static int memory_address_length (rtx addr);
static int ix86_flags_dependant (rtx, rtx, enum attr_type);
static int ix86_agi_dependant (rtx, rtx, enum attr_type);
-static enum attr_ppro_uops ix86_safe_ppro_uops (rtx);
-static void ix86_dump_ppro_packet (FILE *);
-static void ix86_reorder_insn (rtx *, rtx *);
static struct machine_function * ix86_init_machine_status (void);
static int ix86_split_to_parts (rtx, rtx *, enum machine_mode);
static int ix86_nsaved_regs (void);
static void ix86_emit_save_regs_using_mov (rtx, HOST_WIDE_INT);
static void ix86_emit_restore_regs_using_mov (rtx, HOST_WIDE_INT, int);
static void ix86_output_function_epilogue (FILE *, HOST_WIDE_INT);
-static void ix86_sched_reorder_ppro (rtx *, rtx *);
static HOST_WIDE_INT ix86_GOT_alias_set (void);
static void ix86_adjust_counter (rtx, HOST_WIDE_INT);
static rtx ix86_expand_aligntest (rtx, int);
static void ix86_expand_strlensi_unroll_1 (rtx, rtx, rtx);
static int ix86_issue_rate (void);
static int ix86_adjust_cost (rtx, rtx, rtx, int);
-static void ix86_sched_init (FILE *, int, int);
-static int ix86_sched_reorder (FILE *, int, rtx *, int *, int);
-static int ix86_variable_issue (FILE *, int, rtx, int);
static int ia32_use_dfa_pipeline_interface (void);
static int ia32_multipass_dfa_lookahead (void);
static void ix86_init_mmx_sse_builtins (void);
#define TARGET_SCHED_ADJUST_COST ix86_adjust_cost
#undef TARGET_SCHED_ISSUE_RATE
#define TARGET_SCHED_ISSUE_RATE ix86_issue_rate
-#undef TARGET_SCHED_VARIABLE_ISSUE
-#define TARGET_SCHED_VARIABLE_ISSUE ix86_variable_issue
-#undef TARGET_SCHED_INIT
-#define TARGET_SCHED_INIT ix86_sched_init
-#undef TARGET_SCHED_REORDER
-#define TARGET_SCHED_REORDER ix86_sched_reorder
#undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE
#define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE \
ia32_use_dfa_pipeline_interface
return cost;
}
-static union
-{
- struct ppro_sched_data
- {
- rtx decode[3];
- int issued_this_cycle;
- } ppro;
-} ix86_sched_data;
-
-static enum attr_ppro_uops
-ix86_safe_ppro_uops (rtx insn)
-{
- if (recog_memoized (insn) >= 0)
- return get_attr_ppro_uops (insn);
- else
- return PPRO_UOPS_MANY;
-}
-
-static void
-ix86_dump_ppro_packet (FILE *dump)
-{
- if (ix86_sched_data.ppro.decode[0])
- {
- fprintf (dump, "PPRO packet: %d",
- INSN_UID (ix86_sched_data.ppro.decode[0]));
- if (ix86_sched_data.ppro.decode[1])
- fprintf (dump, " %d", INSN_UID (ix86_sched_data.ppro.decode[1]));
- if (ix86_sched_data.ppro.decode[2])
- fprintf (dump, " %d", INSN_UID (ix86_sched_data.ppro.decode[2]));
- fputc ('\n', dump);
- }
-}
-
-/* We're beginning a new block. Initialize data structures as necessary. */
-
-static void
-ix86_sched_init (FILE *dump ATTRIBUTE_UNUSED,
- int sched_verbose ATTRIBUTE_UNUSED,
- int veclen ATTRIBUTE_UNUSED)
-{
- memset (&ix86_sched_data, 0, sizeof (ix86_sched_data));
-}
-
-/* Shift INSN to SLOT, and shift everything else down. */
-
-static void
-ix86_reorder_insn (rtx *insnp, rtx *slot)
-{
- if (insnp != slot)
- {
- rtx insn = *insnp;
- do
- insnp[0] = insnp[1];
- while (++insnp != slot);
- *insnp = insn;
- }
-}
-
-static void
-ix86_sched_reorder_ppro (rtx *ready, rtx *e_ready)
-{
- rtx decode[3];
- enum attr_ppro_uops cur_uops;
- int issued_this_cycle;
- rtx *insnp;
- int i;
-
- /* At this point .ppro.decode contains the state of the three
- decoders from last "cycle". That is, those insns that were
- actually independent. But here we're scheduling for the
- decoder, and we may find things that are decodable in the
- same cycle. */
-
- memcpy (decode, ix86_sched_data.ppro.decode, sizeof (decode));
- issued_this_cycle = 0;
-
- insnp = e_ready;
- cur_uops = ix86_safe_ppro_uops (*insnp);
-
- /* If the decoders are empty, and we've a complex insn at the
- head of the priority queue, let it issue without complaint. */
- if (decode[0] == NULL)
- {
- if (cur_uops == PPRO_UOPS_MANY)
- {
- decode[0] = *insnp;
- goto ppro_done;
- }
-
- /* Otherwise, search for a 2-4 uop unsn to issue. */
- while (cur_uops != PPRO_UOPS_FEW)
- {
- if (insnp == ready)
- break;
- cur_uops = ix86_safe_ppro_uops (*--insnp);
- }
-
- /* If so, move it to the head of the line. */
- if (cur_uops == PPRO_UOPS_FEW)
- ix86_reorder_insn (insnp, e_ready);
-
- /* Issue the head of the queue. */
- issued_this_cycle = 1;
- decode[0] = *e_ready--;
- }
-
- /* Look for simple insns to fill in the other two slots. */
- for (i = 1; i < 3; ++i)
- if (decode[i] == NULL)
- {
- if (ready > e_ready)
- goto ppro_done;
-
- insnp = e_ready;
- cur_uops = ix86_safe_ppro_uops (*insnp);
- while (cur_uops != PPRO_UOPS_ONE)
- {
- if (insnp == ready)
- break;
- cur_uops = ix86_safe_ppro_uops (*--insnp);
- }
-
- /* Found one. Move it to the head of the queue and issue it. */
- if (cur_uops == PPRO_UOPS_ONE)
- {
- ix86_reorder_insn (insnp, e_ready);
- decode[i] = *e_ready--;
- issued_this_cycle++;
- continue;
- }
-
- /* ??? Didn't find one. Ideally, here we would do a lazy split
- of 2-uop insns, issue one and queue the other. */
- }
-
- ppro_done:
- if (issued_this_cycle == 0)
- issued_this_cycle = 1;
- ix86_sched_data.ppro.issued_this_cycle = issued_this_cycle;
-}
-
-/* We are about to being issuing insns for this clock cycle.
- Override the default sort algorithm to better slot instructions. */
-static int
-ix86_sched_reorder (FILE *dump ATTRIBUTE_UNUSED,
- int sched_verbose ATTRIBUTE_UNUSED, rtx *ready,
- int *n_readyp, int clock_var ATTRIBUTE_UNUSED)
-{
- int n_ready = *n_readyp;
- rtx *e_ready = ready + n_ready - 1;
-
- /* Make sure to go ahead and initialize key items in
- ix86_sched_data if we are not going to bother trying to
- reorder the ready queue. */
- if (n_ready < 2)
- {
- ix86_sched_data.ppro.issued_this_cycle = 1;
- goto out;
- }
-
- switch (ix86_tune)
- {
- default:
- break;
-
- case PROCESSOR_PENTIUMPRO:
- ix86_sched_reorder_ppro (ready, e_ready);
- break;
- }
-
-out:
- return ix86_issue_rate ();
-}
-
-/* We are about to issue INSN. Return the number of insns left on the
- ready queue that can be issued this cycle. */
-
-static int
-ix86_variable_issue (FILE *dump, int sched_verbose, rtx insn,
- int can_issue_more)
-{
- int i;
- switch (ix86_tune)
- {
- default:
- return can_issue_more - 1;
-
- case PROCESSOR_PENTIUMPRO:
- {
- enum attr_ppro_uops uops = ix86_safe_ppro_uops (insn);
-
- if (uops == PPRO_UOPS_MANY)
- {
- if (sched_verbose)
- ix86_dump_ppro_packet (dump);
- ix86_sched_data.ppro.decode[0] = insn;
- ix86_sched_data.ppro.decode[1] = NULL;
- ix86_sched_data.ppro.decode[2] = NULL;
- if (sched_verbose)
- ix86_dump_ppro_packet (dump);
- ix86_sched_data.ppro.decode[0] = NULL;
- }
- else if (uops == PPRO_UOPS_FEW)
- {
- if (sched_verbose)
- ix86_dump_ppro_packet (dump);
- ix86_sched_data.ppro.decode[0] = insn;
- ix86_sched_data.ppro.decode[1] = NULL;
- ix86_sched_data.ppro.decode[2] = NULL;
- }
- else
- {
- for (i = 0; i < 3; ++i)
- if (ix86_sched_data.ppro.decode[i] == NULL)
- {
- ix86_sched_data.ppro.decode[i] = insn;
- break;
- }
- if (i == 3)
- abort ();
- if (i == 2)
- {
- if (sched_verbose)
- ix86_dump_ppro_packet (dump);
- ix86_sched_data.ppro.decode[0] = NULL;
- ix86_sched_data.ppro.decode[1] = NULL;
- ix86_sched_data.ppro.decode[2] = NULL;
- }
- }
- }
- return --ix86_sched_data.ppro.issued_this_cycle;
- }
-}
-
static int
ia32_use_dfa_pipeline_interface (void)
{
- if (TARGET_PENTIUM || TARGET_ATHLON_K8)
+ if (TARGET_PENTIUM
+ || TARGET_PENTIUMPRO
+ || TARGET_ATHLON_K8)
return 1;
return 0;
}
{
if (ix86_tune == PROCESSOR_PENTIUM)
return 2;
+
+ if (ix86_tune == PROCESSOR_PENTIUMPRO)
+ return 1;
+
else
- return 0;
+ return 0;
}
\f
(const_int 1)
(const_int 0)))
-;; Set when 0f opcode prefix is used.
+;; Set when REX opcode prefix is used.
(define_attr "prefix_rex" ""
(cond [(and (eq_attr "mode" "DI")
(eq_attr "type" "!push,pop,call,callv,leave,ibr"))
"fnstsw\t%0"
[(set_attr "length" "2")
(set_attr "mode" "SI")
- (set_attr "unit" "i387")
- (set_attr "ppro_uops" "few")])
+ (set_attr "unit" "i387")])
;; FP compares, step 3
;; Get ax into flags, general case.
"sahf"
[(set_attr "length" "1")
(set_attr "athlon_decode" "vector")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "one")])
+ (set_attr "mode" "SI")])
;; Pentium Pro can do steps 1 through 3 in one go.
(set_attr "pent_pair" "np")
(set_attr "athlon_decode" "vector")
(set_attr "mode" "SI")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "few")])
+ (set_attr "modrm" "0")])
(define_expand "movhi"
[(set (match_operand:HI 0 "nonimmediate_operand" "")
[(set_attr "type" "imov")
(set_attr "pent_pair" "np")
(set_attr "mode" "HI")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "few")])
+ (set_attr "modrm" "0")])
(define_insn "*swaphi_2"
[(set (match_operand:HI 0 "register_operand" "+r")
[(set_attr "type" "imov")
(set_attr "pent_pair" "np")
(set_attr "mode" "SI")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "few")])
+ (set_attr "modrm" "0")])
(define_expand "movstricthi"
[(set (strict_low_part (match_operand:HI 0 "nonimmediate_operand" ""))
[(set_attr "type" "imov")
(set_attr "pent_pair" "np")
(set_attr "mode" "QI")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "few")])
+ (set_attr "modrm" "0")])
(define_expand "movstrictqi"
[(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" ""))
(set_attr "pent_pair" "np")
(set_attr "athlon_decode" "vector")
(set_attr "mode" "DI")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "few")])
+ (set_attr "modrm" "0")])
(define_expand "movsf"
"fnstcw\t%0"
[(set_attr "length" "2")
(set_attr "mode" "HI")
- (set_attr "unit" "i387")
- (set_attr "ppro_uops" "few")])
+ (set_attr "unit" "i387")])
(define_insn "x86_fldcw_1"
[(set (reg:HI 18)
[(set_attr "length" "2")
(set_attr "mode" "HI")
(set_attr "unit" "i387")
- (set_attr "athlon_decode" "vector")
- (set_attr "ppro_uops" "few")])
+ (set_attr "athlon_decode" "vector")])
\f
;; Conversion between fixed point and floating point.
"adc{q}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "mode" "DI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "DI")])
(define_insn "*adddi3_cc_rex64"
[(set (reg:CC 17)
"adc{b}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "mode" "QI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "QI")])
(define_insn "addhi3_carry"
[(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
"adc{w}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "mode" "HI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "HI")])
(define_insn "addsi3_carry"
[(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
"adc{l}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "SI")])
(define_insn "*addsi3_carry_zext"
[(set (match_operand:DI 0 "register_operand" "=r")
"adc{l}\t{%2, %k0|%k0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "SI")])
(define_insn "*addsi3_cc"
[(set (reg:CC 17)
"sbb{q}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "DI")])
(define_insn "*subdi_1_rex64"
"sbb{b}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "QI")])
(define_insn "subhi3_carry"
"sbb{w}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "HI")])
(define_insn "subsi3_carry"
"sbb{l}\t{%2, %0|%0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "SI")])
(define_insn "subsi3_carry_zext"
"sbb{l}\t{%2, %k0|%k0, %2}"
[(set_attr "type" "alu")
(set_attr "pent_pair" "pu")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "SI")])
(define_expand "subsi3"
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"mul{q}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set_attr "length_immediate" "0")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"mul{l}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set_attr "length_immediate" "0")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"mul{q}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set_attr "length_immediate" "0")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
"GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM"
"mul{l}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set_attr "length_immediate" "0")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"mul{l}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set_attr "length_immediate" "0")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"imul{q}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
(const_string "vector")
"GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM"
"imul{l}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
(const_string "vector")
&& (GET_CODE (operands[1]) != MEM || GET_CODE (operands[2]) != MEM)"
"imul{l}\t%2"
[(set_attr "type" "imul")
- (set_attr "ppro_uops" "few")
(set (attr "athlon_decode")
(if_then_else (eq_attr "cpu" "athlon")
(const_string "vector")
"TARGET_QIMODE_MATH"
"idiv{b}\t%2"
[(set_attr "type" "idiv")
- (set_attr "mode" "QI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "QI")])
(define_insn "udivqi3"
[(set (match_operand:QI 0 "register_operand" "=a")
"TARGET_QIMODE_MATH"
"div{b}\t%2"
[(set_attr "type" "idiv")
- (set_attr "mode" "QI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "QI")])
;; The patterns that match these are at the end of this file.
"TARGET_64BIT"
"idiv{q}\t%2"
[(set_attr "type" "idiv")
- (set_attr "mode" "DI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "DI")])
(define_split
[(set (match_operand:DI 0 "register_operand" "")
""
"idiv{l}\t%2"
[(set_attr "type" "idiv")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "SI")])
(define_split
[(set (match_operand:SI 0 "register_operand" "")
"TARGET_64BIT"
"div{q}\t%2"
[(set_attr "type" "idiv")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "DI")])
(define_split
""
"div{l}\t%2"
[(set_attr "type" "idiv")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "SI")])
(define_split
""
"div{w}\t%2"
[(set_attr "type" "idiv")
- (set_attr "mode" "HI")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "HI")])
;; We can not use div/idiv for double division, because it causes
;; "division by zero" on the overflow and that's not what we expect
; (clobber (reg:CC 17))]
; ""
; "div{l}\t{%2, %0|%0, %2}"
-; [(set_attr "type" "idiv")
-; (set_attr "ppro_uops" "few")])
+; [(set_attr "type" "idiv")])
\f
;;- Logical AND instructions
"TARGET_80387 && reload_completed"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "SF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "SF")])
(define_insn "*negdf2_1"
[(set (match_operand:DF 0 "register_operand" "=f")
"TARGET_80387 && reload_completed"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "DF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "DF")])
(define_insn "*negextendsfdf2"
[(set (match_operand:DF 0 "register_operand" "=f")
"TARGET_80387"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "DF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "DF")])
(define_insn "*negxf2_1"
[(set (match_operand:XF 0 "register_operand" "=f")
"TARGET_80387 && reload_completed"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "XF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "XF")])
(define_insn "*negextenddfxf2"
[(set (match_operand:XF 0 "register_operand" "=f")
"TARGET_80387"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "XF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "XF")])
(define_insn "*negextendsfxf2"
[(set (match_operand:XF 0 "register_operand" "=f")
"TARGET_80387"
"fchs"
[(set_attr "type" "fsgn")
- (set_attr "mode" "XF")
- (set_attr "ppro_uops" "few")])
+ (set_attr "mode" "XF")])
\f
;; Absolute value instructions
(set_attr "prefix_0f" "1")
(set_attr "mode" "SI")
(set_attr "pent_pair" "np")
- (set_attr "athlon_decode" "vector")
- (set_attr "ppro_uops" "few")])
+ (set_attr "athlon_decode" "vector")])
(define_expand "x86_shift_adj_1"
[(set (reg:CCZ 17)
[(set_attr "type" "ishift")
(set_attr "prefix_0f" "1")
(set_attr "pent_pair" "np")
- (set_attr "ppro_uops" "few")
(set_attr "mode" "SI")])
(define_expand "x86_shift_adj_3"
else
return "dec{l}\t%1\;%+jne\t%l0";
}
- [(set_attr "ppro_uops" "many")
- (set (attr "length")
+ [(set (attr "length")
(if_then_else (and (eq_attr "alternative" "0")
(and (ge (minus (match_dup 0) (pc))
(const_int -126))
"nop"
[(set_attr "length" "1")
(set_attr "length_immediate" "0")
- (set_attr "modrm" "0")
- (set_attr "ppro_uops" "one")])
+ (set_attr "modrm" "0")])
;; Align to 16-byte boundary, max skip in op0. Used to avoid
;; branch prediction penalty for the third jump in a 16-byte
(ctz:SI (match_dup 1)))]
""
"bsf{l}\t{%1, %0|%0, %1}"
- [(set_attr "prefix_0f" "1")
- (set_attr "ppro_uops" "few")])
+ [(set_attr "prefix_0f" "1")])
(define_insn "ctzsi2"
[(set (match_operand:SI 0 "register_operand" "=r")
(clobber (reg:CC 17))]
""
"bsf{l}\t{%1, %0|%0, %1}"
- [(set_attr "prefix_0f" "1")
- (set_attr "ppro_uops" "few")])
+ [(set_attr "prefix_0f" "1")])
(define_expand "clzsi2"
[(parallel
(clobber (reg:CC 17))]
""
"bsr{l}\t{%1, %0|%0, %1}"
- [(set_attr "prefix_0f" "1")
- (set_attr "ppro_uops" "few")])
+ [(set_attr "prefix_0f" "1")])
\f
;; Thread-local storage patterns for ELF.
;;
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "ppro_uops" "many")
(set_attr "mode" "SI")])
(define_insn "*fop_sf_3"
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "ppro_uops" "many")
(set_attr "mode" "SI")])
(define_insn "*fop_df_1_nosse"
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "ppro_uops" "many")
(set_attr "mode" "SI")])
(define_insn "*fop_df_3"
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "ppro_uops" "many")
(set_attr "mode" "SI")])
(define_insn "*fop_df_4"
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "many")])
+ (set_attr "mode" "SI")])
(define_insn "*fop_xf_3"
[(set (match_operand:XF 0 "register_operand" "=f,f")
]
(const_string "fop")))
(set_attr "fp_int_src" "true")
- (set_attr "mode" "SI")
- (set_attr "ppro_uops" "many")])
+ (set_attr "mode" "SI")])
(define_insn "*fop_xf_4"
[(set (match_operand:XF 0 "register_operand" "=f,f")
-;; Pentium Pro/PII Scheduling
-;; Copyright (C) 2002 Free Software Foundation, Inc.
+;; Scheduling for the Intel P6 family of processors
+;; Copyright (C) 2004 Free Software Foundation, Inc.
;;
;; This file is part of GCC.
;;
;; the Free Software Foundation, 59 Temple Place - Suite 330,
;; Boston, MA 02111-1307, USA. */
-;; Categorize how many uops an ia32 instruction evaluates to:
-;; one -- an instruction with 1 uop can be decoded by any of the
-;; three decoders.
-;; few -- an instruction with 1 to 4 uops can be decoded only by
-;; decoder 0.
-;; many -- a complex instruction may take an unspecified number of
-;; cycles to decode in decoder 0.
-
-(define_attr "ppro_uops" "one,few,many"
- (cond [(eq_attr "type" "other,multi,call,callv,fpspc,str")
- (const_string "many")
- (eq_attr "type" "icmov,fcmov,str,cld,leave")
- (const_string "few")
- (eq_attr "type" "imov")
- (if_then_else (eq_attr "memory" "store,both")
- (const_string "few")
- (const_string "one"))
- (eq_attr "memory" "!none")
- (const_string "few")
- ]
- (const_string "one")))
-
-;;
-;; The PPro has an out-of-order core, but the instruction decoders are
-;; naturally in-order and asymmetric. We get best performance by scheduling
-;; for the decoders, for in doing so we give the oo execution unit the
-;; most choices.
-;;
-;; Rough readiness numbers. Fine tuning happens in i386.c.
-;;
-;; p0 describes port 0.
-;; p01 describes ports 0 and 1 as a pair; alu insns can issue to either.
-;; p2 describes port 2 for loads.
-;; p34 describes ports 3 and 4 for stores.
-;; fpu describes the fpu accessed via port 0.
-;; ??? It is less than clear if there are separate fadd and fmul units
-;; that could operate in parallel.
-;;
-;; ??? fxch isn't handled; not an issue until sched3 after reg-stack is real.
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "ishift,rotate,ishift1,rotate1,lea,ibr,cld"))
- 1 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "imul"))
- 4 1)
-
-;; ??? Does the divider lock out the pipe while it works,
-;; or is there a disconnected unit?
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "idiv"))
- 17 17)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fop,fsgn,fistp"))
- 3 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fcmov"))
- 2 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fcmp"))
- 1 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fmov"))
- 1 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fmul"))
- 5 1)
-
-(define_function_unit "ppro_p0" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fdiv,fpspc"))
- 56 1)
-
-(define_function_unit "ppro_p01" 2 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "!imov,fmov"))
- 1 1)
-
-(define_function_unit "ppro_p01" 2 0
- (and (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "imov,fmov"))
- (eq_attr "memory" "none"))
- 1 1)
-
-(define_function_unit "ppro_p2" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (ior (eq_attr "type" "pop,leave")
- (eq_attr "memory" "load,both")))
- 3 1)
-
-(define_function_unit "ppro_p34" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (ior (eq_attr "type" "push")
- (eq_attr "memory" "store,both")))
- 1 1)
-
-(define_function_unit "fpu" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fop,fsgn,fmov,fcmp,fcmov,fistp"))
- 1 1)
-
-(define_function_unit "fpu" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fmul"))
- 5 2)
-
-(define_function_unit "fpu" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "fdiv,fpspc"))
- 56 56)
-
-;; imul uses the fpu. ??? does it have the same throughput as fmul?
-(define_function_unit "fpu" 1 0
- (and (eq_attr "cpu" "pentiumpro")
- (eq_attr "type" "imul"))
- 4 1)
+;; The P6 familiy includes the Pentium Pro, Pentium II, Pentium III, Celeron
+;; and Xeon lines of CPUs. The DFA scheduler description in this file is
+;; based on information that can be found in the following three documents:
+;;
+;; "P6 Family of Processors Hardware Developer's Manual",
+;; Intel, September 1999.
+;;
+;; "Intel Architecture Optimization Manual",
+;; Intel, 1999 (Order Number: 245127-001).
+;;
+;; "How to optimize for the Pentium family of microprocessors",
+;; by Agner Fog, PhD.
+;;
+;; The P6 pipeline has three major components:
+;; 1) the FETCH/DECODE unit, an in-order issue front-end
+;; 2) the DISPATCH/EXECUTE unit, which is the out-of-order core
+;; 3) the RETIRE unit, an in-order retirement unit
+;;
+;; So, the P6 CPUs have out-of-order cores, but the instruction decoder and
+;; retirement unit are naturally in-order.
+;;
+;; BUS INTERFACE UNIT
+;; / \
+;; L1 ICACHE L1 DCACHE
+;; / | \ | \
+;; DECODER0 DECODER1 DECODER2 DISP/EXEC RETIRE
+;; \ | / | |
+;; INSTRUCTION POOL __________|_______/
+;; (inc. reorder buffer)
+;;
+;; Since the P6 CPUs execute instructions out-of-order, the most important
+;; consideration in performance tuning is making sure enough micro-ops are
+;; ready for execution in the out-of-order core, while not stalling the
+;; decoder.
+;;
+;; TODO:
+;; - Find a less crude way to model complex instructions, in
+;; particular how many cycles they take to be decoded.
+;; - Include decoder latencies in the total reservation latencies.
+;; This isn't necessary right now because we assume for every
+;; instruction that it never blocks a decoder.
+;; - Figure out where the p0 and p1 reservations come from. These
+;; appear not to be in the manual (e.g. why is cld "(p0+p1)*2"
+;; better than "(p0|p1)*4" ???)
+;; - Lots more because I'm sure this is still far from optimal :-)
+
+;; The ppro_idiv and ppro_fdiv automata are used to model issue
+;; latencies of idiv and fdiv type insns.
+(define_automaton "ppro_decoder,ppro_core,ppro_idiv,ppro_fdiv,ppro_load,ppro_store")
+
+;; Simple instructions of the register-register form have only one uop.
+;; Load instructions are also only one uop. Store instructions decode to
+;; two uops, and simple read-modify instructions also take two uops.
+;; Simple instructions of the register-memory form have two to three uops.
+;; Simple read-modify-write instructions have four uops. The rules for
+;; the decoder are simple:
+;; - an instruction with 1 uop can be decoded by any of the three
+;; decoders in one cycle.
+;; - an instruction with 1 to 4 uops can be decoded only by decoder 0
+;; but still in only one cycle.
+;; - a complex (microcode) instruction can also only be decoded by
+;; decoder 0, and this takes an unspecified number of cycles.
+;;
+;; The goal is to schedule such that we have a few-one-one uops sequence
+;; in each cycle, to decode as many instructions per cycle as possible.
+(define_cpu_unit "decoder0" "ppro_decoder")
+(define_cpu_unit "decoder1" "ppro_decoder")
+(define_cpu_unit "decoder2" "ppro_decoder")
+
+;; We first wish to find an instruction for decoder0, so exclude
+;; decoder1 and decoder2 from being reserved until decoder 0 is
+;; reserved.
+(presence_set "decoder1" "decoder0")
+(presence_set "decoder2" "decoder0")
+
+;; Most instructions can be decoded on any of the three decoders.
+(define_reservation "decodern" "(decoder0|decoder1|decoder2)")
+
+;; The out-of-order core has five pipelines. During each cycle, the core
+;; may dispatch zero or one uop on the port of any of the five pipelines
+;; so the maximum number of dispatched uops per cycle is 5. In practicer,
+;; 3 uops per cycle is more realistic.
+;;
+;; Two of the five pipelines contain several execution units:
+;;
+;; Port 0 Port 1 Port 2 Port 3 Port 4
+;; ALU ALU LOAD SAC SDA
+;; FPU JUE
+;; AGU MMX
+;; MMX P3FPU
+;; P3FPU
+;;
+;; (SAC=Store Address Calculation, SDA=Store Data Unit, P3FPU = SSE unit,
+;; JUE = Jump Execution Unit, AGU = Addres Generation Unit)
+;;
+(define_cpu_unit "p0,p1" "ppro_core")
+(define_cpu_unit "p2" "ppro_load")
+(define_cpu_unit "p3,p4" "ppro_store")
+(define_cpu_unit "idiv" "ppro_idiv")
+(define_cpu_unit "fdiv" "ppro_fdiv")
+
+;; Only the irregular instructions have to be modeled here. A load
+;; increases the latency by 2 or 3, or by nothing if the manual gives
+;; a latency already. Store latencies are not accounted for.
+;;
+;; The simple instructions follow a very regular pattern of 1 uop per
+;; reg-reg operation, 1 uop per load on port 2. and 2 uops per store
+;; on port 4 and port 3. These instructions are modelled at the bottom
+;; of this file.
+;;
+;; For microcoded instructions we don't know how many uops are produced.
+;; These instructions are the "complex" ones in the Intel manuals. All
+;; we _do_ know is that they typically produce four or more uops, so
+;; they can only be decoded on decoder0. Modelling their latencies
+;; doesn't make sense because we don't know how these instructions are
+;; executed in the core. So we just model that they can only be decoded
+;; on decoder 0, and say that it takes a little while before the result
+;; is availale.
+(define_insn_reservation "ppro_complex_insn" 6
+ (eq_attr "type" "other,multi,call,callv,str")
+ "decoder0")
+
+;; imov with memory operands does not use the integer units.
+(define_insn_reservation "ppro_imov" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "imov")))
+ "decodern,(p0|p1)")
+
+(define_insn_reservation "ppro_imov_load" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "imov")))
+ "decodern,p2")
+
+(define_insn_reservation "ppro_imov_store" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (eq_attr "type" "imov")))
+ "decoder0,p4+p3")
+
+;; imovx always decodes to one uop, and also doesn't use the integer
+;; units if it has memory operands.
+(define_insn_reservation "ppro_imovx" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "imovx")))
+ "decodern,(p0|p1)")
+
+(define_insn_reservation "ppro_imovx_load" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "imovx")))
+ "decodern,p2")
+
+;; lea executes on port 0 with latency one and throughput 1.
+(define_insn_reservation "ppro_lea" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "lea")))
+ "decodern,p0")
+
+;; Shift and rotate execute on port 0 with latency and throughput 1.
+;; The load and store units need to be reserved when memory operands
+;; are involved.
+(define_insn_reservation "ppro_shift_rotate" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "ishift,ishift1,rotate,rotate1")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_shift_rotate_mem" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "!none")
+ (eq_attr "type" "ishift,ishift1,rotate,rotate1")))
+ "decoder0,p2+p0,p4+p3")
+
+(define_insn_reservation "ppro_cld" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (eq_attr "type" "cld"))
+ "decoder0,(p0+p1)*2")
+
+;; The P6 has a sophisticated branch prediction mechanism to miminize
+;; latencies due to branching. In particular, it has a fast way to
+;; execute branches that are taken multiple times (such as in loops).
+;; Branches not taken suffer no penalty, and correctly predicted
+;; branches cost only one fetch cycle. Mispredicted branches are very
+;; costly: typically 15 cycles and possibly as many as 26 cycles.
+;;
+;; Unfortunatetely all this makes it quite difficult to properly model
+;; the latencies for the compiler. Here I've made the choice to be
+;; optimistic and assume branches are often predicted correctly, so
+;; they have latency 1, and the decoders are not blocked.
+;;
+;; In addition, the model assumes a branch always decodes to only 1 uop,
+;; which is not exactly true because there are a few instructions that
+;; decode to 2 uops or microcode. But this probably gives the best
+;; results because we can assume these instructions can decode on all
+;; decoders.
+(define_insn_reservation "ppro_branch" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "ibr")))
+ "decodern,p1")
+
+;; ??? Indirect branches probably have worse latency than this.
+(define_insn_reservation "ppro_indirect_branch" 6
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "!none")
+ (eq_attr "type" "ibr")))
+ "decoder0,p2+p1")
+
+(define_insn_reservation "ppro_leave" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (eq_attr "type" "leave"))
+ "decoder0,p2+(p0|p1),(p0|p1)")
+
+;; imul has throughput one, but latency 4, and can only execute on port 0.
+(define_insn_reservation "ppro_imul" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "imul")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_imul_mem" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "!none")
+ (eq_attr "type" "imul")))
+ "decoder0,p2+p0")
+
+;; div and idiv are very similar, so we model them the same.
+;; QI, HI, and SI have issue latency 12, 21, and 37, respectively.
+;; These issue latencies are modelled via the ppro_div automaton.
+(define_insn_reservation "ppro_idiv_QI" 19
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "QI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,(p0+idiv)*2,(p0|p1)+idiv,idiv*9")
+
+(define_insn_reservation "ppro_idiv_QI_load" 19
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "QI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,p2+p0+idiv,p0+idiv,(p0|p1)+idiv,idiv*9")
+
+(define_insn_reservation "ppro_idiv_HI" 23
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "HI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,(p0+idiv)*3,(p0|p1)+idiv,idiv*17")
+
+(define_insn_reservation "ppro_idiv_HI_load" 23
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "HI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,p2+p0+idiv,p0+idiv,(p0|p1)+idiv,idiv*18")
+
+(define_insn_reservation "ppro_idiv_SI" 39
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,(p0+idiv)*3,(p0|p1)+idiv,idiv*33")
+
+(define_insn_reservation "ppro_idiv_SI_load" 39
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SI")
+ (eq_attr "type" "idiv"))))
+ "decoder0,p2+p0+idiv,p0+idiv,(p0|p1)+idiv,idiv*34")
+
+;; Floating point operations always execute on port 0.
+;; ??? where do these latencies come from? fadd has latency 3 and
+;; has throughput "1/cycle (align with FADD)". What do they
+;; mean and how can we model that?
+(define_insn_reservation "ppro_fop" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none,unknown")
+ (eq_attr "type" "fop")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_fop_load" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "fop")))
+ "decoder0,p2+p0,p0")
+
+(define_insn_reservation "ppro_fop_store" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (eq_attr "type" "fop")))
+ "decoder0,p0,p0,p0+p4+p3")
+
+(define_insn_reservation "ppro_fop_both" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "both")
+ (eq_attr "type" "fop")))
+ "decoder0,p2+p0,p0+p4+p3")
+
+(define_insn_reservation "ppro_fsgn" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (eq_attr "type" "fsgn"))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_fistp" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (eq_attr "type" "fistp"))
+ "decoder0,p0*2,p4+p3")
+
+(define_insn_reservation "ppro_fcmov" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (eq_attr "type" "fcmov"))
+ "decoder0,p0*2")
+
+(define_insn_reservation "ppro_fcmp" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "fcmp")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_fcmp_load" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "fcmp")))
+ "decoder0,p2+p0")
+
+(define_insn_reservation "ppro_fmov" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "fmov")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_fmov_load" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "!XF")
+ (eq_attr "type" "fmov"))))
+ "decodern,p2")
+
+(define_insn_reservation "ppro_fmov_XF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "XF")
+ (eq_attr "type" "fmov"))))
+ "decoder0,(p2+p0)*2")
+
+(define_insn_reservation "ppro_fmov_store" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (and (eq_attr "mode" "!XF")
+ (eq_attr "type" "fmov"))))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_fmov_XF_store" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (and (eq_attr "mode" "XF")
+ (eq_attr "type" "fmov"))))
+ "decoder0,(p0+p4),(p0+p3)")
+
+;; fmul executes on port 0 with latency 5. It has issue latency 2,
+;; but we don't model this.
+(define_insn_reservation "ppro_fmul" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "fmul")))
+ "decoder0,p0*2")
+
+(define_insn_reservation "ppro_fmul_load" 6
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "fmul")))
+ "decoder0,p2+p0,p0")
+
+;; fdiv latencies depend on the mode of the operands. XFmode gives
+;; a latency of 38 cycles, DFmode gives 32, and SFmode gives latency 18.
+;; Division by a power of 2 takes only 9 cycles, but we cannot model
+;; that. Throughput is equal to latency - 1, which we model using the
+;; ppro_div automaton.
+(define_insn_reservation "ppro_fdiv_SF" 18
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decodern,p0+fdiv,fdiv*16")
+
+(define_insn_reservation "ppro_fdiv_SF_load" 19
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decoder0,p2+p0+fdiv,fdiv*16")
+
+(define_insn_reservation "ppro_fdiv_DF" 32
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "DF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decodern,p0+fdiv,fdiv*30")
+
+(define_insn_reservation "ppro_fdiv_DF_load" 33
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "DF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decoder0,p2+p0+fdiv,fdiv*30")
+
+(define_insn_reservation "ppro_fdiv_XF" 38
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "XF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decodern,p0+fdiv,fdiv*36")
+
+(define_insn_reservation "ppro_fdiv_XF_load" 39
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "XF")
+ (eq_attr "type" "fdiv,fpspc"))))
+ "decoder0,p2+p0+fdiv,fdiv*36")
+
+;; MMX instructions can execute on either port 0 or port 1 with a
+;; throughput of 1/cycle.
+;; on port 0: - ALU (latency 1)
+;; - Multiplier Unit (latency 3)
+;; on port 1: - ALU (latency 1)
+;; - Shift Unit (latency 1)
+;;
+;; MMX instructions are either of the type reg-reg, or read-modify, and
+;; except for mmxshft and mmxmul they can execute on port 0 or port 1,
+;; so they behave as "simple" instructions that need no special modelling.
+;; We only have to model mmxshft and mmxmul.
+(define_insn_reservation "ppro_mmx_shft" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "mmxshft")))
+ "decodern,p1")
+
+(define_insn_reservation "ppro_mmx_shft_load" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "mmxshft")))
+ "decoder0,p2+p1")
+
+(define_insn_reservation "ppro_mmx_mul" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "mmxmul")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_mmx_mul_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (eq_attr "type" "mmxmul")))
+ "decoder0,p2+p0")
+
+(define_insn_reservation "ppro_sse_mmxcvt" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "mode" "DI")
+ (eq_attr "type" "mmxcvt")))
+ "decodern,p1")
+
+;; FIXME: These are Pentium III only, but we cannot tell here if
+;; we're generating code for PentiumPro/Pentium II or Pentium III
+;; (define_insn_reservation "ppro_sse_mmxshft" 2
+;; (and (eq_attr "cpu" "pentiumpro")
+;; (and (eq_attr "mode" "DI")
+;; (eq_attr "type" "mmxshft")))
+;; "decodern,p0")
+
+;; SSE is very complicated, and takes a bit more effort.
+;; ??? I assumed that all SSE instructions decode on decoder0,
+;; but is this correct?
+
+;; The sfence instruction.
+(define_insn_reservation "ppro_sse_sfence" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "unknown")
+ (eq_attr "type" "sse")))
+ "decoder0,p4+p3")
+
+;; FIXME: This reservation is all wrong when we're scheduling sqrtss.
+(define_insn_reservation "ppro_sse_SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "sse")))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_sse_add_SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "sseadd"))))
+ "decodern,p1")
+
+(define_insn_reservation "ppro_sse_add_SF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "sseadd"))))
+ "decoder0,p2+p1")
+
+(define_insn_reservation "ppro_sse_cmp_SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssecmp"))))
+ "decoder0,p1")
+
+(define_insn_reservation "ppro_sse_cmp_SF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssecmp"))))
+ "decoder0,p2+p1")
+
+(define_insn_reservation "ppro_sse_comi_SF" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssecomi"))))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_sse_comi_SF_load" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssecomi"))))
+ "decoder0,p2+p0")
+
+(define_insn_reservation "ppro_sse_mul_SF" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssemul"))))
+ "decodern,p0")
+
+(define_insn_reservation "ppro_sse_mul_SF_load" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssemul"))))
+ "decoder0,p2+p0")
+
+;; FIXME: ssediv doesn't close p0 for 17 cycles, surely???
+(define_insn_reservation "ppro_sse_div_SF" 18
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssediv"))))
+ "decoder0,p0*17")
+
+(define_insn_reservation "ppro_sse_div_SF_load" 18
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssediv"))))
+ "decoder0,(p2+p0),p0*16")
+
+(define_insn_reservation "ppro_sse_icvt_SF" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "sseicvt")))
+ "decoder0,(p2+p1)*2")
+
+(define_insn_reservation "ppro_sse_icvt_SI" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "mode" "SI")
+ (eq_attr "type" "sseicvt")))
+ "decoder0,(p2+p1)")
+
+(define_insn_reservation "ppro_sse_mov_SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,(p0|p1)")
+
+(define_insn_reservation "ppro_sse_mov_SF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,p2+(p0|p1)")
+
+(define_insn_reservation "ppro_sse_mov_SF_store" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (and (eq_attr "mode" "SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,p4+p3")
+
+(define_insn_reservation "ppro_sse_V4SF" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "sse")))
+ "decoder0,p1*2")
+
+(define_insn_reservation "ppro_sse_add_V4SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "sseadd"))))
+ "decoder0,p1*2")
+
+(define_insn_reservation "ppro_sse_add_V4SF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "sseadd"))))
+ "decoder0,(p2+p1)*2")
+
+(define_insn_reservation "ppro_sse_cmp_V4SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssecmp"))))
+ "decoder0,p1*2")
+
+(define_insn_reservation "ppro_sse_cmp_V4SF_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssecmp"))))
+ "decoder0,(p2+p1)*2")
+
+(define_insn_reservation "ppro_sse_cvt_V4SF" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none,unknown")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssecvt"))))
+ "decoder0,p1*2")
+
+(define_insn_reservation "ppro_sse_cvt_V4SF_other" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "!none,unknown")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssecmp"))))
+ "decoder0,p1,p4+p3")
+
+(define_insn_reservation "ppro_sse_mul_V4SF" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssemul"))))
+ "decoder0,p0*2")
+
+(define_insn_reservation "ppro_sse_mul_V4SF_load" 5
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssemul"))))
+ "decoder0,(p2+p0)*2")
+
+;; FIXME: p0 really closed this long???
+(define_insn_reservation "ppro_sse_div_V4SF" 48
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssediv"))))
+ "decoder0,p0*34")
+
+(define_insn_reservation "ppro_sse_div_V4SF_load" 48
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssediv"))))
+ "decoder0,(p2+p0)*2,p0*32")
+
+(define_insn_reservation "ppro_sse_log_V4SF" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "sselog"))))
+ "decodern,p1")
+
+(define_insn_reservation "ppro_sse_log_V4SF_load" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "sselog"))))
+ "decoder0,(p2+p1)")
+
+(define_insn_reservation "ppro_sse_mov_V4SF" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,(p0|p1)*2")
+
+(define_insn_reservation "ppro_sse_mov_V4SF_load" 2
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,p2*2")
+
+(define_insn_reservation "ppro_sse_mov_V4SF_store" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (and (eq_attr "mode" "V4SF")
+ (eq_attr "type" "ssemov"))))
+ "decoder0,(p4+p3)*2")
+
+;; All other instructions are modelled as simple instructions.
+;; We have already modelled all i387 floating point instructions, so all
+;; other instructions execute on either port 0 or port 1. This includes
+;; the ALU units, and the MMX units.
+;;
+;; reg-reg instructions produce 1 uop so they can be decoded on any of
+;; the three decoders.
+(define_insn_reservation "ppro_insn" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "none,unknown")
+ (eq_attr "type" "alu,alu1,negnot,incdec,icmp,test,setcc,icmov,push,pop,fxch,sseiadd,sseishft,sseimul,mmx,mmxadd,mmxcmp")))
+ "decodern,(p0|p1)")
+
+;; read-modify and register-memory instructions have 2 or three uops,
+;; so they have to be decoded on decoder0.
+(define_insn_reservation "ppro_insn_load" 3
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "load")
+ (eq_attr "type" "alu,alu1,negnot,incdec,icmp,test,setcc,icmov,push,pop,fxch,sseiadd,sseishft,sseimul,mmx,mmxadd,mmxcmp")))
+ "decoder0,p2+(p0|p1)")
+
+(define_insn_reservation "ppro_insn_store" 1
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "store")
+ (eq_attr "type" "alu,alu1,negnot,incdec,icmp,test,setcc,icmov,push,pop,fxch,sseiadd,sseishft,sseimul,mmx,mmxadd,mmxcmp")))
+ "decoder0,(p0|p1),p4+p3")
+
+;; read-modify-store instructions produce 4 uops so they have to be
+;; decoded on decoder0 as well.
+(define_insn_reservation "ppro_insn_both" 4
+ (and (eq_attr "cpu" "pentiumpro")
+ (and (eq_attr "memory" "both")
+ (eq_attr "type" "alu,alu1,negnot,incdec,icmp,test,setcc,icmov,push,pop,fxch,sseiadd,sseishft,sseimul,mmx,mmxadd,mmxcmp")))
+ "decoder0,p2+(p0|p1),p4+p3")
+