gcc_assert (REGNO (x) == REG_A0 || REGNO (x) == REG_A1);
fprintf (file, "%s.x", reg_names[REGNO (x)]);
}
+ else if (code == 'v')
+ {
+ if (REGNO (x) == REG_A0)
+ fprintf (file, "AV0");
+ else if (REGNO (x) == REG_A1)
+ fprintf (file, "AV1");
+ else
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ }
else if (code == 'D')
{
fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
BFIN_BUILTIN_MIN_1X16,
BFIN_BUILTIN_MAX_1X16,
+ BFIN_BUILTIN_SUM_2X16,
BFIN_BUILTIN_DIFFHL_2X16,
BFIN_BUILTIN_DIFFLH_2X16,
BFIN_BUILTIN_SSADD_1X32,
BFIN_BUILTIN_SSSUB_1X32,
BFIN_BUILTIN_NORM_1X32,
+ BFIN_BUILTIN_ROUND_1X32,
BFIN_BUILTIN_NEG_1X32,
+ BFIN_BUILTIN_ABS_1X32,
BFIN_BUILTIN_MIN_1X32,
BFIN_BUILTIN_MAX_1X32,
BFIN_BUILTIN_MULT_1X32,
+ BFIN_BUILTIN_MULT_1X32X32,
+ BFIN_BUILTIN_MULT_1X32X32NS,
BFIN_BUILTIN_MULHISILL,
BFIN_BUILTIN_MULHISILH,
BFIN_BUILTIN_LSHIFT_2X16,
BFIN_BUILTIN_SSASHIFT_1X16,
BFIN_BUILTIN_SSASHIFT_2X16,
+ BFIN_BUILTIN_SSASHIFT_1X32,
BFIN_BUILTIN_CPLX_MUL_16,
BFIN_BUILTIN_CPLX_MAC_16,
def_builtin ("__builtin_bfin_norm_fr1x16", short_ftype_int,
BFIN_BUILTIN_NORM_1X16);
+ def_builtin ("__builtin_bfin_sum_fr2x16", short_ftype_v2hi,
+ BFIN_BUILTIN_SUM_2X16);
def_builtin ("__builtin_bfin_diff_hl_fr2x16", short_ftype_v2hi,
BFIN_BUILTIN_DIFFHL_2X16);
def_builtin ("__builtin_bfin_diff_lh_fr2x16", short_ftype_v2hi,
BFIN_BUILTIN_SSSUB_1X32);
def_builtin ("__builtin_bfin_negate_fr1x32", int_ftype_int,
BFIN_BUILTIN_NEG_1X32);
+ def_builtin ("__builtin_bfin_abs_fr1x32", int_ftype_int,
+ BFIN_BUILTIN_ABS_1X32);
def_builtin ("__builtin_bfin_norm_fr1x32", short_ftype_int,
BFIN_BUILTIN_NORM_1X32);
+ def_builtin ("__builtin_bfin_round_fr1x32", short_ftype_int,
+ BFIN_BUILTIN_ROUND_1X32);
def_builtin ("__builtin_bfin_mult_fr1x32", int_ftype_short_short,
BFIN_BUILTIN_MULT_1X32);
+ def_builtin ("__builtin_bfin_mult_fr1x32x32", int_ftype_int_int,
+ BFIN_BUILTIN_MULT_1X32X32);
+ def_builtin ("__builtin_bfin_mult_fr1x32x32NS", int_ftype_int_int,
+ BFIN_BUILTIN_MULT_1X32X32NS);
/* Shifts. */
def_builtin ("__builtin_bfin_shl_fr1x16", short_ftype_int_int,
BFIN_BUILTIN_LSHIFT_1X16);
def_builtin ("__builtin_bfin_lshl_fr2x16", v2hi_ftype_v2hi_int,
BFIN_BUILTIN_LSHIFT_2X16);
+ def_builtin ("__builtin_bfin_shl_fr1x32", int_ftype_int_int,
+ BFIN_BUILTIN_SSASHIFT_1X32);
/* Complex numbers. */
def_builtin ("__builtin_bfin_cmplx_mul", v2hi_ftype_v2hi_v2hi,
{ CODE_FOR_ssashifthi3, "__builtin_bfin_shl_fr1x16", BFIN_BUILTIN_SSASHIFT_1X16, -1 },
{ CODE_FOR_lshiftv2hi3, "__builtin_bfin_lshl_fr2x16", BFIN_BUILTIN_LSHIFT_2X16, -1 },
{ CODE_FOR_lshifthi3, "__builtin_bfin_lshl_fr1x16", BFIN_BUILTIN_LSHIFT_1X16, -1 },
+ { CODE_FOR_ssashiftsi3, "__builtin_bfin_shl_fr1x32", BFIN_BUILTIN_SSASHIFT_1X32, -1 },
{ CODE_FOR_sminhi3, "__builtin_bfin_min_fr1x16", BFIN_BUILTIN_MIN_1X16, -1 },
{ CODE_FOR_smaxhi3, "__builtin_bfin_max_fr1x16", BFIN_BUILTIN_MAX_1X16, -1 },
{ CODE_FOR_abshi2, "__builtin_bfin_abs_fr1x16", BFIN_BUILTIN_ABS_1X16, 0 },
{ CODE_FOR_signbitssi2, "__builtin_bfin_norm_fr1x32", BFIN_BUILTIN_NORM_1X32, 0 },
+ { CODE_FOR_ssroundsi2, "__builtin_bfin_round_fr1x32", BFIN_BUILTIN_ROUND_1X32, 0 },
{ CODE_FOR_ssnegsi2, "__builtin_bfin_negate_fr1x32", BFIN_BUILTIN_NEG_1X32, 0 },
+ { CODE_FOR_ssabssi2, "__builtin_bfin_abs_fr1x32", BFIN_BUILTIN_ABS_1X32, 0 },
{ CODE_FOR_movv2hi_hi_low, "__builtin_bfin_extract_lo", BFIN_BUILTIN_EXTRACTLO, 0 },
{ CODE_FOR_movv2hi_hi_high, "__builtin_bfin_extract_hi", BFIN_BUILTIN_EXTRACTHI, 0 },
{ CODE_FOR_ssnegv2hi2, "__builtin_bfin_negate_fr2x16", BFIN_BUILTIN_NEG_2X16, 0 },
- { CODE_FOR_absv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
+ { CODE_FOR_ssabsv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
};
/* Errors in the source file can cause expand_expr to return const0_rtx
tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
tree arg0, arg1, arg2;
- rtx op0, op1, op2, accvec, pat, tmp1, tmp2;
+ rtx op0, op1, op2, accvec, pat, tmp1, tmp2, a0reg, a1reg;
enum machine_mode tmode, mode0;
switch (fcode)
case BFIN_BUILTIN_DIFFHL_2X16:
case BFIN_BUILTIN_DIFFLH_2X16:
+ case BFIN_BUILTIN_SUM_2X16:
arg0 = CALL_EXPR_ARG (exp, 0);
op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
- icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16
- ? CODE_FOR_subhilov2hi3 : CODE_FOR_sublohiv2hi3);
+ icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16 ? CODE_FOR_subhilov2hi3
+ : fcode == BFIN_BUILTIN_DIFFLH_2X16 ? CODE_FOR_sublohiv2hi3
+ : CODE_FOR_ssaddhilov2hi3);
tmode = insn_data[icode].operand[0].mode;
mode0 = insn_data[icode].operand[1].mode;
emit_insn (pat);
return target;
+ case BFIN_BUILTIN_MULT_1X32X32:
+ case BFIN_BUILTIN_MULT_1X32X32NS:
+ arg0 = CALL_EXPR_ARG (exp, 0);
+ arg1 = CALL_EXPR_ARG (exp, 1);
+ op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
+ op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
+ if (! target
+ || !register_operand (target, SImode))
+ target = gen_reg_rtx (SImode);
+
+ a1reg = gen_rtx_REG (PDImode, REG_A1);
+ a0reg = gen_rtx_REG (PDImode, REG_A0);
+ tmp1 = gen_lowpart (V2HImode, op0);
+ tmp2 = gen_lowpart (V2HImode, op1);
+ emit_insn (gen_flag_macinit1hi (a1reg,
+ gen_lowpart (HImode, op0),
+ gen_lowpart (HImode, op1),
+ GEN_INT (MACFLAG_FU)));
+ emit_insn (gen_lshrpdi3 (a1reg, a1reg, GEN_INT (16)));
+
+ if (fcode == BFIN_BUILTIN_MULT_1X32X32)
+ emit_insn (gen_flag_mul_macv2hi_parts_acconly (a0reg, a1reg, tmp1, tmp2,
+ const1_rtx, const1_rtx,
+ const1_rtx, const0_rtx, a1reg,
+ const0_rtx, GEN_INT (MACFLAG_NONE),
+ GEN_INT (MACFLAG_M)));
+ else
+ {
+ /* For saturating multiplication, there's exactly one special case
+ to be handled: multiplying the smallest negative value with
+ itself. Due to shift correction in fractional multiplies, this
+ can overflow. Iff this happens, OP2 will contain 1, which, when
+ added in 32 bits to the smallest negative, wraps to the largest
+ positive, which is the result we want. */
+ op2 = gen_reg_rtx (V2HImode);
+ emit_insn (gen_packv2hi (op2, tmp1, tmp2, const0_rtx, const0_rtx));
+ emit_insn (gen_movsibi (gen_rtx_REG (BImode, REG_CC),
+ gen_lowpart (SImode, op2)));
+ emit_insn (gen_flag_mul_macv2hi_parts_acconly_andcc0 (a0reg, a1reg, tmp1, tmp2,
+ const1_rtx, const1_rtx,
+ const1_rtx, const0_rtx, a1reg,
+ const0_rtx, GEN_INT (MACFLAG_NONE),
+ GEN_INT (MACFLAG_M)));
+ op2 = gen_reg_rtx (SImode);
+ emit_insn (gen_movbisi (op2, gen_rtx_REG (BImode, REG_CC)));
+ }
+ emit_insn (gen_flag_machi_parts_acconly (a1reg, tmp2, tmp1,
+ const1_rtx, const0_rtx,
+ a1reg, const0_rtx, GEN_INT (MACFLAG_M)));
+ emit_insn (gen_ashrpdi3 (a1reg, a1reg, GEN_INT (15)));
+ emit_insn (gen_sum_of_accumulators (target, a0reg, a0reg, a1reg));
+ if (fcode == BFIN_BUILTIN_MULT_1X32X32NS)
+ emit_insn (gen_addsi3 (target, target, op2));
+ return target;
+
case BFIN_BUILTIN_CPLX_MUL_16:
arg0 = CALL_EXPR_ARG (exp, 0);
arg1 = CALL_EXPR_ARG (exp, 1);
"%0 = abs %1%!"
[(set_attr "type" "dsp32")])
+(define_insn "ssabssi2"
+ [(set (match_operand:SI 0 "register_operand" "=d")
+ (ss_abs:SI (match_operand:SI 1 "register_operand" "d")))]
+ ""
+ "%0 = abs %1%!"
+ [(set_attr "type" "dsp32")])
+
(define_insn "negsi2"
[(set (match_operand:SI 0 "register_operand" "=d")
(neg:SI (match_operand:SI 1 "register_operand" "d")))]
"%h0 = signbits %1%!"
[(set_attr "type" "dsp32")])
+(define_insn "ssroundsi2"
+ [(set (match_operand:HI 0 "register_operand" "=d")
+ (truncate:HI
+ (lshiftrt:SI (ss_plus:SI (match_operand:SI 1 "register_operand" "d")
+ (const_int 32768))
+ (const_int 16))))]
+ ""
+ "%h0 = %1 (RND)%!"
+ [(set_attr "type" "dsp32")])
+
(define_insn "smaxhi3"
[(set (match_operand:HI 0 "register_operand" "=d")
(smax:HI (match_operand:HI 1 "register_operand" "d")
operands[3] = gen_rtx_NE (BImode, operands[1], const0_rtx);
})
+;; Same as above, but and CC with the overflow bit generated by the first
+;; multiplication.
+(define_insn "flag_mul_macv2hi_parts_acconly_andcc0"
+ [(set (match_operand:PDI 0 "register_operand" "=B,e,e")
+ (unspec:PDI [(vec_select:HI
+ (match_operand:V2HI 2 "register_operand" "d,d,d")
+ (parallel [(match_operand 4 "const01_operand" "P0P1,P0P1,P0P1")]))
+ (vec_select:HI
+ (match_operand:V2HI 3 "register_operand" "d,d,d")
+ (parallel [(match_operand 6 "const01_operand" "P0P1,P0P1,P0P1")]))
+ (match_operand 10 "const_int_operand" "PB,PA,PA")]
+ UNSPEC_MUL_WITH_FLAG))
+ (set (match_operand:PDI 1 "register_operand" "=B,e,e")
+ (unspec:PDI [(vec_select:HI
+ (match_dup 2)
+ (parallel [(match_operand 5 "const01_operand" "P0P1,P0P1,P0P1")]))
+ (vec_select:HI
+ (match_dup 3)
+ (parallel [(match_operand 7 "const01_operand" "P0P1,P0P1,P0P1")]))
+ (match_operand:PDI 8 "register_operand" "1,1,1")
+ (match_operand 9 "const01_operand" "P0P1,P0P1,P0P1")
+ (match_operand 11 "const_int_operand" "PA,PB,PA")]
+ UNSPEC_MAC_WITH_FLAG))
+ (set (reg:BI REG_CC)
+ (and:BI (reg:BI REG_CC)
+ (unspec:BI [(vec_select:HI (match_dup 2) (parallel [(match_dup 4)]))
+ (vec_select:HI (match_dup 3) (parallel [(match_dup 6)]))
+ (match_dup 10)]
+ UNSPEC_MUL_WITH_FLAG)))]
+ "MACFLAGS_MATCH_P (INTVAL (operands[10]), INTVAL (operands[11]))"
+{
+ rtx xops[6];
+ const char *templates[] = {
+ "%0 = %h2 * %h3, %1 %b4 %h2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %h3, %1 %b4 %h2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %h3, %1 %b4 %d2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %h3, %1 %b4 %d2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %d3, %1 %b4 %h2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %d3, %1 %b4 %h2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %d3, %1 %b4 %d2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %d3, %1 %b4 %d2 * %h3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %h3, %1 %b4 %h2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %h3, %1 %b4 %h2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %h3, %1 %b4 %d2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %h3, %1 %b4 %d2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %d3, %1 %b4 %h2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %d3, %1 %b4 %h2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %h2 * %d3, %1 %b4 %d2 * %d3 %M5;\n\tCC &= %v0;",
+ "%0 = %d2 * %d3, %1 %b4 %d2 * %d3 %M5;\n\tCC &= %v0;" };
+ int alt = (INTVAL (operands[4]) + (INTVAL (operands[5]) << 1)
+ + (INTVAL (operands[6]) << 2) + (INTVAL (operands[7]) << 3));
+ xops[0] = operands[0];
+ xops[1] = operands[1];
+ xops[2] = operands[2];
+ xops[3] = operands[3];
+ xops[4] = operands[9];
+ xops[5] = which_alternative == 0 ? operands[10] : operands[11];
+ output_asm_insn (templates[alt], xops);
+ return "";
+}
+ [(set_attr "type" "misc")
+ (set_attr "length" "6")
+ (set_attr "seq_insns" "multi")])
(define_expand "bge"
[(set (match_dup 1) (match_dup 2))
"%0 = - %1 (V)%!"
[(set_attr "type" "dsp32")])
-(define_insn "absv2hi2"
+(define_insn "ssabsv2hi2"
[(set (match_operand:V2HI 0 "register_operand" "=d")
- (abs:V2HI (match_operand:V2HI 1 "register_operand" "d")))]
+ (ss_abs:V2HI (match_operand:V2HI 1 "register_operand" "d")))]
""
"%0 = ABS %1 (V)%!"
[(set_attr "type" "dsp32")])
(define_insn "ssashiftv2hi3"
[(set (match_operand:V2HI 0 "register_operand" "=d,d,d")
(if_then_else:V2HI
- (lt (match_operand:SI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
+ (lt (match_operand:HI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
(ashiftrt:V2HI (match_operand:V2HI 1 "register_operand" "d,d,d")
(match_dup 2))
(ss_ashift:V2HI (match_dup 1) (match_dup 2))))]
(define_insn "ssashifthi3"
[(set (match_operand:HI 0 "register_operand" "=d,d,d")
(if_then_else:HI
- (lt (match_operand:SI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
+ (lt (match_operand:HI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
(ashiftrt:HI (match_operand:HI 1 "register_operand" "d,d,d")
(match_dup 2))
(ss_ashift:HI (match_dup 1) (match_dup 2))))]
%0 = %1 >>> %N2 (V,S)%!"
[(set_attr "type" "dsp32")])
+(define_insn "ssashiftsi3"
+ [(set (match_operand:SI 0 "register_operand" "=d,d,d")
+ (if_then_else:SI
+ (lt (match_operand:HI 2 "reg_or_const_int_operand" "d,Ku5,Ks5") (const_int 0))
+ (ashiftrt:SI (match_operand:HI 1 "register_operand" "d,d,d")
+ (match_dup 2))
+ (ss_ashift:SI (match_dup 1) (match_dup 2))))]
+ ""
+ "@
+ %0 = ASHIFT %1 BY %h2 (S)%!
+ %0 = %1 << %2 (S)%!
+ %0 = %1 >>> %N2 (S)%!"
+ [(set_attr "type" "dsp32")])
+
(define_insn "lshiftv2hi3"
[(set (match_operand:V2HI 0 "register_operand" "=d,d,d")
(if_then_else:V2HI
- (lt (match_operand:SI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
+ (lt (match_operand:HI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
(lshiftrt:V2HI (match_operand:V2HI 1 "register_operand" "d,d,d")
(match_dup 2))
(ashift:V2HI (match_dup 1) (match_dup 2))))]
(define_insn "lshifthi3"
[(set (match_operand:HI 0 "register_operand" "=d,d,d")
(if_then_else:HI
- (lt (match_operand:SI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
+ (lt (match_operand:HI 2 "vec_shift_operand" "d,Ku4,Ks4") (const_int 0))
(lshiftrt:HI (match_operand:HI 1 "register_operand" "d,d,d")
(match_dup 2))
(ashift:HI (match_dup 1) (match_dup 2))))]
projects / gcc.git / commitdiff