT h;
+
+ mul_double (opint[0], 0, opint[1], 0, &l, &h);
+
+#if HOST_BITS_PER_WIDE_INT > 64
+# error fixme
+#endif
+
+ return build_int_cst (long_integer_type_node, h);
+ }
+
+ case 1:
+ opint[1] = opint[0];
+ /* FALLTHRU */
+ case 2:
+ /* Note that (X*1) >> 64 == 0. */
+ if (opint[1] == 0 || opint[1] == 1)
+ return build_int_cst (long_integer_type_node, 0);
+ break;
+ }
+ return NULL;
+}
+
+static tree
+alpha_fold_vector_minmax (enum tree_code code, tree op[], tree vtype)
+{
+ tree op0 = fold_convert (vtype, op[0]);
+ tree op1 = fold_convert (vtype, op[1]);
+ tree val = fold (build2 (code, vtype, op0, op1));
+ return fold_convert (long_integer_type_node, val);
+}
+
+static tree
+alpha_fold_builtin_perr (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp = 0;
+ int i;
+
+ if (op_const != 3)
+ return NULL;
+
+ for (i = 0; i < 8; ++i)
+ {
+ unsigned HOST_WIDE_INT a = (opint[0] >> (i * 8)) & 0xff;
+ unsigned HOST_WIDE_INT b = (opint[1] >> (i * 8)) & 0xff;
+ if (a >= b)
+ temp += a - b;
+ else
+ temp += b - a;
+ }
+
+ return build_int_cst (long_integer_type_node, temp);
+}
+
+static tree
+alpha_fold_builtin_pklb (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp;
+
+ if (op_const == 0)
+ return NULL;
+
+ temp = opint[0] & 0xff;
+ temp |= (opint[0] >> 24) & 0xff00;
+
+ return build_int_cst (long_integer_type_node, temp);
+}
+
+static tree
+alpha_fold_builtin_pkwb (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp;
+
+ if (op_const == 0)
+ return NULL;
+
+ temp = opint[0] & 0xff;
+ temp |= (opint[0] >> 8) & 0xff00;
+ temp |= (opint[0] >> 16) & 0xff0000;
+ temp |= (opint[0] >> 24) & 0xff000000;
+
+ return build_int_cst (long_integer_type_node, temp);
+}
+
+static tree
+alpha_fold_builtin_unpkbl (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp;
+
+ if (op_const == 0)
+ return NULL;
+
+ temp = opint[0] & 0xff;
+ temp |= (opint[0] & 0xff00) << 24;
+
+ return build_int_cst (long_integer_type_node, temp);
+}
+
+static tree
+alpha_fold_builtin_unpkbw (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp;
+
+ if (op_const == 0)
+ return NULL;
+
+ temp = opint[0] & 0xff;
+ temp |= (opint[0] & 0x0000ff00) << 8;
+ temp |= (opint[0] & 0x00ff0000) << 16;
+ temp |= (opint[0] & 0xff000000) << 24;
+
+ return build_int_cst (long_integer_type_node, temp);
+}
- ftype = build_function_type_list (long_integer_type_node,
- long_integer_type_node,
- long_integer_type_node, NULL_TREE);
+static tree
+alpha_fold_builtin_cttz (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp;
- p = two_arg_builtins;
- for (i = 0; i < ARRAY_SIZE (two_arg_builtins); ++i, ++p)
- if ((target_flags & p->target_mask) == p->target_mask)
- lang_hooks.builtin_function (p->name, ftype, p->code, BUILT_IN_MD,
- NULL, NULL_TREE);
+ if (op_const == 0)
+ return NULL;
- ftype = build_function_type (ptr_type_node, void_list_node);
- lang_hooks.builtin_function ("__builtin_thread_pointer", ftype,
- ALPHA_BUILTIN_THREAD_POINTER, BUILT_IN_MD,
- NULL, NULL_TREE);
+ if (opint[0] == 0)
+ temp = 64;
+ else
+ temp = exact_log2 (opint[0] & -opint[0]);
- ftype = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
- lang_hooks.builtin_function ("__builtin_set_thread_pointer", ftype,
- ALPHA_BUILTIN_SET_THREAD_POINTER, BUILT_IN_MD,
- NULL, NULL_TREE);
+ return build_int_cst (long_integer_type_node, temp);
}
-/* Expand an expression EXP that calls a built-in function,
- with result going to TARGET if that's convenient
- (and in mode MODE if that's convenient).
- SUBTARGET may be used as the target for computing one of EXP's operands.
- IGNORE is nonzero if the value is to be ignored. */
-
-static rtx
-alpha_expand_builtin (tree exp, rtx target,
- rtx subtarget ATTRIBUTE_UNUSED,
- enum machine_mode mode ATTRIBUTE_UNUSED,
- int ignore ATTRIBUTE_UNUSED)
+static tree
+alpha_fold_builtin_ctlz (unsigned HOST_WIDE_INT opint[], long op_const)
{
-#define MAX_ARGS 2
+ unsigned HOST_WIDE_INT temp;
- tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
- unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
- tree arglist = TREE_OPERAND (exp, 1);
- enum insn_code icode;
- rtx op[MAX_ARGS], pat;
- int arity;
- bool nonvoid;
+ if (op_const == 0)
+ return NULL;
- if (fcode >= ALPHA_BUILTIN_max)
- internal_error ("bad builtin fcode");
- icode = code_for_builtin[fcode];
- if (icode == 0)
- internal_error ("bad builtin fcode");
+ if (opint[0] == 0)
+ temp = 64;
+ else
+ temp = 64 - floor_log2 (opint[0]) - 1;
- nonvoid = TREE_TYPE (TREE_TYPE (fndecl)) != void_type_node;
+ return build_int_cst (long_integer_type_node, temp);
+}
- for (arglist = TREE_OPERAND (exp, 1), arity = 0;
- arglist;
- arglist = TREE_CHAIN (arglist), arity++)
- {
- const struct insn_operand_data *insn_op;
+static tree
+alpha_fold_builtin_ctpop (unsigned HOST_WIDE_INT opint[], long op_const)
+{
+ unsigned HOST_WIDE_INT temp, op;
- tree arg = TREE_VALUE (arglist);
- if (arg == error_mark_node)
- return NULL_RTX;
- if (arity > MAX_ARGS)
- return NULL_RTX;
+ if (op_const == 0)
+ return NULL;
- insn_op = &insn_data[icode].operand[arity + nonvoid];
+ op = opint[0];
+ temp = 0;
+ while (op)
+ temp++, op &= op - 1;
- op[arity] = expand_expr (arg, NULL_RTX, insn_op->mode, 0);
+ return build_int_cst (long_integer_type_node, temp);
+}
- if (!(*insn_op->predicate) (op[arity], insn_op->mode))
- op[arity] = copy_to_mode_reg (insn_op->mode, op[arity]);
- }
+/* Fold one of our builtin functions. */
- if (nonvoid)
- {
- enum machine_mode tmode = insn_data[icode].operand[0].mode;
- if (!target
- || GET_MODE (target) != tmode
- || !(*insn_data[icode].operand[0].predicate) (target, tmode))
- target = gen_reg_rtx (tmode);
- }
+static tree
+alpha_fold_builtin (tree fndecl, tree arglist, bool ignore ATTRIBUTE_UNUSED)
+{
+ tree op[MAX_ARGS], t;
+ unsigned HOST_WIDE_INT opint[MAX_ARGS];
+ long op_const = 0, arity = 0;
- switch (arity)
+ for (t = arglist; t ; t = TREE_CHAIN (t), ++arity)
{
- case 0:
- pat = GEN_FCN (icode) (target);
- break;
- case 1:
- if (nonvoid)
- pat = GEN_FCN (icode) (target, op[0]);
- else
- pat = GEN_FCN (icode) (op[0]);
- break;
- case 2:
- pat = GEN_FCN (icode) (target, op[0], op[1]);
- break;
+ tree arg = TREE_VALUE (t);
+ if (arg == error_mark_node)
+ return NULL;
+ if (arity >= MAX_ARGS)
+ return NULL;
+
+ op[arity] = arg;
+ opint[arity] = 0;
+ if (TREE_CODE (arg) == INTEGER_CST)
+ {
+ op_const |= 1L << arity;
+ opint[arity] = int_cst_value (arg);
+ }
+ }
+
+ switch (DECL_FUNCTION_CODE (fndecl))
+ {
+ case ALPHA_BUILTIN_CMPBGE:
+ return alpha_fold_builtin_cmpbge (opint, op_const);
+
+ case ALPHA_BUILTIN_EXTBL:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0x01, false);
+ case ALPHA_BUILTIN_EXTWL:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0x03, false);
+ case ALPHA_BUILTIN_EXTLL:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0x0f, false);
+ case ALPHA_BUILTIN_EXTQL:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0xff, false);
+ case ALPHA_BUILTIN_EXTWH:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0x03, true);
+ case ALPHA_BUILTIN_EXTLH:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0x0f, true);
+ case ALPHA_BUILTIN_EXTQH:
+ return alpha_fold_builtin_extxx (op, opint, op_const, 0xff, true);
+
+ case ALPHA_BUILTIN_INSBL:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0x01, false);
+ case ALPHA_BUILTIN_INSWL:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0x03, false);
+ case ALPHA_BUILTIN_INSLL:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0x0f, false);
+ case ALPHA_BUILTIN_INSQL:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0xff, false);
+ case ALPHA_BUILTIN_INSWH:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0x03, true);
+ case ALPHA_BUILTIN_INSLH:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0x0f, true);
+ case ALPHA_BUILTIN_INSQH:
+ return alpha_fold_builtin_insxx (op, opint, op_const, 0xff, true);
+
+ case ALPHA_BUILTIN_MSKBL:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0x01, false);
+ case ALPHA_BUILTIN_MSKWL:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0x03, false);
+ case ALPHA_BUILTIN_MSKLL:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0x0f, false);
+ case ALPHA_BUILTIN_MSKQL:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0xff, false);
+ case ALPHA_BUILTIN_MSKWH:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0x03, true);
+ case ALPHA_BUILTIN_MSKLH:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0x0f, true);
+ case ALPHA_BUILTIN_MSKQH:
+ return alpha_fold_builtin_mskxx (op, opint, op_const, 0xff, true);
+
+ case ALPHA_BUILTIN_UMULH:
+ return alpha_fold_builtin_umulh (opint, op_const);
+
+ case ALPHA_BUILTIN_ZAP:
+ opint[1] ^= 0xff;
+ /* FALLTHRU */
+ case ALPHA_BUILTIN_ZAPNOT:
+ return alpha_fold_builtin_zapnot (op, opint, op_const);
+
+ case ALPHA_BUILTIN_MINUB8:
+ return alpha_fold_vector_minmax (MIN_EXPR, op, alpha_v8qi_u);
+ case ALPHA_BUILTIN_MINSB8:
+ return alpha_fold_vector_minmax (MIN_EXPR, op, alpha_v8qi_s);
+ case ALPHA_BUILTIN_MINUW4:
+ return alpha_fold_vector_minmax (MIN_EXPR, op, alpha_v4hi_u);
+ case ALPHA_BUILTIN_MINSW4:
+ return alpha_fold_vector_minmax (MIN_EXPR, op, alpha_v4hi_s);
+ case ALPHA_BUILTIN_MAXUB8:
+ return alpha_fold_vector_minmax (MAX_EXPR, op, alpha_v8qi_u);
+ case ALPHA_BUILTIN_MAXSB8:
+ return alpha_fold_vector_minmax (MAX_EXPR, op, alpha_v8qi_s);
+ case ALPHA_BUILTIN_MAXUW4:
+ return alpha_fold_vector_minmax (MAX_EXPR, op, alpha_v4hi_u);
+ case ALPHA_BUILTIN_MAXSW4:
+ return alpha_fold_vector_minmax (MAX_EXPR, op, alpha_v4hi_s);
+
+ case ALPHA_BUILTIN_PERR:
+ return alpha_fold_builtin_perr (opint, op_const);
+ case ALPHA_BUILTIN_PKLB:
+ return alpha_fold_builtin_pklb (opint, op_const);
+ case ALPHA_BUILTIN_PKWB:
+ return alpha_fold_builtin_pkwb (opint, op_const);
+ case ALPHA_BUILTIN_UNPKBL:
+ return alpha_fold_builtin_unpkbl (opint, op_const);
+ case ALPHA_BUILTIN_UNPKBW:
+ return alpha_fold_builtin_unpkbw (opint, op_const);
+
+ case ALPHA_BUILTIN_CTTZ:
+ return alpha_fold_builtin_cttz (opint, op_const);
+ case ALPHA_BUILTIN_CTLZ:
+ return alpha_fold_builtin_ctlz (opint, op_const);
+ case ALPHA_BUILTIN_CTPOP:
+ return alpha_fold_builtin_ctpop (opint, op_const);
+
+ case ALPHA_BUILTIN_AMASK:
+ case ALPHA_BUILTIN_IMPLVER:
+ case ALPHA_BUILTIN_RPCC:
+ case ALPHA_BUILTIN_THREAD_POINTER:
+ case ALPHA_BUILTIN_SET_THREAD_POINTER:
+ /* None of these are foldable at compile-time. */
default:
- abort ();
+ return NULL;
}
- if (!pat)
- return NULL_RTX;
- emit_insn (pat);
-
- if (nonvoid)
- return target;
- else
- return const0_rtx;
}
\f
/* This page contains routines that are used to determine what the function
break;
imask |= 1UL << regno;
}
-
- /* Glibc likes to use $31 as an unwind stopper for crt0. To
- avoid hackery in unwind-dw2.c, we need to actively store a
- zero in the prologue of _Unwind_RaiseException et al. */
- imask |= 1UL << 31;
}
-
+
/* If any register spilled, then spill the return address also. */
/* ??? This is required by the Digital stack unwind specification
and isn't needed if we're doing Dwarf2 unwinding. */
ret = alpha_sa_size ();
ret += ALPHA_ROUND (current_function_outgoing_args_size);
- if (from == FRAME_POINTER_REGNUM)
- ;
- else if (from == ARG_POINTER_REGNUM)
- ret += (ALPHA_ROUND (get_frame_size ()
- + current_function_pretend_args_size)
- - current_function_pretend_args_size);
- else
- abort ();
+ switch (from)
+ {
+ case FRAME_POINTER_REGNUM:
+ break;
+
+ case ARG_POINTER_REGNUM:
+ ret += (ALPHA_ROUND (get_frame_size ()
+ + current_function_pretend_args_size)
+ - current_function_pretend_args_size);
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
return ret;
}
if (current_function_has_nonlocal_goto)
return 1;
- /* If we need a GP (we have a LDSYM insn or a CALL_INSN), load it first.
+ /* If we need a GP (we have a LDSYM insn or a CALL_INSN), load it first.
Even if we are a static function, we still need to do this in case
our address is taken and passed to something like qsort. */
#define FRP(exp) (start_sequence (), exp, set_frame_related_p ())
+/* Generates a store with the proper unwind info attached. VALUE is
+ stored at BASE_REG+BASE_OFS. If FRAME_BIAS is nonzero, then BASE_REG
+ contains SP+FRAME_BIAS, and that is the unwind info that should be
+ generated. If FRAME_REG != VALUE, then VALUE is being stored on
+ behalf of FRAME_REG, and FRAME_REG should be present in the unwind. */
+
+static void
+emit_frame_store_1 (rtx value, rtx base_reg, HOST_WIDE_INT frame_bias,
+ HOST_WIDE_INT base_ofs, rtx frame_reg)
+{
+ rtx addr, mem, insn;
+
+ addr = plus_constant (base_reg, base_ofs);
+ mem = gen_rtx_MEM (DImode, addr);
+ set_mem_alias_set (mem, alpha_sr_alias_set);
+
+ insn = emit_move_insn (mem, value);
+ RTX_FRAME_RELATED_P (insn) = 1;
+
+ if (frame_bias || value != frame_reg)
+ {
+ if (frame_bias)
+ {
+ addr = plus_constant (stack_pointer_rtx, frame_bias + base_ofs);
+ mem = gen_rtx_MEM (DImode, addr);
+ }
+
+ REG_NOTES (insn)
+ = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
+ gen_rtx_SET (VOIDmode, mem, frame_reg),
+ REG_NOTES (insn));
+ }
+}
+
+static void
+emit_frame_store (unsigned int regno, rtx base_reg,
+ HOST_WIDE_INT frame_bias, HOST_WIDE_INT base_ofs)
+{
+ rtx reg = gen_rtx_REG (DImode, regno);
+ emit_frame_store_1 (reg, base_reg, frame_bias, base_ofs, reg);
+}
+
/* Write function prologue. */
/* On vms we have two kinds of functions:
HOST_WIDE_INT frame_size;
/* Offset from base reg to register save area. */
HOST_WIDE_INT reg_offset;
- rtx sa_reg, mem;
+ rtx sa_reg;
int i;
sa_size = alpha_sa_size ();
frame_size = get_frame_size ();
if (TARGET_ABI_OPEN_VMS)
- frame_size = ALPHA_ROUND (sa_size
+ frame_size = ALPHA_ROUND (sa_size
+ (alpha_procedure_type == PT_STACK ? 8 : 0)
+ frame_size
+ current_function_pretend_args_size);
4096 bytes (we can probably get away without the latter test) and
every 8192 bytes in between. If the frame size is > 32768, we
do this in a loop. Otherwise, we generate the explicit probe
- instructions.
+ instructions.
Note that we are only allowed to adjust sp once in the prologue. */
{
if (frame_size > 4096)
{
- int probed = 4096;
+ int probed;
- do
+ for (probed = 4096; probed < frame_size; probed += 8192)
emit_insn (gen_probe_stack (GEN_INT (TARGET_ABI_UNICOSMK
? -probed + 64
: -probed)));
- while ((probed += 8192) < frame_size);
/* We only have to do this probe if we aren't saving registers. */
- if (sa_size == 0 && probed + 4096 < frame_size)
+ if (sa_size == 0 && frame_size > probed - 4096)
emit_insn (gen_probe_stack (GEN_INT (-frame_size)));
}
/* For NT stack unwind (done by 'reverse execution'), it's
not OK to take the result of a loop, even though the value
is already in ptr, so we reload it via a single operation
- and subtract it to sp.
+ and subtract it to sp.
Yes, that's correct -- we have to reload the whole constant
into a temporary via ldah+lda then subtract from sp. */
if (!TARGET_ABI_UNICOSMK)
{
+ HOST_WIDE_INT sa_bias = 0;
+
/* Cope with very large offsets to the register save area. */
sa_reg = stack_pointer_rtx;
if (reg_offset + sa_size > 0x8000)
{
int low = ((reg_offset & 0xffff) ^ 0x8000) - 0x8000;
- HOST_WIDE_INT bias;
+ rtx sa_bias_rtx;
if (low + sa_size <= 0x8000)
- bias = reg_offset - low, reg_offset = low;
- else
- bias = reg_offset, reg_offset = 0;
+ sa_bias = reg_offset - low, reg_offset = low;
+ else
+ sa_bias = reg_offset, reg_offset = 0;
sa_reg = gen_rtx_REG (DImode, 24);
- FRP (emit_insn (gen_adddi3 (sa_reg, stack_pointer_rtx,
- GEN_INT (bias))));
+ sa_bias_rtx = GEN_INT (sa_bias);
+
+ if (add_operand (sa_bias_rtx, DImode))
+ emit_insn (gen_adddi3 (sa_reg, stack_pointer_rtx, sa_bias_rtx));
+ else
+ {
+ emit_move_insn (sa_reg, sa_bias_rtx);
+ emit_insn (gen_adddi3 (sa_reg, stack_pointer_rtx, sa_reg));
+ }
}
-
+
/* Save regs in stack order. Beginning with VMS PV. */
if (TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_STACK)
- {
- mem = gen_rtx_MEM (DImode, stack_pointer_rtx);
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DImode, REG_PV)));
- }
+ emit_frame_store (REG_PV, stack_pointer_rtx, 0, 0);
/* Save register RA next. */
if (imask & (1UL << REG_RA))
{
- mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DImode, REG_RA)));
+ emit_frame_store (REG_RA, sa_reg, sa_bias, reg_offset);
imask &= ~(1UL << REG_RA);
reg_offset += 8;
}
for (i = 0; i < 31; i++)
if (imask & (1UL << i))
{
- mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DImode, i)));
+ emit_frame_store (i, sa_reg, sa_bias, reg_offset);
reg_offset += 8;
}
- /* Store a zero if requested for unwinding. */
- if (imask & (1UL << 31))
- {
- rtx insn, t;
-
- mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- insn = emit_move_insn (mem, const0_rtx);
-
- RTX_FRAME_RELATED_P (insn) = 1;
- t = gen_rtx_REG (Pmode, 31);
- t = gen_rtx_SET (VOIDmode, mem, t);
- t = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, t, REG_NOTES (insn));
- REG_NOTES (insn) = t;
-
- reg_offset += 8;
- }
-
for (i = 0; i < 31; i++)
if (fmask & (1UL << i))
{
- mem = gen_rtx_MEM (DFmode, plus_constant (sa_reg, reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DFmode, i+32)));
+ emit_frame_store (i+32, sa_reg, sa_bias, reg_offset);
reg_offset += 8;
}
}
for (i = 9; i < 15; i++)
if (imask & (1UL << i))
{
- mem = gen_rtx_MEM (DImode, plus_constant(hard_frame_pointer_rtx,
- reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DImode, i)));
+ emit_frame_store (i, hard_frame_pointer_rtx, 0, reg_offset);
reg_offset -= 8;
}
for (i = 2; i < 10; i++)
if (fmask & (1UL << i))
{
- mem = gen_rtx_MEM (DFmode, plus_constant (hard_frame_pointer_rtx,
- reg_offset));
- set_mem_alias_set (mem, alpha_sr_alias_set);
- FRP (emit_move_insn (mem, gen_rtx_REG (DFmode, i+32)));
+ emit_frame_store (i+32, hard_frame_pointer_rtx, 0, reg_offset);
reg_offset -= 8;
}
}
if (current_function_outgoing_args_size != 0)
{
rtx seq
- = emit_move_insn (stack_pointer_rtx,
+ = emit_move_insn (stack_pointer_rtx,
plus_constant
(hard_frame_pointer_rtx,
- (ALPHA_ROUND
(current_function_outgoing_args_size))));
-
+
/* Only set FRAME_RELATED_P on the stack adjustment we just emitted
if ! frame_pointer_needed. Setting the bit will change the CFA
computation rule to use sp again, which would be wrong if we had
(clobber:BLK (scratch)), but this doesn't work for fp insns. So we
have to prevent all such scheduling with a blockage.
- Linux, on the other hand, never bothered to implement OSF/1's
+ Linux, on the other hand, never bothered to implement OSF/1's
exception handling, and so doesn't care about such things. Anyone
planning to use dwarf2 frame-unwind info can also omit the blockage. */
emit_insn (gen_blockage ());
}
+/* Count the number of .file directives, so that .loc is up to date. */
+int num_source_filenames = 0;
+
/* Output the textual info surrounding the prologue. */
void
frame_size = get_frame_size ();
if (TARGET_ABI_OPEN_VMS)
- frame_size = ALPHA_ROUND (sa_size
+ frame_size = ALPHA_ROUND (sa_size
+ (alpha_procedure_type == PT_STACK ? 8 : 0)
+ frame_size
+ current_function_pretend_args_size);
ASM_OUTPUT_SOURCE_FILENAME (file,
DECL_SOURCE_FILE (current_function_decl));
#endif
-#ifdef ASM_OUTPUT_SOURCE_LINE
+#ifdef SDB_OUTPUT_SOURCE_LINE
if (debug_info_level != DINFO_LEVEL_TERSE)
- ASM_OUTPUT_SOURCE_LINE (file,
- DECL_SOURCE_LINE (current_function_decl), 0);
+ SDB_OUTPUT_SOURCE_LINE (file,
+ DECL_SOURCE_LINE (current_function_decl));
#endif
}
#if TARGET_ABI_OPEN_VMS
/* Ifdef'ed cause link_section are only available then. */
- readonly_data_section ();
+ switch_to_section (readonly_data_section);
fprintf (file, "\t.align 3\n");
assemble_name (file, fnname); fputs ("..na:\n", file);
fputs ("\t.ascii \"", file);
assemble_name (file, fnname);
fputs ("\\0\"\n", file);
alpha_need_linkage (fnname, 1);
- text_section ();
+ switch_to_section (text_section);
#endif
}
/* Write function epilogue. */
-/* ??? At some point we will want to support full unwind, and so will
+/* ??? At some point we will want to support full unwind, and so will
need to mark the epilogue as well. At the moment, we just confuse
dwarf2out. */
#undef FRP
frame_size = get_frame_size ();
if (TARGET_ABI_OPEN_VMS)
- frame_size = ALPHA_ROUND (sa_size
+ frame_size = ALPHA_ROUND (sa_size
+ (alpha_procedure_type == PT_STACK ? 8 : 0)
+ frame_size
+ current_function_pretend_args_size);
if (low + sa_size <= 0x8000)
bias = reg_offset - low, reg_offset = low;
- else
+ else
bias = reg_offset, reg_offset = 0;
sa_reg = gen_rtx_REG (DImode, 22);
FRP (emit_move_insn (sa_reg, sa_reg_exp));
}
-
+
/* Restore registers in order, excepting a true frame pointer. */
mem = gen_rtx_MEM (DImode, plus_constant (sa_reg, reg_offset));
reg_offset += 8;
}
- if (imask & (1UL << 31))
- reg_offset += 8;
-
for (i = 0; i < 31; ++i)
if (fmask & (1UL << i))
{
else
{
rtx tmp = gen_rtx_REG (DImode, 23);
- FRP (sp_adj2 = alpha_emit_set_const (tmp, DImode, frame_size, 3));
+ FRP (sp_adj2 = alpha_emit_set_const (tmp, DImode, frame_size,
+ 3, false));
if (!sp_adj2)
{
/* We can't drop new things to memory this late, afaik,
so build it up by pieces. */
FRP (sp_adj2 = alpha_emit_set_long_const (tmp, frame_size,
-(frame_size < 0)));
- if (!sp_adj2)
- abort ();
+ gcc_assert (sp_adj2);
}
}
FRP (emit_move_insn (stack_pointer_rtx,
gen_rtx_PLUS (DImode, sp_adj1, sp_adj2)));
}
- else
+ else
{
if (TARGET_ABI_OPEN_VMS && alpha_procedure_type == PT_REGISTER)
{
insn_locators_initialize ();
shorten_branches (insn);
final_start_function (insn, file, 1);
- final (insn, file, 1, 0);
+ final (insn, file, 1);
final_end_function ();
}
#endif /* TARGET_ABI_OSF */
int sdb_label_count = 0;
-/* Next label # for each statement. */
-
-static int sym_lineno = 0;
-
-/* Count the number of .file directives, so that .loc is up to date. */
-
-static int num_source_filenames = 0;
-
/* Name of the file containing the current function. */
static const char *current_function_file = "";
alpha_output_filename (FILE *stream, const char *name)
{
static int first_time = TRUE;
- char ltext_label_name[100];
if (first_time)
{
}
else if (write_symbols == DBX_DEBUG)
- {
- ASM_GENERATE_INTERNAL_LABEL (ltext_label_name, "Ltext", 0);
- fprintf (stream, "%s", ASM_STABS_OP);
- output_quoted_string (stream, name);
- fprintf (stream, ",%d,0,0,%s\n", N_SOL, <ext_label_name[1]);
- }
+ /* dbxout.c will emit an appropriate .stabs directive. */
+ return;
else if (name != current_function_file
&& strcmp (name, current_function_file) != 0)
fprintf (stream, "\n");
}
}
-
-/* Emit a linenumber to a stream. */
-
-void
-alpha_output_lineno (FILE *stream, int line)
-{
- if (write_symbols == DBX_DEBUG)
- {
- /* mips-tfile doesn't understand .stabd directives. */
- ++sym_lineno;
- fprintf (stream, "$LM%d:\n%s%d,0,%d,$LM%d\n",
- sym_lineno, ASM_STABN_OP, N_SLINE, line, sym_lineno);
- }
- else
- fprintf (stream, "\n\t.loc\t%d %d\n", num_source_filenames, line);
-}
\f
/* Structure to show the current status of registers and memory. */
case NEG: case NOT: case SIGN_EXTEND: case ZERO_EXTEND:
case TRUNCATE: case FLOAT_EXTEND: case FLOAT_TRUNCATE: case FLOAT:
case FIX: case UNSIGNED_FLOAT: case UNSIGNED_FIX: case ABS:
- case SQRT: case FFS:
+ case SQRT: case FFS:
summarize_insn (XEXP (x, 0), sum, 0);
break;
break;
default:
- abort ();
+ gcc_unreachable ();
}
}
}
shadow.used.fp = 0;
shadow.used.mem = 0;
shadow.defd = shadow.used;
-
+
for (i = get_insns (); i ; i = NEXT_INSN (i))
{
if (GET_CODE (i) == NOTE)
switch (GET_CODE (i))
{
case INSN:
- /* Annoyingly, get_attr_trap will abort on these. */
+ /* Annoyingly, get_attr_trap will die on these. */
if (GET_CODE (PATTERN (i)) == USE
|| GET_CODE (PATTERN (i)) == CLOBBER)
break;
|| (sum.defd.mem & shadow.used.mem))
{
/* (a) would be violated (also takes care of (b)) */
- if (get_attr_trap (i) == TRAP_YES
- && ((sum.defd.i & sum.used.i)
- || (sum.defd.fp & sum.used.fp)))
- abort ();
+ gcc_assert (get_attr_trap (i) != TRAP_YES
+ || (!(sum.defd.i & sum.used.i)
+ && !(sum.defd.fp & sum.used.fp)));
goto close_shadow;
}
goto close_shadow;
default:
- abort ();
+ gcc_unreachable ();
}
}
else
\f
/* Alpha can only issue instruction groups simultaneously if they are
suitably aligned. This is very processor-specific. */
+/* There are a number of entries in alphaev4_insn_pipe and alphaev5_insn_pipe
+ that are marked "fake". These instructions do not exist on that target,
+ but it is possible to see these insns with deranged combinations of
+ command-line options, such as "-mtune=ev4 -mmax". Instead of aborting,
+ choose a result at random. */
enum alphaev4_pipe {
EV4_STOP = 0,
switch (get_attr_type (insn))
{
case TYPE_ILD:
+ case TYPE_LDSYM:
case TYPE_FLD:
+ case TYPE_LD_L:
return EV4_IBX;
- case TYPE_LDSYM:
case TYPE_IADD:
case TYPE_ILOG:
case TYPE_ICMOV:
case TYPE_ICMP:
- case TYPE_IST:
case TYPE_FST:
case TYPE_SHIFT:
case TYPE_IMUL:
case TYPE_FBR:
+ case TYPE_MVI: /* fake */
return EV4_IB0;
+ case TYPE_IST:
case TYPE_MISC:
case TYPE_IBR:
case TYPE_JSR:
case TYPE_FADD:
case TYPE_FDIV:
case TYPE_FMUL:
+ case TYPE_ST_C:
+ case TYPE_MB:
+ case TYPE_FSQRT: /* fake */
+ case TYPE_FTOI: /* fake */
+ case TYPE_ITOF: /* fake */
return EV4_IB1;
default:
- abort ();
+ gcc_unreachable ();
}
}
case TYPE_IMUL:
case TYPE_MISC:
case TYPE_MVI:
+ case TYPE_LD_L:
+ case TYPE_ST_C:
+ case TYPE_MB:
+ case TYPE_FTOI: /* fake */
+ case TYPE_ITOF: /* fake */
return EV5_E0;
case TYPE_IBR:
case TYPE_FCMOV:
case TYPE_FADD:
case TYPE_FDIV:
+ case TYPE_FSQRT: /* fake */
return EV5_FA;
case TYPE_FMUL:
return EV5_FM;
default:
- abort();
+ gcc_unreachable ();
}
}
-/* IN_USE is a mask of the slots currently filled within the insn group.
+/* IN_USE is a mask of the slots currently filled within the insn group.
The mask bits come from alphaev4_pipe above. If EV4_IBX is set, then
- the insn in EV4_IB0 can be swapped by the hardware into EV4_IB1.
+ the insn in EV4_IB0 can be swapped by the hardware into EV4_IB1.
LEN is, of course, the length of the group in bytes. */
if (in_use)
goto done;
- /* If this is a completely unrecognized insn, its an asm.
+ /* If this is a completely unrecognized insn, it's an asm.
We don't know how long it is, so record length as -1 to
signal a needed realignment. */
if (recog_memoized (insn) < 0)
break;
default:
- abort();
+ gcc_unreachable ();
}
len += 4;
-
+
/* Haifa doesn't do well scheduling branches. */
if (GET_CODE (insn) == JUMP_INSN)
goto next_and_done;
return insn;
}
-/* IN_USE is a mask of the slots currently filled within the insn group.
+/* IN_USE is a mask of the slots currently filled within the insn group.
The mask bits come from alphaev5_pipe above. If EV5_E01 is set, then
- the insn in EV5_E0 can be swapped by the hardware into EV5_E1.
+ the insn in EV5_E0 can be swapped by the hardware into EV5_E1.
LEN is, of course, the length of the group in bytes. */
if (in_use)
goto done;
- /* If this is a completely unrecognized insn, its an asm.
+ /* If this is a completely unrecognized insn, it's an asm.
We don't know how long it is, so record length as -1 to
signal a needed realignment. */
if (recog_memoized (insn) < 0)
len = get_attr_length (insn);
goto next_and_done;
- /* ??? Most of the places below, we would like to abort, as
- it would indicate an error either in Haifa, or in the
- scheduling description. Unfortunately, Haifa never
- schedules the last instruction of the BB, so we don't
- have an accurate TI bit to go off. */
+ /* ??? Most of the places below, we would like to assert never
+ happen, as it would indicate an error either in Haifa, or
+ in the scheduling description. Unfortunately, Haifa never
+ schedules the last instruction of the BB, so we don't have
+ an accurate TI bit to go off. */
case EV5_E01:
if (in_use & EV5_E0)
{
break;
default:
- abort();
+ gcc_unreachable ();
}
len += 4;
-
+
/* Haifa doesn't do well scheduling branches. */
/* ??? If this is predicted not-taken, slotting continues, except
that no more IBR, FBR, or JSR insns may be slotted. */
unsigned int align;
/* OFS is the offset of the current insn in the insn group. */
int ofs;
- int prev_in_use, in_use, len;
+ int prev_in_use, in_use, len, ldgp;
rtx i, next;
/* Let shorten branches care for assigning alignments to code labels. */
if (GET_CODE (i) == NOTE)
i = next_nonnote_insn (i);
+ ldgp = alpha_function_needs_gp ? 8 : 0;
+
while (i)
{
next = (*next_group) (i, &in_use, &len);
else if (ofs & (new_align-1))
ofs = (ofs | (new_align-1)) + 1;
- if (len != 0)
- abort();
+ gcc_assert (!len);
}
/* Handle complex instructions special. */
}
}
+ /* We may not insert padding inside the initial ldgp sequence. */
+ else if (ldgp > 0)
+ ldgp -= len;
+
/* If the group won't fit in the same INT16 as the previous,
we need to add padding to keep the group together. Rather
than simply leaving the insn filling to the assembler, we
else
where = i;
- do
+ do
emit_insn_before ((*next_nop)(&prev_in_use), where);
while (--nop_count);
ofs = 0;
&& alpha_tp != ALPHA_TP_INSN
&& flag_schedule_insns_after_reload)
{
- if (alpha_cpu == PROCESSOR_EV4)
+ if (alpha_tune == PROCESSOR_EV4)
alpha_align_insns (8, alphaev4_next_group, alphaev4_next_nop);
- else if (alpha_cpu == PROCESSOR_EV5)
+ else if (alpha_tune == PROCESSOR_EV5)
alpha_align_insns (16, alphaev5_next_group, alphaev5_next_nop);
}
}
if (TARGET_EXPLICIT_RELOCS)
fputs ("\t.set nomacro\n", asm_out_file);
if (TARGET_SUPPORT_ARCH | TARGET_BWX | TARGET_MAX | TARGET_FIX | TARGET_CIX)
- fprintf (asm_out_file,
- "\t.arch %s\n",
- TARGET_CPU_EV6 ? "ev6"
- : (TARGET_CPU_EV5
- ? (TARGET_MAX ? "pca56" : TARGET_BWX ? "ev56" : "ev5")
- : "ev4"));
+ {
+ const char *arch;
+
+ if (alpha_cpu == PROCESSOR_EV6 || TARGET_FIX || TARGET_CIX)
+ arch = "ev6";
+ else if (TARGET_MAX)
+ arch = "pca56";
+ else if (TARGET_BWX)
+ arch = "ev56";
+ else if (alpha_cpu == PROCESSOR_EV5)
+ arch = "ev5";
+ else
+ arch = "ev4";
+
+ fprintf (asm_out_file, "\t.arch %s\n", arch);
+ }
}
#endif
#ifdef OBJECT_FORMAT_ELF
-/* Switch to the section to which we should output X. The only thing
- special we do here is to honor small data. */
+/* Return a section for X. The only special thing we do here is to
+ honor small data. */
-static void
+static section *
alpha_elf_select_rtx_section (enum machine_mode mode, rtx x,
unsigned HOST_WIDE_INT align)
{
if (TARGET_SMALL_DATA && GET_MODE_SIZE (mode) <= g_switch_value)
/* ??? Consider using mergeable sdata sections. */
- sdata_section ();
+ return sdata_section;
else
- default_elf_select_rtx_section (mode, x, align);
+ return default_elf_select_rtx_section (mode, x, align);
}
#endif /* OBJECT_FORMAT_ELF */
if (!alpha_funcs_tree)
alpha_funcs_tree = splay_tree_new_ggc ((splay_tree_compare_fn)
splay_tree_compare_pointers);
-
+
cfaf = (struct alpha_funcs *) ggc_alloc (sizeof (struct alpha_funcs));
cfaf->links = 0;
al->rkind = KIND_CODEADDR;
else
al->rkind = KIND_LINKAGE;
-
+
if (lflag)
return gen_rtx_MEM (Pmode, plus_constant (al->linkage, 8));
else
splay_tree_node node;
struct alpha_funcs *func;
- link_section ();
+ fprintf (stream, "\t.link\n");
fprintf (stream, "\t.align 3\n");
+ in_section = NULL;
+
node = splay_tree_lookup (alpha_funcs_tree, (splay_tree_key) fundecl);
func = (struct alpha_funcs *) node->value;
the section; 0 if the default should be used. */
static void
-vms_asm_named_section (const char *name, unsigned int flags)
+vms_asm_named_section (const char *name, unsigned int flags,
+ tree decl ATTRIBUTE_UNUSED)
{
fputc ('\n', asm_out_file);
fprintf (asm_out_file, ".section\t%s", name);
/* Record an element in the table of global constructors. SYMBOL is
a SYMBOL_REF of the function to be called; PRIORITY is a number
- between 0 and MAX_INIT_PRIORITY.
+ between 0 and MAX_INIT_PRIORITY.
Differs from default_ctors_section_asm_out_constructor in that the
width of the .ctors entry is always 64 bits, rather than the 32 bits
static void
vms_asm_out_constructor (rtx symbol, int priority ATTRIBUTE_UNUSED)
{
- ctors_section ();
+ switch_to_section (ctors_section);
assemble_align (BITS_PER_WORD);
assemble_integer (symbol, UNITS_PER_WORD, BITS_PER_WORD, 1);
}
static void
vms_asm_out_destructor (rtx symbol, int priority ATTRIBUTE_UNUSED)
{
- dtors_section ();
+ switch_to_section (dtors_section);
assemble_align (BITS_PER_WORD);
assemble_integer (symbol, UNITS_PER_WORD, BITS_PER_WORD, 1);
}
unicosmk_initial_elimination_offset (int from, int to)
{
int fixed_size;
-
+
fixed_size = alpha_sa_size();
if (fixed_size != 0)
fixed_size += 48;
if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
- return -fixed_size;
+ return -fixed_size;
else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
return 0;
else if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
+ ALPHA_ROUND (get_frame_size()));
else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
return (ALPHA_ROUND (fixed_size)
- + ALPHA_ROUND (get_frame_size()
+ + ALPHA_ROUND (get_frame_size()
+ current_function_outgoing_args_size));
else
- abort ();
+ gcc_unreachable ();
}
/* Output the module name for .ident and .end directives. We have to strip
unsigned len = strlen (name);
char *clean_name = alloca (len + 2);
char *ptr = clean_name;
-
+
/* CAM only accepts module names that start with a letter or '$'. We
prefix the module name with a '$' if necessary. */
tree name_tree;
printf ("T3E__: common %s\n", name);
- common_section ();
+ in_section = NULL;
fputs("\t.endp\n\n\t.psect ", file);
assemble_name(file, name);
fprintf(file, ",%d,common\n", floor_log2 (align / BITS_PER_UNIT));
#define SECTION_MAIN (SECTION_PUBLIC << 1)
static int current_section_align;
+/* A get_unnamed_section callback for switching to the text section. */
+
+static void
+unicosmk_output_text_section_asm_op (const void *data ATTRIBUTE_UNUSED)
+{
+ static int count = 0;
+ fprintf (asm_out_file, "\t.endp\n\n\t.psect\tgcc@text___%d,code\n", count++);
+}
+
+/* A get_unnamed_section callback for switching to the data section. */
+
+static void
+unicosmk_output_data_section_asm_op (const void *data ATTRIBUTE_UNUSED)
+{
+ static int count = 1;
+ fprintf (asm_out_file, "\t.endp\n\n\t.psect\tgcc@data___%d,data\n", count++);
+}
+
+/* Implement TARGET_ASM_INIT_SECTIONS.
+
+ The Cray assembler is really weird with respect to sections. It has only
+ named sections and you can't reopen a section once it has been closed.
+ This means that we have to generate unique names whenever we want to
+ reenter the text or the data section. */
+
+static void
+unicosmk_init_sections (void)
+{
+ text_section = get_unnamed_section (SECTION_CODE,
+ unicosmk_output_text_section_asm_op,
+ NULL);
+ data_section = get_unnamed_section (SECTION_WRITE,
+ unicosmk_output_data_section_asm_op,
+ NULL);
+ readonly_data_section = data_section;
+}
+
static unsigned int
unicosmk_section_type_flags (tree decl, const char *name,
int reloc ATTRIBUTE_UNUSED)
const char *name;
int len;
- if (!decl)
- abort ();
+ gcc_assert (decl);
name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
name = default_strip_name_encoding (name);
{
char *string;
- /* It is essential that we prefix the section name here because
- otherwise the section names generated for constructors and
+ /* It is essential that we prefix the section name here because
+ otherwise the section names generated for constructors and
destructors confuse collect2. */
string = alloca (len + 6);
the section; 0 if the default should be used. */
static void
-unicosmk_asm_named_section (const char *name, unsigned int flags)
+unicosmk_asm_named_section (const char *name, unsigned int flags,
+ tree decl ATTRIBUTE_UNUSED)
{
const char *kind;
/* Output an alignment directive. We have to use the macro 'gcc@code@align'
in code sections because .align fill unused space with zeroes. */
-
+
void
unicosmk_output_align (FILE *file, int align)
{
unicosmk_defer_case_vector (rtx lab, rtx vec)
{
struct machine_function *machine = cfun->machine;
-
+
vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec);
machine->addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec,
- machine->addr_list);
+ machine->addr_list);
}
/* Output a case vector. */
if (machine->addr_list == NULL_RTX)
return;
- data_section ();
+ switch_to_section (data_section);
for (t = machine->addr_list; t; t = XEXP (t, 1))
unicosmk_output_addr_vec (file, XEXP (t, 0));
}
static const char *
unicosmk_ssib_name (void)
{
- /* This is ok since CAM won't be able to deal with names longer than that
+ /* This is ok since CAM won't be able to deal with names longer than that
anyway. */
static char name[256];
int len;
x = DECL_RTL (cfun->decl);
- if (GET_CODE (x) != MEM)
- abort ();
+ gcc_assert (GET_CODE (x) == MEM);
x = XEXP (x, 0);
- if (GET_CODE (x) != SYMBOL_REF)
- abort ();
+ gcc_assert (GET_CODE (x) == SYMBOL_REF);
fnname = XSTR (x, 0);
len = strlen (fnname);
return name;
}
-/* Set up the dynamic subprogram information block (DSIB) and update the
- frame pointer register ($15) for subroutines which have a frame. If the
+/* Set up the dynamic subprogram information block (DSIB) and update the
+ frame pointer register ($15) for subroutines which have a frame. If the
subroutine doesn't have a frame, simply increment $15. */
static void
rtx ciw;
struct machine_function *machine = cfun->machine;
- ssib_section ();
+ in_section = NULL;
fprintf (file, "\t.endp\n\n\t.psect\t%s%s,data\n", user_label_prefix,
unicosmk_ssib_name ());
+ strlen (current_function_name ())/8 + 5);
}
-static char unicosmk_section_buf[100];
-
-char *
-unicosmk_text_section (void)
-{
- static int count = 0;
- sprintf (unicosmk_section_buf, "\t.endp\n\n\t.psect\tgcc@text___%d,code",
- count++);
- return unicosmk_section_buf;
-}
-
-char *
-unicosmk_data_section (void)
-{
- static int count = 1;
- sprintf (unicosmk_section_buf, "\t.endp\n\n\t.psect\tgcc@data___%d,data",
- count++);
- return unicosmk_section_buf;
-}
-
/* The Cray assembler doesn't accept extern declarations for symbols which
are defined in the same file. We have to keep track of all global
symbols which are referenced and/or defined in a source file and output
len = strlen (user_label_prefix);
for (p = unicosmk_extern_head; p != 0; p = p->next)
{
- /* We have to strip the encoding and possibly remove user_label_prefix
+ /* We have to strip the encoding and possibly remove user_label_prefix
from the identifier in order to handle -fleading-underscore and
explicit asm names correctly (cf. gcc.dg/asm-names-1.c). */
real_name = default_strip_name_encoding (p->name);
if (len && p->name[0] == '*'
&& !memcmp (real_name, user_label_prefix, len))
real_name += len;
-
+
name_tree = get_identifier (real_name);
if (! TREE_ASM_WRITTEN (name_tree))
{
}
}
}
-
+
/* Record an extern. */
void
const char *name;
};
-/* List of identifiers which have been replaced by DEX expressions. The DEX
+/* List of identifiers which have been replaced by DEX expressions. The DEX
number is determined by the position in the list. */
-static struct unicosmk_dex *unicosmk_dex_list = NULL;
+static struct unicosmk_dex *unicosmk_dex_list = NULL;
/* The number of elements in the DEX list. */
struct unicosmk_dex *dex;
const char *name;
int i;
-
+
if (GET_CODE (x) != SYMBOL_REF)
return 0;
return i;
--i;
}
-
+
dex = (struct unicosmk_dex *) xmalloc (sizeof (struct unicosmk_dex));
dex->name = name;
dex->next = unicosmk_dex_list;
putc ('\n', file);
--i;
}
-
+
fprintf (file, "\t.dexend\n");
}
/* Output text that to appear at the beginning of an assembler file. */
-static void
+static void
unicosmk_file_start (void)
{
int i;
unicosmk_output_externs (asm_out_file);
- /* Output dex definitions used for functions whose names conflict with
+ /* Output dex definitions used for functions whose names conflict with
register names. */
unicosmk_output_dex (asm_out_file);
#define TARGET_INIT_BUILTINS alpha_init_builtins
#undef TARGET_EXPAND_BUILTIN
#define TARGET_EXPAND_BUILTIN alpha_expand_builtin
+#undef TARGET_FOLD_BUILTIN
+#define TARGET_FOLD_BUILTIN alpha_fold_builtin
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL alpha_function_ok_for_sibcall
#undef TARGET_CANNOT_COPY_INSN_P
#define TARGET_CANNOT_COPY_INSN_P alpha_cannot_copy_insn_p
+#undef TARGET_CANNOT_FORCE_CONST_MEM
+#define TARGET_CANNOT_FORCE_CONST_MEM alpha_cannot_force_const_mem
#if TARGET_ABI_OSF
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK alpha_output_mi_thunk_osf
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true
+#undef TARGET_STDARG_OPTIMIZE_HOOK
+#define TARGET_STDARG_OPTIMIZE_HOOK alpha_stdarg_optimize_hook
#endif
#undef TARGET_RTX_COSTS
#define TARGET_SPLIT_COMPLEX_ARG alpha_split_complex_arg
#undef TARGET_GIMPLIFY_VA_ARG_EXPR
#define TARGET_GIMPLIFY_VA_ARG_EXPR alpha_gimplify_va_arg
+#undef TARGET_ARG_PARTIAL_BYTES
+#define TARGET_ARG_PARTIAL_BYTES alpha_arg_partial_bytes
+
+#undef TARGET_SCALAR_MODE_SUPPORTED_P
+#define TARGET_SCALAR_MODE_SUPPORTED_P alpha_scalar_mode_supported_p
+#undef TARGET_VECTOR_MODE_SUPPORTED_P
+#define TARGET_VECTOR_MODE_SUPPORTED_P alpha_vector_mode_supported_p
#undef TARGET_BUILD_BUILTIN_VA_LIST
#define TARGET_BUILD_BUILTIN_VA_LIST alpha_build_builtin_va_list
+/* The Alpha architecture does not require sequential consistency. See
+ http://www.cs.umd.edu/~pugh/java/memoryModel/AlphaReordering.html
+ for an example of how it can be violated in practice. */
+#undef TARGET_RELAXED_ORDERING
+#define TARGET_RELAXED_ORDERING true
+
+#undef TARGET_DEFAULT_TARGET_FLAGS
+#define TARGET_DEFAULT_TARGET_FLAGS \
+ (TARGET_DEFAULT | TARGET_CPU_DEFAULT | TARGET_DEFAULT_EXPLICIT_RELOCS)
+#undef TARGET_HANDLE_OPTION
+#define TARGET_HANDLE_OPTION alpha_handle_option
+
+#ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
+#undef TARGET_MANGLE_FUNDAMENTAL_TYPE
+#define TARGET_MANGLE_FUNDAMENTAL_TYPE alpha_mangle_fundamental_type
+#endif
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
#include "gt-alpha.h"
-