static void sparc_print_operand (FILE *, rtx, int);
static void sparc_print_operand_address (FILE *, rtx);
static reg_class_t sparc_secondary_reload (bool, rtx, reg_class_t,
- enum machine_mode, secondary_reload_info *);
+ enum machine_mode,
+ secondary_reload_info *);
\f
#ifdef SUBTARGET_ATTRIBUTE_TABLE
/* Table of valid machine attributes. */
gcc_unreachable ();
};
+ if (sparc_memory_model == SMM_DEFAULT)
+ {
+ /* Choose the memory model for the operating system. */
+ enum sparc_memory_model_type os_default = SUBTARGET_DEFAULT_MEMORY_MODEL;
+ if (os_default != SMM_DEFAULT)
+ sparc_memory_model = os_default;
+ /* Choose the most relaxed model for the processor. */
+ else if (TARGET_V9)
+ sparc_memory_model = SMM_RMO;
+ else if (TARGET_V8)
+ sparc_memory_model = SMM_PSO;
+ else
+ sparc_memory_model = SMM_SC;
+ }
+
#ifdef TARGET_DEFAULT_LONG_DOUBLE_128
if (!(target_flags_explicit & MASK_LONG_DOUBLE_128))
target_flags |= MASK_LONG_DOUBLE_128;
else if (size <= 8192)
{
insn = emit_insn (gen_stack_pointer_inc (GEN_INT (-4096)));
- /* %sp is still the CFA register. */
RTX_FRAME_RELATED_P (insn) = 1;
+
+ /* %sp is still the CFA register. */
insn = emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
}
else
else if (size <= 8192)
{
emit_window_save (GEN_INT (-4096));
+
/* %sp is not the CFA register anymore. */
emit_insn (gen_stack_pointer_inc (GEN_INT (4096 - size)));
+
+ /* Make sure no %fp-based store is issued until after the frame is
+ established. The offset between the frame pointer and the stack
+ pointer is calculated relative to the value of the stack pointer
+ at the end of the function prologue, and moving instructions that
+ access the stack via the frame pointer between the instructions
+ that decrement the stack pointer could result in accessing the
+ register window save area, which is volatile. */
+ emit_insn (gen_frame_blockage ());
}
else
{
return 0;
}
-/* Returns assembly code to perform a DImode shift using
- a 64-bit global or out register on SPARC-V8+. */
+/* Output a wide shift instruction in V8+ mode. INSN is the instruction,
+ OPERANDS are its operands and OPCODE is the mnemonic to be used. */
+
const char *
-output_v8plus_shift (rtx *operands, rtx insn, const char *opcode)
+output_v8plus_shift (rtx insn, rtx *operands, const char *opcode)
{
static char asm_code[60];
output_asm_insn ("or\t%L1, %3, %3", operands);
}
- strcpy(asm_code, opcode);
+ strcpy (asm_code, opcode);
if (which_alternative != 2)
return strcat (asm_code, "\t%0, %2, %L0\n\tsrlx\t%L0, 32, %H0");
else
- return strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
+ return
+ strcat (asm_code, "\t%3, %2, %3\n\tsrlx\t%3, 32, %H0\n\tmov\t%3, %L0");
}
\f
/* Output rtl to increment the profiler label LABELNO
}
#endif
+/* Expand a membar instruction for various use cases. Both the LOAD_STORE
+ and BEFORE_AFTER arguments of the form X_Y. They are two-bit masks where
+ bit 0 indicates that X is true, and bit 1 indicates Y is true. */
+
+void
+sparc_emit_membar_for_model (enum memmodel model,
+ int load_store, int before_after)
+{
+ /* Bits for the MEMBAR mmask field. */
+ const int LoadLoad = 1;
+ const int StoreLoad = 2;
+ const int LoadStore = 4;
+ const int StoreStore = 8;
+
+ int mm = 0, implied = 0;
+
+ switch (sparc_memory_model)
+ {
+ case SMM_SC:
+ /* Sequential Consistency. All memory transactions are immediately
+ visible in sequential execution order. No barriers needed. */
+ implied = LoadLoad | StoreLoad | LoadStore | StoreStore;
+ break;
+
+ case SMM_TSO:
+ /* Total Store Ordering: all memory transactions with store semantics
+ are followed by an implied StoreStore. */
+ implied |= StoreStore;
+ /* FALLTHRU */
+
+ case SMM_PSO:
+ /* Partial Store Ordering: all memory transactions with load semantics
+ are followed by an implied LoadLoad | LoadStore. */
+ implied |= LoadLoad | LoadStore;
+
+ /* If we're not looking for a raw barrer (before+after), then atomic
+ operations get the benefit of being both load and store. */
+ if (load_store == 3 && before_after == 2)
+ implied |= StoreLoad | StoreStore;
+ /* FALLTHRU */
+
+ case SMM_RMO:
+ /* Relaxed Memory Ordering: no implicit bits. */
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ if (before_after & 1)
+ {
+ if (model == MEMMODEL_ACQUIRE
+ || model == MEMMODEL_ACQ_REL
+ || model == MEMMODEL_SEQ_CST)
+ {
+ if (load_store & 1)
+ mm |= LoadLoad | LoadStore;
+ if (load_store & 2)
+ mm |= StoreLoad | StoreStore;
+ }
+ }
+ if (before_after & 2)
+ {
+ if (model == MEMMODEL_RELEASE
+ || model == MEMMODEL_ACQ_REL
+ || model == MEMMODEL_SEQ_CST)
+ {
+ if (load_store & 1)
+ mm |= LoadLoad | StoreLoad;
+ if (load_store & 2)
+ mm |= LoadStore | StoreStore;
+ }
+ }
+
+ /* Remove the bits implied by the system memory model. */
+ mm &= ~implied;
+
+ /* For raw barriers (before+after), always emit a barrier.
+ This will become a compile-time barrier if needed. */
+ if (mm || before_after == 3)
+ emit_insn (gen_membar (GEN_INT (mm)));
+}
+
/* Expand code to perform a 8 or 16-bit compare and swap by doing 32-bit
compare and swap on the word containing the byte or half-word. */
-void
-sparc_expand_compare_and_swap_12 (rtx result, rtx mem, rtx oldval, rtx newval)
+static void
+sparc_expand_compare_and_swap_12 (rtx bool_result, rtx result, rtx mem,
+ rtx oldval, rtx newval)
{
rtx addr1 = force_reg (Pmode, XEXP (mem, 0));
rtx addr = gen_reg_rtx (Pmode);
set_mem_alias_set (memsi, ALIAS_SET_MEMORY_BARRIER);
MEM_VOLATILE_P (memsi) = MEM_VOLATILE_P (mem);
- val = force_reg (SImode, memsi);
+ val = copy_to_reg (memsi);
emit_insn (gen_rtx_SET (VOIDmode, off,
gen_rtx_XOR (SImode, off,
emit_insn (gen_rtx_SET (VOIDmode, newvalue,
gen_rtx_IOR (SImode, newv, val)));
- emit_insn (gen_sync_compare_and_swapsi (res, memsi, oldvalue, newvalue));
+ emit_move_insn (bool_result, const1_rtx);
+
+ emit_insn (gen_atomic_compare_and_swapsi_1 (res, memsi, oldvalue, newvalue));
emit_cmp_and_jump_insns (res, oldvalue, EQ, NULL, SImode, 0, end_label);
gen_rtx_AND (SImode, gen_rtx_NOT (SImode, mask),
res)));
+ emit_move_insn (bool_result, const0_rtx);
+
cc = gen_compare_reg_1 (NE, resv, val);
emit_insn (gen_rtx_SET (VOIDmode, val, resv));
emit_move_insn (result, gen_lowpart (GET_MODE (result), res));
}
+/* Expand code to perform a compare-and-swap. */
+
+void
+sparc_expand_compare_and_swap (rtx operands[])
+{
+ rtx bval, retval, mem, oldval, newval;
+ enum machine_mode mode;
+ enum memmodel model;
+
+ bval = operands[0];
+ retval = operands[1];
+ mem = operands[2];
+ oldval = operands[3];
+ newval = operands[4];
+ model = (enum memmodel) INTVAL (operands[6]);
+ mode = GET_MODE (mem);
+
+ sparc_emit_membar_for_model (model, 3, 1);
+
+ if (reg_overlap_mentioned_p (retval, oldval))
+ oldval = copy_to_reg (oldval);
+
+ if (mode == QImode || mode == HImode)
+ sparc_expand_compare_and_swap_12 (bval, retval, mem, oldval, newval);
+ else
+ {
+ rtx (*gen) (rtx, rtx, rtx, rtx);
+ rtx x;
+
+ if (mode == SImode)
+ gen = gen_atomic_compare_and_swapsi_1;
+ else
+ gen = gen_atomic_compare_and_swapdi_1;
+ emit_insn (gen (retval, mem, oldval, newval));
+
+ x = emit_store_flag (bval, EQ, retval, oldval, mode, 1, 1);
+ if (x != bval)
+ convert_move (bval, x, 1);
+ }
+
+ sparc_emit_membar_for_model (model, 3, 2);
+}
+
void
sparc_expand_vec_perm_bmask (enum machine_mode vmode, rtx sel)
{
global_regs[SPARC_GSR_REG] = 1;
}
-/* Implement TARGET_PREFERRED_RELOAD_CLASS
+/* Implement TARGET_PREFERRED_RELOAD_CLASS:
- We can't load constants into FP registers.
- We can't load FP constants into integer registers when soft-float,
return rclass;
}
+/* Output a wide multiply instruction in V8+ mode. INSN is the instruction,
+ OPERANDS are its operands and OPCODE is the mnemonic to be used. */
+
const char *
-output_v8plus_mult (rtx insn, rtx *operands, const char *name)
+output_v8plus_mult (rtx insn, rtx *operands, const char *opcode)
{
char mulstr[32];
if (which_alternative == 1)
{
output_asm_insn ("or\t%L1, %H1, %H1", operands);
- sprintf (mulstr, "%s\t%%H1, %%2, %%L0", name);
+ sprintf (mulstr, "%s\t%%H1, %%2, %%L0", opcode);
output_asm_insn (mulstr, operands);
return "srlx\t%L0, 32, %H0";
}
{
output_asm_insn ("sllx\t%H1, 32, %3", operands);
output_asm_insn ("or\t%L1, %3, %3", operands);
- sprintf (mulstr, "%s\t%%3, %%2, %%3", name);
+ sprintf (mulstr, "%s\t%%3, %%2, %%3", opcode);
output_asm_insn (mulstr, operands);
output_asm_insn ("srlx\t%3, 32, %H0", operands);
return "mov\t%3, %L0";
if (which_alternative == 1)
{
output_asm_insn ("or\t%L1, %H1, %H1", operands);
- sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", name);
+ sprintf (mulstr, "%s\t%%H1, %%H1, %%L0", opcode);
output_asm_insn (mulstr, operands);
return "srlx\t%L0, 32, %H0";
}
{
output_asm_insn ("sllx\t%H1, 32, %3", operands);
output_asm_insn ("or\t%L1, %3, %3", operands);
- sprintf (mulstr, "%s\t%%3, %%3, %%3", name);
+ sprintf (mulstr, "%s\t%%3, %%3, %%3", opcode);
output_asm_insn (mulstr, operands);
output_asm_insn ("srlx\t%3, 32, %H0", operands);
return "mov\t%3, %L0";
output_asm_insn ("or\t%L1, %H1, %H1", operands);
output_asm_insn ("sllx\t%H2, 32, %L1", operands);
output_asm_insn ("or\t%L2, %L1, %L1", operands);
- sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", name);
+ sprintf (mulstr, "%s\t%%H1, %%L1, %%L0", opcode);
output_asm_insn (mulstr, operands);
return "srlx\t%L0, 32, %H0";
}
output_asm_insn ("sllx\t%H2, 32, %4", operands);
output_asm_insn ("or\t%L1, %3, %3", operands);
output_asm_insn ("or\t%L2, %4, %4", operands);
- sprintf (mulstr, "%s\t%%3, %%4, %%3", name);
+ sprintf (mulstr, "%s\t%%3, %%4, %%3", opcode);
output_asm_insn (mulstr, operands);
output_asm_insn ("srlx\t%3, 32, %H0", operands);
return "mov\t%3, %L0";
}
}
+/* Subroutine of sparc_expand_vector_init. Emit code to initialize
+ all fields of TARGET to ELT by means of VIS2 BSHUFFLE insn. MODE
+ and INNER_MODE are the modes describing TARGET. */
+
static void
vector_init_bshuffle (rtx target, rtx elt, enum machine_mode mode,
enum machine_mode inner_mode)
{
- rtx t1, final_insn;
- int bmask;
+ rtx t1, final_insn;
+ int bmask;
- t1 = gen_reg_rtx (mode);
+ t1 = gen_reg_rtx (mode);
- elt = convert_modes (SImode, inner_mode, elt, true);
- emit_move_insn (gen_lowpart(SImode, t1), elt);
+ elt = convert_modes (SImode, inner_mode, elt, true);
+ emit_move_insn (gen_lowpart(SImode, t1), elt);
- switch (mode)
- {
- case V2SImode:
- final_insn = gen_bshufflev2si_vis (target, t1, t1);
- bmask = 0x45674567;
- break;
- case V4HImode:
- final_insn = gen_bshufflev4hi_vis (target, t1, t1);
- bmask = 0x67676767;
- break;
- case V8QImode:
- final_insn = gen_bshufflev8qi_vis (target, t1, t1);
- bmask = 0x77777777;
- break;
- default:
- gcc_unreachable ();
- }
+ switch (mode)
+ {
+ case V2SImode:
+ final_insn = gen_bshufflev2si_vis (target, t1, t1);
+ bmask = 0x45674567;
+ break;
+ case V4HImode:
+ final_insn = gen_bshufflev4hi_vis (target, t1, t1);
+ bmask = 0x67676767;
+ break;
+ case V8QImode:
+ final_insn = gen_bshufflev8qi_vis (target, t1, t1);
+ bmask = 0x77777777;
+ break;
+ default:
+ gcc_unreachable ();
+ }
- emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
- force_reg (SImode, GEN_INT (bmask))));
- emit_insn (final_insn);
+ emit_insn (gen_bmasksi_vis (gen_reg_rtx (SImode), CONST0_RTX (SImode),
+ force_reg (SImode, GEN_INT (bmask))));
+ emit_insn (final_insn);
}
static void
emit_insn (gen_faligndatav4hi_vis (target, t1, target));
}
+/* Emit code to initialize TARGET to values for individual fields VALS. */
+
void
sparc_expand_vector_init (rtx target, rtx vals)
{
mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
for (i = 0; i < n_elts; i++)
emit_move_insn (adjust_address_nv (mem, inner_mode,
- i * GET_MODE_SIZE (inner_mode)),
+ i * GET_MODE_SIZE (inner_mode)),
XVECEXP (vals, 0, i));
emit_move_insn (target, mem);
}
+/* Implement TARGET_SECONDARY_RELOAD. */
+
static reg_class_t
sparc_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
enum machine_mode mode, secondary_reload_info *sri)
return NO_REGS;
}
+/* Emit code to conditionally move either OPERANDS[2] or OPERANDS[3] into
+ OPERANDS[0] in MODE. OPERANDS[1] is the operator of the condition. */
+
bool
sparc_expand_conditional_move (enum machine_mode mode, rtx *operands)
{
rtx cc_reg, dst, cmp;
cmp = operands[1];
- cmp_mode = GET_MODE (XEXP (cmp, 0));
- if (cmp_mode == DImode && !TARGET_ARCH64)
+ if (GET_MODE (XEXP (cmp, 0)) == DImode && !TARGET_ARCH64)
return false;
- dst = operands[0];
+ if (GET_MODE (XEXP (cmp, 0)) == TFmode && !TARGET_HARD_QUAD)
+ cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
+
+ cmp_mode = GET_MODE (XEXP (cmp, 0));
+ rc = GET_CODE (cmp);
+ dst = operands[0];
if (! rtx_equal_p (operands[2], dst)
&& ! rtx_equal_p (operands[3], dst))
{
rc = reverse_condition (rc);
}
- if (cmp_mode == TFmode && !TARGET_HARD_QUAD)
- cmp = sparc_emit_float_lib_cmp (XEXP (cmp, 0), XEXP (cmp, 1), rc);
-
if (XEXP (cmp, 1) == const0_rtx
&& GET_CODE (XEXP (cmp, 0)) == REG
&& cmp_mode == DImode
return true;
}
+/* Emit code to conditionally move a combination of OPERANDS[1] and OPERANDS[2]
+ into OPERANDS[0] in MODE, depending on the outcome of the comparison of
+ OPERANDS[4] and OPERANDS[5]. OPERANDS[3] is the operator of the condition.
+ FCODE is the machine code to be used for OPERANDS[3] and CCODE the machine
+ code to be used for the condition mask. */
+
void
sparc_expand_vcond (enum machine_mode mode, rtx *operands, int ccode, int fcode)
{
emit_insn (gen_rtx_SET (VOIDmode, operands[0], bshuf));
}
+/* On sparc, any mode which naturally allocates into the float
+ registers should return 4 here. */
+
+unsigned int
+sparc_regmode_natural_size (enum machine_mode mode)
+{
+ int size = UNITS_PER_WORD;
+
+ if (TARGET_ARCH64)
+ {
+ enum mode_class mclass = GET_MODE_CLASS (mode);
+
+ if (mclass == MODE_FLOAT || mclass == MODE_VECTOR_INT)
+ size = 4;
+ }
+
+ return size;
+}
+
+/* Return TRUE if it is a good idea to tie two pseudo registers
+ when one has mode MODE1 and one has mode MODE2.
+ If HARD_REGNO_MODE_OK could produce different values for MODE1 and MODE2,
+ for any hard reg, then this must be FALSE for correct output.
+
+ For V9 we have to deal with the fact that only the lower 32 floating
+ point registers are 32-bit addressable. */
+
+bool
+sparc_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
+{
+ enum mode_class mclass1, mclass2;
+ unsigned short size1, size2;
+
+ if (mode1 == mode2)
+ return true;
+
+ mclass1 = GET_MODE_CLASS (mode1);
+ mclass2 = GET_MODE_CLASS (mode2);
+ if (mclass1 != mclass2)
+ return false;
+
+ if (! TARGET_V9)
+ return true;
+
+ /* Classes are the same and we are V9 so we have to deal with upper
+ vs. lower floating point registers. If one of the modes is a
+ 4-byte mode, and the other is not, we have to mark them as not
+ tieable because only the lower 32 floating point register are
+ addressable 32-bits at a time.
+
+ We can't just test explicitly for SFmode, otherwise we won't
+ cover the vector mode cases properly. */
+
+ if (mclass1 != MODE_FLOAT && mclass1 != MODE_VECTOR_INT)
+ return true;
+
+ size1 = GET_MODE_SIZE (mode1);
+ size2 = GET_MODE_SIZE (mode2);
+ if ((size1 > 4 && size2 == 4)
+ || (size2 > 4 && size1 == 4))
+ return false;
+
+ return true;
+}
+
#include "gt-sparc.h"