/* Expand the basic unary and binary arithmetic operations, for GNU compiler.
Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
- 2011 Free Software Foundation, Inc.
+ 2011, 2012 Free Software Foundation, Inc.
This file is part of GCC.
case ABS_EXPR:
return trapv ? absv_optab : abs_optab;
- case VEC_EXTRACT_EVEN_EXPR:
- return vec_extract_even_optab;
-
- case VEC_EXTRACT_ODD_EXPR:
- return vec_extract_odd_optab;
-
- case VEC_INTERLEAVE_HIGH_EXPR:
- return vec_interleave_high_optab;
-
- case VEC_INTERLEAVE_LOW_EXPR:
- return vec_interleave_low_optab;
-
default:
return NULL;
}
calculated at compile time. The arguments and return value are
otherwise the same as for expand_binop. */
-static rtx
+rtx
simplify_expand_binop (enum machine_mode mode, optab binoptab,
rtx op0, rtx op1, rtx target, int unsignedp,
enum optab_methods methods)
}
}
- /* Certain vector operations can be implemented with vector permutation. */
- if (VECTOR_MODE_P (mode))
- {
- enum tree_code tcode = ERROR_MARK;
- rtx sel;
-
- if (binoptab == vec_interleave_high_optab)
- tcode = VEC_INTERLEAVE_HIGH_EXPR;
- else if (binoptab == vec_interleave_low_optab)
- tcode = VEC_INTERLEAVE_LOW_EXPR;
- else if (binoptab == vec_extract_even_optab)
- tcode = VEC_EXTRACT_EVEN_EXPR;
- else if (binoptab == vec_extract_odd_optab)
- tcode = VEC_EXTRACT_ODD_EXPR;
-
- if (tcode != ERROR_MARK
- && can_vec_perm_for_code_p (tcode, mode, &sel))
- {
- temp = expand_vec_perm (mode, op0, op1, sel, target);
- gcc_assert (temp != NULL);
- return temp;
- }
- }
-
/* Look for a wider mode of the same class for which we think we
can open-code the operation. Check for a widening multiply at the
wider mode as well. */
{
rtx temp = emit_move_insn (target, xtarget);
- set_unique_reg_note (temp,
- REG_EQUAL,
- gen_rtx_fmt_ee (binoptab->code, mode,
- copy_rtx (xop0),
- copy_rtx (xop1)));
+ set_dst_reg_note (temp, REG_EQUAL,
+ gen_rtx_fmt_ee (binoptab->code, mode,
+ copy_rtx (xop0),
+ copy_rtx (xop1)),
+ target);
}
else
target = xtarget;
if (optab_handler (mov_optab, mode) != CODE_FOR_nothing)
{
temp = emit_move_insn (target ? target : product, product);
- set_unique_reg_note (temp,
- REG_EQUAL,
- gen_rtx_fmt_ee (MULT, mode,
- copy_rtx (op0),
- copy_rtx (op1)));
+ set_dst_reg_note (temp,
+ REG_EQUAL,
+ gen_rtx_fmt_ee (MULT, mode,
+ copy_rtx (op0),
+ copy_rtx (op1)),
+ target ? target : product);
}
return product;
}
gen_lowpart (imode, target), 1, OPTAB_LIB_WIDEN);
target = lowpart_subreg_maybe_copy (mode, temp, imode);
- set_unique_reg_note (get_last_insn (), REG_EQUAL,
- gen_rtx_fmt_e (code, mode, copy_rtx (op0)));
+ set_dst_reg_note (get_last_insn (), REG_EQUAL,
+ gen_rtx_fmt_e (code, mode, copy_rtx (op0)),
+ target);
}
return target;
}
last = emit_move_insn (target, result);
- if (optab_handler (mov_optab, GET_MODE (target)) != CODE_FOR_nothing)
- set_unique_reg_note (last, REG_EQUAL, copy_rtx (equiv));
+ set_dst_reg_note (last, REG_EQUAL, copy_rtx (equiv), target);
if (final_dest != target)
emit_move_insn (final_dest, target);
{
/* Make a place for a REG_NOTE and add it. */
insn = emit_move_insn (to, to);
- set_unique_reg_note (insn,
- REG_EQUAL,
- gen_rtx_fmt_e (UNSIGNED_FIX,
- GET_MODE (to),
- copy_rtx (from)));
+ set_dst_reg_note (insn, REG_EQUAL,
+ gen_rtx_fmt_e (UNSIGNED_FIX, GET_MODE (to),
+ copy_rtx (from)),
+ to);
}
return;
init_optab (udot_prod_optab, UNKNOWN);
init_optab (vec_extract_optab, UNKNOWN);
- init_optab (vec_extract_even_optab, UNKNOWN);
- init_optab (vec_extract_odd_optab, UNKNOWN);
- init_optab (vec_interleave_high_optab, UNKNOWN);
- init_optab (vec_interleave_low_optab, UNKNOWN);
init_optab (vec_set_optab, UNKNOWN);
init_optab (vec_init_optab, UNKNOWN);
init_optab (vec_shl_optab, UNKNOWN);
init_sync_libfuncs_1 (sync_old_add_optab, "__sync_fetch_and_add", max);
init_sync_libfuncs_1 (sync_old_sub_optab, "__sync_fetch_and_sub", max);
- init_sync_libfuncs_1 (sync_old_ior_optab, "__sync_fetch_and_ior", max);
+ init_sync_libfuncs_1 (sync_old_ior_optab, "__sync_fetch_and_or", max);
init_sync_libfuncs_1 (sync_old_and_optab, "__sync_fetch_and_and", max);
init_sync_libfuncs_1 (sync_old_xor_optab, "__sync_fetch_and_xor", max);
init_sync_libfuncs_1 (sync_old_nand_optab, "__sync_fetch_and_nand", max);
init_sync_libfuncs_1 (sync_new_add_optab, "__sync_add_and_fetch", max);
init_sync_libfuncs_1 (sync_new_sub_optab, "__sync_sub_and_fetch", max);
- init_sync_libfuncs_1 (sync_new_ior_optab, "__sync_ior_and_fetch", max);
+ init_sync_libfuncs_1 (sync_new_ior_optab, "__sync_or_and_fetch", max);
init_sync_libfuncs_1 (sync_new_and_optab, "__sync_and_and_fetch", max);
init_sync_libfuncs_1 (sync_new_xor_optab, "__sync_xor_and_fetch", max);
init_sync_libfuncs_1 (sync_new_nand_optab, "__sync_nand_and_fetch", max);
return true;
}
-/* Return true if we can implement VEC_INTERLEAVE_{HIGH,LOW}_EXPR or
- VEC_EXTRACT_{EVEN,ODD}_EXPR with VEC_PERM_EXPR for this target.
- If PSEL is non-null, return the selector for the permutation. */
-
-bool
-can_vec_perm_for_code_p (enum tree_code code, enum machine_mode mode,
- rtx *psel)
-{
- bool need_sel_test = false;
- enum insn_code icode;
-
- /* If the target doesn't implement a vector mode for the vector type,
- then no operations are supported. */
- if (!VECTOR_MODE_P (mode))
- return false;
-
- /* Do as many tests as possible without reqiring the selector. */
- icode = direct_optab_handler (vec_perm_optab, mode);
- if (icode == CODE_FOR_nothing && GET_MODE_INNER (mode) != QImode)
- {
- enum machine_mode qimode
- = mode_for_vector (QImode, GET_MODE_SIZE (mode));
- if (VECTOR_MODE_P (qimode))
- icode = direct_optab_handler (vec_perm_optab, qimode);
- }
- if (icode == CODE_FOR_nothing)
- {
- icode = direct_optab_handler (vec_perm_const_optab, mode);
- if (icode != CODE_FOR_nothing
- && targetm.vectorize.vec_perm_const_ok != NULL)
- need_sel_test = true;
- }
- if (icode == CODE_FOR_nothing)
- return false;
-
- /* If the selector is required, or if we need to test it, build it. */
- if (psel || need_sel_test)
- {
- int i, nelt = GET_MODE_NUNITS (mode), alt = 0;
- unsigned char *data = XALLOCAVEC (unsigned char, nelt);
-
- switch (code)
- {
- case VEC_EXTRACT_ODD_EXPR:
- alt = 1;
- /* FALLTHRU */
- case VEC_EXTRACT_EVEN_EXPR:
- for (i = 0; i < nelt; ++i)
- data[i] = i * 2 + alt;
- break;
-
- case VEC_INTERLEAVE_HIGH_EXPR:
- case VEC_INTERLEAVE_LOW_EXPR:
- if ((BYTES_BIG_ENDIAN != 0) ^ (code == VEC_INTERLEAVE_HIGH_EXPR))
- alt = nelt / 2;
- for (i = 0; i < nelt / 2; ++i)
- {
- data[i * 2] = i + alt;
- data[i * 2 + 1] = i + nelt + alt;
- }
- break;
-
- default:
- gcc_unreachable ();
- }
-
- if (need_sel_test
- && !targetm.vectorize.vec_perm_const_ok (mode, data))
- return false;
-
- if (psel)
- {
- rtvec vec = rtvec_alloc (nelt);
- enum machine_mode imode = mode;
-
- for (i = 0; i < nelt; ++i)
- RTVEC_ELT (vec, i) = GEN_INT (data[i]);
-
- if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT)
- {
- imode = int_mode_for_mode (GET_MODE_INNER (mode));
- imode = mode_for_vector (imode, nelt);
- gcc_assert (GET_MODE_CLASS (imode) == MODE_VECTOR_INT);
- }
-
- *psel = gen_rtx_CONST_VECTOR (imode, vec);
- }
- }
-
- return true;
-}
-
/* A subroutine of expand_vec_perm for expanding one vec_perm insn. */
static rtx
}
/* If the input is a constant, expand it specially. */
- if (CONSTANT_P (sel))
+ gcc_assert (GET_MODE_CLASS (GET_MODE (sel)) == MODE_VECTOR_INT);
+ if (GET_CODE (sel) == CONST_VECTOR)
{
icode = direct_optab_handler (vec_perm_const_optab, mode);
if (icode != CODE_FOR_nothing)
{
unsigned int j, this_e;
- this_e = INTVAL (XVECEXP (sel, 0, i));
+ this_e = INTVAL (CONST_VECTOR_ELT (sel, i));
this_e &= 2 * e - 1;
this_e *= u;
rtx addr;
addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
- return emit_library_call_value (libfunc, target, LCT_NORMAL,
+ return emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
mode, 2, addr, ptr_mode,
val, mode);
}
return NULL_RTX;
}
+/* This function tries to implement an atomic test-and-set operation
+ using the atomic_test_and_set instruction pattern. A boolean value
+ is returned from the operation, using TARGET if possible. */
+
#ifndef HAVE_atomic_test_and_set
#define HAVE_atomic_test_and_set 0
-#define gen_atomic_test_and_set(x,y,z) (gcc_unreachable (), NULL_RTX)
+#define CODE_FOR_atomic_test_and_set CODE_FOR_nothing
#endif
+static rtx
+maybe_emit_atomic_test_and_set (rtx target, rtx mem, enum memmodel model)
+{
+ enum machine_mode pat_bool_mode;
+ struct expand_operand ops[3];
+
+ if (!HAVE_atomic_test_and_set)
+ return NULL_RTX;
+
+ /* While we always get QImode from __atomic_test_and_set, we get
+ other memory modes from __sync_lock_test_and_set. Note that we
+ use no endian adjustment here. This matches the 4.6 behavior
+ in the Sparc backend. */
+ gcc_checking_assert
+ (insn_data[CODE_FOR_atomic_test_and_set].operand[1].mode == QImode);
+ if (GET_MODE (mem) != QImode)
+ mem = adjust_address_nv (mem, QImode, 0);
+
+ pat_bool_mode = insn_data[CODE_FOR_atomic_test_and_set].operand[0].mode;
+ create_output_operand (&ops[0], target, pat_bool_mode);
+ create_fixed_operand (&ops[1], mem);
+ create_integer_operand (&ops[2], model);
+
+ if (maybe_expand_insn (CODE_FOR_atomic_test_and_set, 3, ops))
+ return ops[0].value;
+ return NULL_RTX;
+}
+
/* This function expands the legacy _sync_lock test_and_set operation which is
generally an atomic exchange. Some limited targets only allow the
constant 1 to be stored. This is an ACQUIRE operation.
/* Try an atomic_exchange first. */
ret = maybe_emit_atomic_exchange (target, mem, val, MEMMODEL_ACQUIRE);
+ if (ret)
+ return ret;
- if (!ret)
- ret = maybe_emit_sync_lock_test_and_set (target, mem, val,
- MEMMODEL_ACQUIRE);
- if (!ret)
- ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, val);
+ ret = maybe_emit_sync_lock_test_and_set (target, mem, val, MEMMODEL_ACQUIRE);
+ if (ret)
+ return ret;
+
+ ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, val);
+ if (ret)
+ return ret;
/* If there are no other options, try atomic_test_and_set if the value
being stored is 1. */
- if (!ret && val == const1_rtx && HAVE_atomic_test_and_set)
- {
- ret = gen_atomic_test_and_set (target, mem, GEN_INT (MEMMODEL_ACQUIRE));
- emit_insn (ret);
- }
+ if (val == const1_rtx)
+ ret = maybe_emit_atomic_test_and_set (target, mem, MEMMODEL_ACQUIRE);
return ret;
}
expand_atomic_test_and_set (rtx target, rtx mem, enum memmodel model)
{
enum machine_mode mode = GET_MODE (mem);
- rtx ret = NULL_RTX;
+ rtx ret, trueval, subtarget;
- if (target == NULL_RTX)
- target = gen_reg_rtx (mode);
+ ret = maybe_emit_atomic_test_and_set (target, mem, model);
+ if (ret)
+ return ret;
- if (HAVE_atomic_test_and_set)
+ /* Be binary compatible with non-default settings of trueval, and different
+ cpu revisions. E.g. one revision may have atomic-test-and-set, but
+ another only has atomic-exchange. */
+ if (targetm.atomic_test_and_set_trueval == 1)
+ {
+ trueval = const1_rtx;
+ subtarget = target ? target : gen_reg_rtx (mode);
+ }
+ else
{
- ret = gen_atomic_test_and_set (target, mem, GEN_INT (MEMMODEL_ACQUIRE));
- emit_insn (ret);
- return ret;
+ trueval = gen_int_mode (targetm.atomic_test_and_set_trueval, mode);
+ subtarget = gen_reg_rtx (mode);
}
- /* If there is no test and set, try exchange, then a compare_and_swap loop,
- then __sync_test_and_set. */
- ret = maybe_emit_atomic_exchange (target, mem, const1_rtx, model);
+ /* Try the atomic-exchange optab... */
+ ret = maybe_emit_atomic_exchange (subtarget, mem, trueval, model);
+ /* ... then an atomic-compare-and-swap loop ... */
if (!ret)
- ret = maybe_emit_compare_and_swap_exchange_loop (target, mem, const1_rtx);
+ ret = maybe_emit_compare_and_swap_exchange_loop (subtarget, mem, trueval);
+ /* ... before trying the vaguely defined legacy lock_test_and_set. */
if (!ret)
- ret = maybe_emit_sync_lock_test_and_set (target, mem, const1_rtx, model);
+ ret = maybe_emit_sync_lock_test_and_set (subtarget, mem, trueval, model);
- if (ret)
- return ret;
+ /* Recall that the legacy lock_test_and_set optab was allowed to do magic
+ things with the value 1. Thus we try again without trueval. */
+ if (!ret && targetm.atomic_test_and_set_trueval != 1)
+ ret = maybe_emit_sync_lock_test_and_set (subtarget, mem, const1_rtx, model);
- /* Failing all else, assume a single threaded environment and simply perform
- the operation. */
- emit_move_insn (target, mem);
- emit_move_insn (mem, const1_rtx);
- return target;
+ /* Failing all else, assume a single threaded environment and simply
+ perform the operation. */
+ if (!ret)
+ {
+ emit_move_insn (subtarget, mem);
+ emit_move_insn (mem, trueval);
+ ret = subtarget;
+ }
+
+ /* Recall that have to return a boolean value; rectify if trueval
+ is not exactly one. */
+ if (targetm.atomic_test_and_set_trueval != 1)
+ ret = emit_store_flag_force (target, NE, ret, const0_rtx, mode, 0, 1);
+
+ return ret;
}
/* This function expands the atomic exchange operation:
if (libfunc != NULL)
{
rtx addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
- target_oval = emit_library_call_value (libfunc, target_oval, LCT_NORMAL,
+ target_oval = emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
mode, 3, addr, ptr_mode,
expected, mode, desired, mode);
/* Issue val = compare_and_swap (mem, 0, 0).
This may cause the occasional harmless store of 0 when the value is
already 0, but it seems to be OK according to the standards guys. */
- expand_atomic_compare_and_swap (NULL, &target, mem, const0_rtx,
- const0_rtx, false, model, model);
- return target;
+ if (expand_atomic_compare_and_swap (NULL, &target, mem, const0_rtx,
+ const0_rtx, false, model, model))
+ return target;
+ else
+ /* Otherwise there is no atomic load, leave the library call. */
+ return NULL_RTX;
}
/* Otherwise assume loads are atomic, and emit the proper barriers. */
}
}
+/* See if there is a more optimal way to implement the operation "*MEM CODE VAL"
+ using memory order MODEL. If AFTER is true the operation needs to return
+ the value of *MEM after the operation, otherwise the previous value.
+ TARGET is an optional place to place the result. The result is unused if
+ it is const0_rtx.
+ Return the result if there is a better sequence, otherwise NULL_RTX. */
+
+static rtx
+maybe_optimize_fetch_op (rtx target, rtx mem, rtx val, enum rtx_code code,
+ enum memmodel model, bool after)
+{
+ /* If the value is prefetched, or not used, it may be possible to replace
+ the sequence with a native exchange operation. */
+ if (!after || target == const0_rtx)
+ {
+ /* fetch_and (&x, 0, m) can be replaced with exchange (&x, 0, m). */
+ if (code == AND && val == const0_rtx)
+ {
+ if (target == const0_rtx)
+ target = gen_reg_rtx (GET_MODE (mem));
+ return maybe_emit_atomic_exchange (target, mem, val, model);
+ }
+
+ /* fetch_or (&x, -1, m) can be replaced with exchange (&x, -1, m). */
+ if (code == IOR && val == constm1_rtx)
+ {
+ if (target == const0_rtx)
+ target = gen_reg_rtx (GET_MODE (mem));
+ return maybe_emit_atomic_exchange (target, mem, val, model);
+ }
+ }
+
+ return NULL_RTX;
+}
+
/* Try to emit an instruction for a specific operation varaition.
OPTAB contains the OP functions.
TARGET is an optional place to return the result. const0_rtx means unused.
get_atomic_op_for_code (&optab, code);
+ /* Check to see if there are any better instructions. */
+ result = maybe_optimize_fetch_op (target, mem, val, code, model, after);
+ if (result)
+ return result;
+
/* Check for the case where the result isn't used and try those patterns. */
if (unused_result)
{
{
/* If the result isn't used, no need to do compensation code. */
if (unused_result)
- return target;
+ return result;
/* Issue compensation code. Fetch_after == fetch_before OP val.
Fetch_before == after REVERSE_OP val. */
result = emit_library_call_value (libfunc, NULL, LCT_NORMAL, mode,
2, addr, ptr_mode, val, mode);
- if (unused_result)
- return target;
- if (fixup)
+ if (!unused_result && fixup)
result = expand_simple_binop (mode, code, result, val, target,
true, OPTAB_LIB_WIDEN);
return result;
return true;
/* If the operand is a memory whose address has no side effects,
- try forcing the address into a register. The check for side
- effects is important because force_reg cannot handle things
- like auto-modified addresses. */
- if (insn_data[(int) icode].operand[opno].allows_mem
- && MEM_P (op->value)
- && !side_effects_p (XEXP (op->value, 0)))
- {
- rtx addr, mem, last;
-
- last = get_last_insn ();
- addr = force_reg (Pmode, XEXP (op->value, 0));
- mem = replace_equiv_address (op->value, addr);
- if (insn_operand_matches (icode, opno, mem))
+ try forcing the address into a non-virtual pseudo register.
+ The check for side effects is important because copy_to_mode_reg
+ cannot handle things like auto-modified addresses. */
+ if (insn_data[(int) icode].operand[opno].allows_mem && MEM_P (op->value))
+ {
+ rtx addr, mem;
+
+ mem = op->value;
+ addr = XEXP (mem, 0);
+ if (!(REG_P (addr) && REGNO (addr) > LAST_VIRTUAL_REGISTER)
+ && !side_effects_p (addr))
{
- op->value = mem;
- return true;
+ rtx last;
+ enum machine_mode mode;
+
+ last = get_last_insn ();
+ mode = targetm.addr_space.address_mode (MEM_ADDR_SPACE (mem));
+ mem = replace_equiv_address (mem, copy_to_mode_reg (mode, addr));
+ if (insn_operand_matches (icode, opno, mem))
+ {
+ op->value = mem;
+ return true;
+ }
+ delete_insns_since (last);
}
- delete_insns_since (last);
}
return false;