FPU: Decide on A input selection a cycle earlier

This moves opsel_a into the reg_type record, meaning that the A
multiplexer input now needs to be decided a cycle earlier.  This helps
timing by eliminating the combinatorial path from r.state and other
things to opsel_a and thence to in_a and result.

This means that some things now take an extra cycle, in particular
some of the exception cases such as when one or both operands are
NaNs.  The NaN handling has been moved out to its own state, which
simplifies the logic for exception cases in other places.  Additions
or subtractions where FRB's exponent is smaller than FRA's will
also take an extra cycle.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
jtag-port
Paul Mackerras 4 years ago
parent b0b3c0dc70
commit fb5115c944

@ -44,7 +44,7 @@ architecture behaviour of fpu is
DO_FRE, DO_FRSQRTE, DO_FRE, DO_FRSQRTE,
DO_FSEL, DO_FSEL,
FRI_1, FRI_1,
ADD_SHIFT, ADD_2, ADD_3, ADD_1, ADD_SHIFT, ADD_2, ADD_3,
CMP_1, CMP_2, CMP_1, CMP_2,
MULT_1, MULT_1,
FMADD_1, FMADD_2, FMADD_3, FMADD_1, FMADD_2, FMADD_3,
@ -65,7 +65,8 @@ architecture behaviour of fpu is
DENORM, DENORM,
RENORM_A, RENORM_A2, RENORM_A, RENORM_A2,
RENORM_B, RENORM_B2, RENORM_B, RENORM_B2,
RENORM_C, RENORM_C2); RENORM_C, RENORM_C2,
NAN_RESULT, EXC_RESULT);


type reg_type is record type reg_type is record
state : state_t; state : state_t;
@ -111,6 +112,12 @@ architecture behaviour of fpu is
first : std_ulogic; first : std_ulogic;
count : unsigned(1 downto 0); count : unsigned(1 downto 0);
doing_ftdiv : std_ulogic_vector(1 downto 0); doing_ftdiv : std_ulogic_vector(1 downto 0);
opsel_a : std_ulogic_vector(1 downto 0);
use_a : std_ulogic;
use_b : std_ulogic;
use_c : std_ulogic;
invalid : std_ulogic;
negate : std_ulogic;
end record; end record;


type lookup_table is array(0 to 1023) of std_ulogic_vector(17 downto 0); type lookup_table is array(0 to 1023) of std_ulogic_vector(17 downto 0);
@ -118,7 +125,6 @@ architecture behaviour of fpu is
signal r, rin : reg_type; signal r, rin : reg_type;


signal fp_result : std_ulogic_vector(63 downto 0); signal fp_result : std_ulogic_vector(63 downto 0);
signal opsel_a : std_ulogic_vector(1 downto 0);
signal opsel_b : std_ulogic_vector(1 downto 0); signal opsel_b : std_ulogic_vector(1 downto 0);
signal opsel_r : std_ulogic_vector(1 downto 0); signal opsel_r : std_ulogic_vector(1 downto 0);
signal opsel_s : std_ulogic_vector(1 downto 0); signal opsel_s : std_ulogic_vector(1 downto 0);
@ -724,7 +730,7 @@ begin
v.update_fprf := '0'; v.update_fprf := '0';
v.shift := to_signed(0, EXP_BITS); v.shift := to_signed(0, EXP_BITS);
v.first := '0'; v.first := '0';
opsel_a <= AIN_R; v.opsel_a := AIN_R;
opsel_ainv <= '0'; opsel_ainv <= '0';
opsel_amask <= '0'; opsel_amask <= '0';
opsel_b <= BIN_ZERO; opsel_b <= BIN_ZERO;
@ -758,6 +764,11 @@ begin
shiftin := '0'; shiftin := '0';
case r.state is case r.state is
when IDLE => when IDLE =>
v.use_a := '0';
v.use_b := '0';
v.use_c := '0';
v.invalid := '0';
v.negate := '0';
if e_in.valid = '1' then if e_in.valid = '1' then
case e_in.insn(5 downto 1) is case e_in.insn(5 downto 1) is
when "00000" => when "00000" =>
@ -770,6 +781,7 @@ begin
elsif e_in.insn(7) = '1' then elsif e_in.insn(7) = '1' then
v.state := DO_MCRFS; v.state := DO_MCRFS;
else else
v.opsel_a := AIN_B;
v.state := DO_FCMP; v.state := DO_FCMP;
end if; end if;
when "00110" => when "00110" =>
@ -789,14 +801,17 @@ begin
v.state := DO_MTFSF; v.state := DO_MTFSF;
end if; end if;
when "01000" => when "01000" =>
v.opsel_a := AIN_B;
if e_in.insn(9 downto 8) /= "11" then if e_in.insn(9 downto 8) /= "11" then
v.state := DO_FMR; v.state := DO_FMR;
else else
v.state := DO_FRI; v.state := DO_FRI;
end if; end if;
when "01100" => when "01100" =>
v.opsel_a := AIN_B;
v.state := DO_FRSP; v.state := DO_FRSP;
when "01110" => when "01110" =>
v.opsel_a := AIN_B;
if int_input = '1' then if int_input = '1' then
-- fcfid[u][s] -- fcfid[u][s]
v.state := DO_FCFID; v.state := DO_FCFID;
@ -805,25 +820,45 @@ begin
end if; end if;
when "01111" => when "01111" =>
v.round_mode := "001"; v.round_mode := "001";
v.opsel_a := AIN_B;
v.state := DO_FCTI; v.state := DO_FCTI;
when "10010" => when "10010" =>
v.opsel_a := AIN_A;
if v.b.mantissa(54) = '0' and v.a.mantissa(54) = '1' then
v.opsel_a := AIN_B;
end if;
v.state := DO_FDIV; v.state := DO_FDIV;
when "10100" | "10101" => when "10100" | "10101" =>
v.opsel_a := AIN_A;
v.state := DO_FADD; v.state := DO_FADD;
when "10110" => when "10110" =>
v.is_sqrt := '1'; v.is_sqrt := '1';
v.opsel_a := AIN_B;
v.state := DO_FSQRT; v.state := DO_FSQRT;
when "10111" => when "10111" =>
v.state := DO_FSEL; v.state := DO_FSEL;
when "11000" => when "11000" =>
v.opsel_a := AIN_B;
v.state := DO_FRE; v.state := DO_FRE;
when "11001" => when "11001" =>
v.is_multiply := '1'; v.is_multiply := '1';
v.opsel_a := AIN_A;
if v.c.mantissa(54) = '0' and v.a.mantissa(54) = '1' then
v.opsel_a := AIN_C;
end if;
v.state := DO_FMUL; v.state := DO_FMUL;
when "11010" => when "11010" =>
v.is_sqrt := '1'; v.is_sqrt := '1';
v.opsel_a := AIN_B;
v.state := DO_FRSQRTE; v.state := DO_FRSQRTE;
when "11100" | "11101" | "11110" | "11111" => when "11100" | "11101" | "11110" | "11111" =>
if v.a.mantissa(54) = '0' then
v.opsel_a := AIN_A;
elsif v.c.mantissa(54) = '0' then
v.opsel_a := AIN_C;
else
v.opsel_a := AIN_B;
end if;
v.state := DO_FMADD; v.state := DO_FMADD;
when others => when others =>
illegal := '1'; illegal := '1';
@ -880,11 +915,10 @@ begin


when DO_FCMP => when DO_FCMP =>
-- fcmp[uo] -- fcmp[uo]
-- r.opsel_a = AIN_B
v.instr_done := '1'; v.instr_done := '1';
v.state := IDLE; v.state := IDLE;
update_fx := '1'; update_fx := '1';
opsel_a <= AIN_B;
opsel_r <= RES_SUM;
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
if (r.a.class = NAN and r.a.mantissa(53) = '0') or if (r.a.class = NAN and r.a.mantissa(53) = '0') or
(r.b.class = NAN and r.b.mantissa(53) = '0') then (r.b.class = NAN and r.b.mantissa(53) = '0') then
@ -930,6 +964,7 @@ begin
-- Prepare to subtract mantissas, put B in R -- Prepare to subtract mantissas, put B in R
v.cr_result := "0000"; v.cr_result := "0000";
v.instr_done := '0'; v.instr_done := '0';
v.opsel_a := AIN_A;
v.state := CMP_1; v.state := CMP_1;
end if; end if;
v.fpscr(FPSCR_FL downto FPSCR_FU) := v.cr_result; v.fpscr(FPSCR_FL downto FPSCR_FU) := v.cr_result;
@ -1017,7 +1052,7 @@ begin
v.state := IDLE; v.state := IDLE;


when DO_FMR => when DO_FMR =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
v.quieten_nan := '0'; v.quieten_nan := '0';
@ -1037,7 +1072,7 @@ begin
v.state := IDLE; v.state := IDLE;


when DO_FRI => -- fri[nzpm] when DO_FRI => -- fri[nzpm]
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
@ -1062,7 +1097,7 @@ begin
end if; end if;


when DO_FRSP => when DO_FRSP =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B, r.shift = 0
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
@ -1092,7 +1127,7 @@ begin
-- instr bit 9: 1=dword 0=word -- instr bit 9: 1=dword 0=word
-- instr bit 8: 1=unsigned 0=signed -- instr bit 8: 1=unsigned 0=signed
-- instr bit 1: 1=round to zero 0=use fpscr[RN] -- instr bit 1: 1=round to zero 0=use fpscr[RN]
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
@ -1130,8 +1165,8 @@ begin
end case; end case;


when DO_FCFID => when DO_FCFID =>
-- r.opsel_a = AIN_B
v.result_sign := '0'; v.result_sign := '0';
opsel_a <= AIN_B;
if r.insn(8) = '0' and r.b.negative = '1' then if r.insn(8) = '0' and r.b.negative = '1' then
-- fcfid[s] with negative operand, set R = -B -- fcfid[s] with negative operand, set R = -B
opsel_ainv <= '1'; opsel_ainv <= '1';
@ -1150,16 +1185,19 @@ begin


when DO_FADD => when DO_FADD =>
-- fadd[s] and fsub[s] -- fadd[s] and fsub[s]
opsel_a <= AIN_A; -- r.opsel_a = AIN_A
v.result_sign := r.a.negative; v.result_sign := r.a.negative;
v.result_class := r.a.class; v.result_class := r.a.class;
v.result_exp := r.a.exponent; v.result_exp := r.a.exponent;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
v.use_a := '1';
v.use_b := '1';
is_add := r.a.negative xor r.b.negative xor r.insn(1); is_add := r.a.negative xor r.b.negative xor r.insn(1);
if r.a.class = FINITE and r.b.class = FINITE then if r.a.class = FINITE and r.b.class = FINITE then
v.is_subtract := not is_add; v.is_subtract := not is_add;
v.add_bsmall := r.exp_cmp; v.add_bsmall := r.exp_cmp;
v.opsel_a := AIN_B;
if r.exp_cmp = '0' then if r.exp_cmp = '0' then
v.shift := r.a.exponent - r.b.exponent; v.shift := r.a.exponent - r.b.exponent;
v.result_sign := r.b.negative xnor r.insn(1); v.result_sign := r.b.negative xnor r.insn(1);
@ -1169,77 +1207,55 @@ begin
v.state := ADD_SHIFT; v.state := ADD_SHIFT;
end if; end if;
else else
opsel_a <= AIN_B; v.state := ADD_1;
v.shift := r.b.exponent - r.a.exponent;
v.result_exp := r.b.exponent;
v.state := ADD_SHIFT;
end if; end if;
else else
if (r.a.class = NAN and r.a.mantissa(53) = '0') or if r.a.class = NAN or r.b.class = NAN then
(r.b.class = NAN and r.b.mantissa(53) = '0') then v.state := NAN_RESULT;
-- Signalling NAN
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
if r.a.class = NAN then
-- nothing to do, result is A
elsif r.b.class = NAN then
v.result_class := NAN;
v.result_sign := r.b.negative;
opsel_a <= AIN_B;
elsif r.a.class = INFINITY and r.b.class = INFINITY and is_add = '0' then elsif r.a.class = INFINITY and r.b.class = INFINITY and is_add = '0' then
-- invalid operation, construct QNaN -- invalid operation, construct QNaN
v.fpscr(FPSCR_VXISI) := '1'; v.fpscr(FPSCR_VXISI) := '1';
qnan_result := '1'; qnan_result := '1';
arith_done := '1';
elsif r.a.class = ZERO and r.b.class = ZERO and is_add = '0' then elsif r.a.class = ZERO and r.b.class = ZERO and is_add = '0' then
-- return -0 for rounding to -infinity -- return -0 for rounding to -infinity
v.result_sign := r.round_mode(1) and r.round_mode(0); v.result_sign := r.round_mode(1) and r.round_mode(0);
arith_done := '1';
elsif r.a.class = INFINITY or r.b.class = ZERO then elsif r.a.class = INFINITY or r.b.class = ZERO then
-- nothing to do, result is A -- result is A
v.opsel_a := AIN_A;
v.state := EXC_RESULT;
else else
-- result is +/- B -- result is +/- B
v.result_sign := r.b.negative xnor r.insn(1); v.opsel_a := AIN_B;
v.result_class := r.b.class; v.negate := not r.insn(1);
v.result_exp := r.b.exponent; v.state := EXC_RESULT;
opsel_a <= AIN_B;
end if; end if;
arith_done := '1';
end if; end if;


when DO_FMUL => when DO_FMUL =>
-- fmul[s] -- fmul[s]
opsel_a <= AIN_A; -- r.opsel_a = AIN_A unless C is denorm and A isn't
v.result_sign := r.a.negative; v.result_sign := r.a.negative xor r.c.negative;
v.result_class := r.a.class; v.result_class := r.a.class;
v.result_exp := r.a.exponent;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
v.use_a := '1';
v.use_c := '1';
if r.a.class = FINITE and r.c.class = FINITE then if r.a.class = FINITE and r.c.class = FINITE then
v.result_sign := r.a.negative xor r.c.negative;
v.result_exp := r.a.exponent + r.c.exponent; v.result_exp := r.a.exponent + r.c.exponent;
-- Renormalize denorm operands -- Renormalize denorm operands
if r.a.mantissa(54) = '0' then if r.a.mantissa(54) = '0' then
v.state := RENORM_A; v.state := RENORM_A;
elsif r.c.mantissa(54) = '0' then elsif r.c.mantissa(54) = '0' then
opsel_a <= AIN_C;
v.state := RENORM_C; v.state := RENORM_C;
else else
f_to_multiply.valid <= '1'; f_to_multiply.valid <= '1';
v.state := MULT_1; v.state := MULT_1;
end if; end if;
else else
if (r.a.class = NAN and r.a.mantissa(53) = '0') or if r.a.class = NAN or r.c.class = NAN then
(r.c.class = NAN and r.c.mantissa(53) = '0') then v.state := NAN_RESULT;
-- Signalling NAN
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
if r.a.class = NAN then
-- result is A
elsif r.c.class = NAN then
v.result_class := NAN;
v.result_sign := r.c.negative;
opsel_a <= AIN_C;
elsif (r.a.class = INFINITY and r.c.class = ZERO) or elsif (r.a.class = INFINITY and r.c.class = ZERO) or
(r.a.class = ZERO and r.c.class = INFINITY) then (r.a.class = ZERO and r.c.class = INFINITY) then
-- invalid operation, construct QNaN -- invalid operation, construct QNaN
@ -1247,22 +1263,22 @@ begin
qnan_result := '1'; qnan_result := '1';
elsif r.a.class = ZERO or r.a.class = INFINITY then elsif r.a.class = ZERO or r.a.class = INFINITY then
-- result is +/- A -- result is +/- A
v.result_sign := r.a.negative xor r.c.negative; arith_done := '1';
else else
-- r.c.class is ZERO or INFINITY -- r.c.class is ZERO or INFINITY
v.result_class := r.c.class; v.opsel_a := AIN_C;
v.result_sign := r.a.negative xor r.c.negative; v.negate := r.a.negative;
v.state := EXC_RESULT;
end if; end if;
arith_done := '1';
end if; end if;


when DO_FDIV => when DO_FDIV =>
opsel_a <= AIN_A; -- r.opsel_a = AIN_A unless B is denorm and A isn't
v.result_sign := r.a.negative;
v.result_class := r.a.class; v.result_class := r.a.class;
v.result_exp := r.a.exponent;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
v.use_a := '1';
v.use_b := '1';
v.result_sign := r.a.negative xor r.b.negative; v.result_sign := r.a.negative xor r.b.negative;
v.result_exp := r.a.exponent - r.b.exponent; v.result_exp := r.a.exponent - r.b.exponent;
v.count := "00"; v.count := "00";
@ -1271,26 +1287,14 @@ begin
if r.a.mantissa(54) = '0' then if r.a.mantissa(54) = '0' then
v.state := RENORM_A; v.state := RENORM_A;
elsif r.b.mantissa(54) = '0' then elsif r.b.mantissa(54) = '0' then
opsel_a <= AIN_B;
v.state := RENORM_B; v.state := RENORM_B;
else else
v.first := '1'; v.first := '1';
v.state := DIV_2; v.state := DIV_2;
end if; end if;
else else
if (r.a.class = NAN and r.a.mantissa(53) = '0') or if r.a.class = NAN or r.b.class = NAN then
(r.b.class = NAN and r.b.mantissa(53) = '0') then v.state := NAN_RESULT;
-- Signalling NAN
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
if r.a.class = NAN then
-- result is A
v.result_sign := r.a.negative;
elsif r.b.class = NAN then
v.result_class := NAN;
v.result_sign := r.b.negative;
opsel_a <= AIN_B;
elsif r.b.class = INFINITY then elsif r.b.class = INFINITY then
if r.a.class = INFINITY then if r.a.class = INFINITY then
v.fpscr(FPSCR_VXIDI) := '1'; v.fpscr(FPSCR_VXIDI) := '1';
@ -1298,6 +1302,7 @@ begin
else else
v.result_class := ZERO; v.result_class := ZERO;
end if; end if;
arith_done := '1';
elsif r.b.class = ZERO then elsif r.b.class = ZERO then
if r.a.class = ZERO then if r.a.class = ZERO then
v.fpscr(FPSCR_VXZDZ) := '1'; v.fpscr(FPSCR_VXZDZ) := '1';
@ -1308,46 +1313,36 @@ begin
end if; end if;
v.result_class := INFINITY; v.result_class := INFINITY;
end if; end if;
-- else r.b.class = FINITE, result_class = r.a.class arith_done := '1';
end if; else -- r.b.class = FINITE, result_class = r.a.class
arith_done := '1'; arith_done := '1';
end if; end if;
end if;


when DO_FSEL => when DO_FSEL =>
opsel_a <= AIN_A;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
if r.a.class = ZERO or (r.a.negative = '0' and r.a.class /= NAN) then if r.a.class = ZERO or (r.a.negative = '0' and r.a.class /= NAN) then
v.result_sign := r.c.negative; v.opsel_a := AIN_C;
v.result_exp := r.c.exponent;
v.result_class := r.c.class;
opsel_a <= AIN_C;
else else
v.result_sign := r.b.negative; v.opsel_a := AIN_B;
v.result_exp := r.b.exponent;
v.result_class := r.b.class;
opsel_a <= AIN_B;
end if; end if;
v.quieten_nan := '0'; v.quieten_nan := '0';
arith_done := '1'; v.state := EXC_RESULT;


when DO_FSQRT => when DO_FSQRT =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
if r.b.class = NAN and r.b.mantissa(53) = '0' then v.use_b := '1';
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
case r.b.class is case r.b.class is
when FINITE => when FINITE =>
v.result_exp := r.b.exponent; v.result_exp := r.b.exponent;
if r.b.negative = '1' then if r.b.negative = '1' then
v.fpscr(FPSCR_VXSQRT) := '1'; v.fpscr(FPSCR_VXSQRT) := '1';
qnan_result := '1'; qnan_result := '1';
arith_done := '1';
elsif r.b.mantissa(54) = '0' then elsif r.b.mantissa(54) = '0' then
v.state := RENORM_B; v.state := RENORM_B;
elsif r.b.exponent(0) = '0' then elsif r.b.exponent(0) = '0' then
@ -1356,7 +1351,9 @@ begin
v.shift := to_signed(1, EXP_BITS); v.shift := to_signed(1, EXP_BITS);
v.state := RENORM_B2; v.state := RENORM_B2;
end if; end if;
when NAN | ZERO => when NAN =>
v.state := NAN_RESULT;
when ZERO =>
-- result is B -- result is B
arith_done := '1'; arith_done := '1';
when INFINITY => when INFINITY =>
@ -1369,15 +1366,12 @@ begin
end case; end case;


when DO_FRE => when DO_FRE =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
if r.b.class = NAN and r.b.mantissa(53) = '0' then v.use_b := '1';
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
case r.b.class is case r.b.class is
when FINITE => when FINITE =>
v.result_exp := - r.b.exponent; v.result_exp := - r.b.exponent;
@ -1387,8 +1381,7 @@ begin
v.state := FRE_1; v.state := FRE_1;
end if; end if;
when NAN => when NAN =>
-- result is B v.state := NAN_RESULT;
arith_done := '1';
when INFINITY => when INFINITY =>
v.result_class := ZERO; v.result_class := ZERO;
arith_done := '1'; arith_done := '1';
@ -1399,15 +1392,12 @@ begin
end case; end case;


when DO_FRSQRTE => when DO_FRSQRTE =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
v.result_class := r.b.class; v.result_class := r.b.class;
v.result_sign := r.b.negative; v.result_sign := r.b.negative;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
if r.b.class = NAN and r.b.mantissa(53) = '0' then v.use_b := '1';
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
v.shift := to_signed(1, EXP_BITS); v.shift := to_signed(1, EXP_BITS);
case r.b.class is case r.b.class is
when FINITE => when FINITE =>
@ -1415,7 +1405,6 @@ begin
if r.b.negative = '1' then if r.b.negative = '1' then
v.fpscr(FPSCR_VXSQRT) := '1'; v.fpscr(FPSCR_VXSQRT) := '1';
qnan_result := '1'; qnan_result := '1';
arith_done := '1';
elsif r.b.mantissa(54) = '0' then elsif r.b.mantissa(54) = '0' then
v.state := RENORM_B; v.state := RENORM_B;
elsif r.b.exponent(0) = '0' then elsif r.b.exponent(0) = '0' then
@ -1424,8 +1413,7 @@ begin
v.state := RENORM_B2; v.state := RENORM_B2;
end if; end if;
when NAN => when NAN =>
-- result is B v.state := NAN_RESULT;
arith_done := '1';
when INFINITY => when INFINITY =>
if r.b.negative = '1' then if r.b.negative = '1' then
v.fpscr(FPSCR_VXSQRT) := '1'; v.fpscr(FPSCR_VXSQRT) := '1';
@ -1442,25 +1430,26 @@ begin


when DO_FMADD => when DO_FMADD =>
-- fmadd, fmsub, fnmadd, fnmsub -- fmadd, fmsub, fnmadd, fnmsub
opsel_a <= AIN_A; -- r.opsel_a = AIN_A if A is denorm, else AIN_C if C is denorm,
-- else AIN_B
v.result_sign := r.a.negative; v.result_sign := r.a.negative;
v.result_class := r.a.class; v.result_class := r.a.class;
v.result_exp := r.a.exponent; v.result_exp := r.a.exponent;
v.fpscr(FPSCR_FR) := '0'; v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FI) := '0'; v.fpscr(FPSCR_FI) := '0';
v.use_a := '1';
v.use_b := '1';
v.use_c := '1';
is_add := r.a.negative xor r.c.negative xor r.b.negative xor r.insn(1); is_add := r.a.negative xor r.c.negative xor r.b.negative xor r.insn(1);
if r.a.class = FINITE and r.c.class = FINITE and if r.a.class = FINITE and r.c.class = FINITE and
(r.b.class = FINITE or r.b.class = ZERO) then (r.b.class = FINITE or r.b.class = ZERO) then
v.is_subtract := not is_add; v.is_subtract := not is_add;
mulexp := r.a.exponent + r.c.exponent; mulexp := r.a.exponent + r.c.exponent;
v.result_exp := mulexp; v.result_exp := mulexp;
opsel_a <= AIN_B;
-- Make sure A and C are normalized -- Make sure A and C are normalized
if r.a.mantissa(54) = '0' then if r.a.mantissa(54) = '0' then
opsel_a <= AIN_A;
v.state := RENORM_A; v.state := RENORM_A;
elsif r.c.mantissa(54) = '0' then elsif r.c.mantissa(54) = '0' then
opsel_a <= AIN_C;
v.state := RENORM_C; v.state := RENORM_C;
elsif r.b.class = ZERO then elsif r.b.class = ZERO then
-- no addend, degenerates to multiply -- no addend, degenerates to multiply
@ -1483,25 +1472,8 @@ begin
v.state := FMADD_2; v.state := FMADD_2;
end if; end if;
else else
if (r.a.class = NAN and r.a.mantissa(53) = '0') or if r.a.class = NAN or r.b.class = NAN or r.c.class = NAN then
(r.b.class = NAN and r.b.mantissa(53) = '0') or v.state := NAN_RESULT;
(r.c.class = NAN and r.c.mantissa(53) = '0') then
-- Signalling NAN
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
if r.a.class = NAN then
-- nothing to do, result is A
elsif r.b.class = NAN then
-- result is B
v.result_class := NAN;
v.result_sign := r.b.negative;
opsel_a <= AIN_B;
elsif r.c.class = NAN then
-- result is C
v.result_class := NAN;
v.result_sign := r.c.negative;
opsel_a <= AIN_C;
elsif (r.a.class = ZERO and r.c.class = INFINITY) or elsif (r.a.class = ZERO and r.c.class = INFINITY) or
(r.a.class = INFINITY and r.c.class = ZERO) then (r.a.class = INFINITY and r.c.class = ZERO) then
-- invalid operation, construct QNaN -- invalid operation, construct QNaN
@ -1516,32 +1488,36 @@ begin
-- result is infinity -- result is infinity
v.result_class := INFINITY; v.result_class := INFINITY;
v.result_sign := r.a.negative xor r.c.negative xor r.insn(2); v.result_sign := r.a.negative xor r.c.negative xor r.insn(2);
arith_done := '1';
end if; end if;
else else
-- Here A is zero, C is zero, or B is infinity -- Here A is zero, C is zero, or B is infinity
-- Result is +/-B in all of those cases -- Result is +/-B in all of those cases
v.result_class := r.b.class; v.opsel_a := AIN_B;
v.result_exp := r.b.exponent; if r.b.class /= ZERO or is_add = '1' then
if v.result_class /= ZERO or is_add = '1' then v.negate := not (r.insn(1) xor r.insn(2));
v.result_sign := not (r.b.negative xor r.insn(1) xor r.insn(2));
else else
-- have to be careful about rule for 0 - 0 result sign -- have to be careful about rule for 0 - 0 result sign
v.result_sign := (r.round_mode(1) and r.round_mode(0)) xor r.insn(2); v.negate := r.b.negative xor (r.round_mode(1) and r.round_mode(0)) xor r.insn(2);
end if; end if;
opsel_a <= AIN_B; v.state := EXC_RESULT;
end if; end if;
arith_done := '1';
end if; end if;


when RENORM_A => when RENORM_A =>
renormalize := '1'; renormalize := '1';
v.state := RENORM_A2; v.state := RENORM_A2;
if r.insn(4) = '1' then
v.opsel_a := AIN_C;
else
v.opsel_a := AIN_B;
end if;


when RENORM_A2 => when RENORM_A2 =>
-- r.opsel_a = AIN_C for fmul/fmadd, AIN_B for fdiv
set_a := '1'; set_a := '1';
v.result_exp := new_exp; v.result_exp := new_exp;
if r.insn(4) = '1' then if r.insn(4) = '1' then
opsel_a <= AIN_C;
if r.c.mantissa(54) = '1' then if r.c.mantissa(54) = '1' then
if r.insn(3) = '0' or r.b.class = ZERO then if r.insn(3) = '0' or r.b.class = ZERO then
v.first := '1'; v.first := '1';
@ -1551,13 +1527,13 @@ begin
if new_exp + 1 >= r.b.exponent then if new_exp + 1 >= r.b.exponent then
v.madd_cmp := '1'; v.madd_cmp := '1';
end if; end if;
v.opsel_a := AIN_B;
v.state := DO_FMADD; v.state := DO_FMADD;
end if; end if;
else else
v.state := RENORM_C; v.state := RENORM_C;
end if; end if;
else else
opsel_a <= AIN_B;
if r.b.mantissa(54) = '1' then if r.b.mantissa(54) = '1' then
v.first := '1'; v.first := '1';
v.state := DIV_2; v.state := DIV_2;
@ -1578,6 +1554,7 @@ begin
else else
v.result_exp := new_exp; v.result_exp := new_exp;
end if; end if;
v.opsel_a := AIN_B;
v.state := LOOKUP; v.state := LOOKUP;


when RENORM_C => when RENORM_C =>
@ -1595,23 +1572,31 @@ begin
if new_exp + 1 >= r.b.exponent then if new_exp + 1 >= r.b.exponent then
v.madd_cmp := '1'; v.madd_cmp := '1';
end if; end if;
v.opsel_a := AIN_B;
v.state := DO_FMADD; v.state := DO_FMADD;
end if; end if;


when ADD_1 =>
-- transferring B to R
v.shift := r.b.exponent - r.a.exponent;
v.result_exp := r.b.exponent;
v.state := ADD_SHIFT;

when ADD_SHIFT => when ADD_SHIFT =>
-- r.shift = - exponent difference -- r.shift = - exponent difference
opsel_r <= RES_SHIFT; opsel_r <= RES_SHIFT;
v.x := s_nz; v.x := s_nz;
set_x := '1'; set_x := '1';
longmask := '0'; longmask := '0';
v.state := ADD_2;

when ADD_2 =>
if r.add_bsmall = '1' then if r.add_bsmall = '1' then
opsel_a <= AIN_A; v.opsel_a := AIN_A;
else else
opsel_a <= AIN_B; v.opsel_a := AIN_B;
end if; end if;
v.state := ADD_2;

when ADD_2 =>
-- r.opsel_a = AIN_A if r.add_bsmall = 1 else AIN_B
opsel_b <= BIN_R; opsel_b <= BIN_R;
opsel_binv <= r.is_subtract; opsel_binv <= r.is_subtract;
carry_in <= r.is_subtract and not r.x; carry_in <= r.is_subtract and not r.x;
@ -1655,7 +1640,7 @@ begin
end if; end if;


when CMP_1 => when CMP_1 =>
opsel_a <= AIN_A; -- r.opsel_a = AIN_A
opsel_b <= BIN_R; opsel_b <= BIN_R;
opsel_binv <= '1'; opsel_binv <= '1';
carry_in <= '1'; carry_in <= '1';
@ -1696,7 +1681,7 @@ begin


when FMADD_2 => when FMADD_2 =>
-- Product is potentially bigger here -- Product is potentially bigger here
-- r.shift = addend exp - product exp + 64 -- r.shift = addend exp - product exp + 64, r.r = r.b.mantissa
set_s := '1'; set_s := '1';
opsel_s <= S_SHIFT; opsel_s <= S_SHIFT;
v.shift := r.shift - to_signed(64, EXP_BITS); v.shift := r.shift - to_signed(64, EXP_BITS);
@ -1757,7 +1742,7 @@ begin
end if; end if;


when LOOKUP => when LOOKUP =>
opsel_a <= AIN_B; -- r.opsel_a = AIN_B
-- wait one cycle for inverse_table[B] lookup -- wait one cycle for inverse_table[B] lookup
v.first := '1'; v.first := '1';
if r.insn(4) = '0' then if r.insn(4) = '0' then
@ -2260,6 +2245,41 @@ begin
opsel_r <= RES_SHIFT; opsel_r <= RES_SHIFT;
arith_done := '1'; arith_done := '1';


when NAN_RESULT =>
if (r.use_a = '1' and r.a.class = NAN and r.a.mantissa(53) = '0') or
(r.use_b = '1' and r.b.class = NAN and r.b.mantissa(53) = '0') or
(r.use_c = '1' and r.c.class = NAN and r.c.mantissa(53) = '0') then
-- Signalling NAN
v.fpscr(FPSCR_VXSNAN) := '1';
invalid := '1';
end if;
if r.use_a = '1' and r.a.class = NAN then
v.opsel_a := AIN_A;
elsif r.use_b = '1' and r.b.class = NAN then
v.opsel_a := AIN_B;
elsif r.use_c = '1' and r.c.class = NAN then
v.opsel_a := AIN_C;
end if;
v.state := EXC_RESULT;

when EXC_RESULT =>
-- r.opsel_a = AIN_A, AIN_B or AIN_C according to which input is the result
case r.opsel_a is
when AIN_B =>
v.result_sign := r.b.negative xor r.negate;
v.result_exp := r.b.exponent;
v.result_class := r.b.class;
when AIN_C =>
v.result_sign := r.c.negative xor r.negate;
v.result_exp := r.c.exponent;
v.result_class := r.c.class;
when others =>
v.result_sign := r.a.negative xor r.negate;
v.result_exp := r.a.exponent;
v.result_class := r.a.class;
end case;
arith_done := '1';

end case; end case;


if zero_divide = '1' then if zero_divide = '1' then
@ -2271,11 +2291,15 @@ begin
v.result_sign := '0'; v.result_sign := '0';
misc_sel <= "0001"; misc_sel <= "0001";
opsel_r <= RES_MISC; opsel_r <= RES_MISC;
arith_done := '1';
end if;
if invalid = '1' then
v.invalid := '1';
end if; end if;
if arith_done = '1' then if arith_done = '1' then
-- Enabled invalid exception doesn't write result or FPRF -- Enabled invalid exception doesn't write result or FPRF
-- Neither does enabled zero-divide exception -- Neither does enabled zero-divide exception
if (invalid and r.fpscr(FPSCR_VE)) = '0' and if (v.invalid and r.fpscr(FPSCR_VE)) = '0' and
(zero_divide and r.fpscr(FPSCR_ZE)) = '0' then (zero_divide and r.fpscr(FPSCR_ZE)) = '0' then
v.writing_back := '1'; v.writing_back := '1';
v.update_fprf := '1'; v.update_fprf := '1';
@ -2355,7 +2379,7 @@ begin
else else
mask := right_mask(unsigned(mshift(5 downto 0))); mask := right_mask(unsigned(mshift(5 downto 0)));
end if; end if;
case opsel_a is case r.opsel_a is
when AIN_R => when AIN_R =>
in_a0 := r.r; in_a0 := r.r;
when AIN_A => when AIN_A =>

Loading…
Cancel
Save