loadstore1/dcache: Send store data one cycle later

This makes timing easier and also means that store floating-point
single precision instructions no longer need to take an extra cycle.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
pull/268/head
Paul Mackerras 4 years ago
parent d1f35705c0
commit 6427cab46f

@ -365,7 +365,7 @@ package common is
virt_mode : std_ulogic; virt_mode : std_ulogic;
priv_mode : std_ulogic; priv_mode : std_ulogic;
addr : std_ulogic_vector(63 downto 0); addr : std_ulogic_vector(63 downto 0);
data : std_ulogic_vector(63 downto 0); data : std_ulogic_vector(63 downto 0); -- valid the cycle after .valid = 1
byte_sel : std_ulogic_vector(7 downto 0); byte_sel : std_ulogic_vector(7 downto 0);
end record; end record;



@ -1306,7 +1306,7 @@ begin
req.real_addr := ra; req.real_addr := ra;
-- Force data to 0 for dcbz -- Force data to 0 for dcbz
if r0.req.dcbz = '0' then if r0.req.dcbz = '0' then
req.data := r0.req.data; req.data := d_in.data;
else else
req.data := (others => '0'); req.data := (others => '0');
end if; end if;

@ -45,7 +45,6 @@ architecture behave of loadstore1 is


-- State machine for unaligned loads/stores -- State machine for unaligned loads/stores
type state_t is (IDLE, -- ready for instruction type state_t is (IDLE, -- ready for instruction
FPR_CONV, -- converting double to float for store
SECOND_REQ, -- send 2nd request of unaligned xfer SECOND_REQ, -- send 2nd request of unaligned xfer
ACK_WAIT, -- waiting for ack from dcache ACK_WAIT, -- waiting for ack from dcache
MMU_LOOKUP, -- waiting for MMU to look up translation MMU_LOOKUP, -- waiting for MMU to look up translation
@ -69,6 +68,8 @@ architecture behave of loadstore1 is
write_reg : gspr_index_t; write_reg : gspr_index_t;
length : std_ulogic_vector(3 downto 0); length : std_ulogic_vector(3 downto 0);
byte_reverse : std_ulogic; byte_reverse : std_ulogic;
byte_offset : unsigned(2 downto 0);
brev_mask : unsigned(2 downto 0);
sign_extend : std_ulogic; sign_extend : std_ulogic;
update : std_ulogic; update : std_ulogic;
update_reg : gpr_index_t; update_reg : gpr_index_t;
@ -103,7 +104,6 @@ architecture behave of loadstore1 is
ld_sp_data : std_ulogic_vector(31 downto 0); ld_sp_data : std_ulogic_vector(31 downto 0);
ld_sp_nz : std_ulogic; ld_sp_nz : std_ulogic;
ld_sp_lz : std_ulogic_vector(5 downto 0); ld_sp_lz : std_ulogic_vector(5 downto 0);
st_sp_data : std_ulogic_vector(31 downto 0);
wr_sel : std_ulogic_vector(1 downto 0); wr_sel : std_ulogic_vector(1 downto 0);
end record; end record;


@ -299,7 +299,6 @@ begin
variable data_permuted : std_ulogic_vector(63 downto 0); variable data_permuted : std_ulogic_vector(63 downto 0);
variable data_trimmed : std_ulogic_vector(63 downto 0); variable data_trimmed : std_ulogic_vector(63 downto 0);
variable store_data : std_ulogic_vector(63 downto 0); variable store_data : std_ulogic_vector(63 downto 0);
variable data_in : std_ulogic_vector(63 downto 0);
variable byte_rev : std_ulogic; variable byte_rev : std_ulogic;
variable length : std_ulogic_vector(3 downto 0); variable length : std_ulogic_vector(3 downto 0);
variable negative : std_ulogic; variable negative : std_ulogic;
@ -311,7 +310,6 @@ begin
variable mmu_mtspr : std_ulogic; variable mmu_mtspr : std_ulogic;
variable itlb_fault : std_ulogic; variable itlb_fault : std_ulogic;
variable misaligned : std_ulogic; variable misaligned : std_ulogic;
variable fp_reg_conv : std_ulogic;
begin begin
v := r; v := r;
req := '0'; req := '0';
@ -320,7 +318,6 @@ begin
sprn := std_ulogic_vector(to_unsigned(decode_spr_num(l_in.insn), 10)); sprn := std_ulogic_vector(to_unsigned(decode_spr_num(l_in.insn), 10));
dsisr := (others => '0'); dsisr := (others => '0');
mmureq := '0'; mmureq := '0';
fp_reg_conv := '0';
v.wr_sel := "11"; v.wr_sel := "11";


write_enable := '0'; write_enable := '0';
@ -366,40 +363,19 @@ begin
end loop; end loop;


if HAS_FPU then if HAS_FPU then
-- Single-precision FP conversion -- Single-precision FP conversion for loads
v.st_sp_data := store_sp_data;
v.ld_sp_data := data_trimmed(31 downto 0); v.ld_sp_data := data_trimmed(31 downto 0);
v.ld_sp_nz := or (data_trimmed(22 downto 0)); v.ld_sp_nz := or (data_trimmed(22 downto 0));
v.ld_sp_lz := count_left_zeroes(data_trimmed(22 downto 0)); v.ld_sp_lz := count_left_zeroes(data_trimmed(22 downto 0));
end if; end if;


-- Byte reversing and rotating for stores. -- Byte reversing and rotating for stores.
-- Done in the first cycle (when l_in.valid = 1) for integer stores -- Done in the second cycle (the cycle after l_in.valid = 1).
-- and DP float stores, and in the second cycle for SP float stores. for i in 0 to 7 loop
store_data := r.store_data; k := (to_unsigned(i, 3) - r.byte_offset) xor r.brev_mask;
if l_in.valid = '1' or (HAS_FPU and r.state = FPR_CONV) then j := to_integer(k) * 8;
if HAS_FPU and r.state = FPR_CONV then store_data(i * 8 + 7 downto i * 8) := r.store_data(j + 7 downto j);
data_in := x"00000000" & r.st_sp_data; end loop;
byte_offset := unsigned(r.addr(2 downto 0));
byte_rev := r.byte_reverse;
length := r.length;
else
data_in := l_in.data;
byte_offset := unsigned(lsu_sum(2 downto 0));
byte_rev := l_in.byte_reverse;
length := l_in.length;
end if;
brev_lenm1 := "000";
if byte_rev = '1' then
brev_lenm1 := unsigned(length(2 downto 0)) - 1;
end if;
for i in 0 to 7 loop
k := (to_unsigned(i, 3) - byte_offset) xor brev_lenm1;
j := to_integer(k) * 8;
store_data(i * 8 + 7 downto i * 8) := data_in(j + 7 downto j);
end loop;
end if;
v.store_data := store_data;


-- compute (addr + 8) & ~7 for the second doubleword when unaligned -- compute (addr + 8) & ~7 for the second doubleword when unaligned
next_addr := std_ulogic_vector(unsigned(r.addr(63 downto 3)) + 1) & "000"; next_addr := std_ulogic_vector(unsigned(r.addr(63 downto 3)) + 1) & "000";
@ -431,14 +407,6 @@ begin
case r.state is case r.state is
when IDLE => when IDLE =>


when FPR_CONV =>
req := '1';
if r.second_bytes /= "00000000" then
v.state := SECOND_REQ;
else
v.state := ACK_WAIT;
end if;

when SECOND_REQ => when SECOND_REQ =>
req := '1'; req := '1';
v.state := ACK_WAIT; v.state := ACK_WAIT;
@ -561,6 +529,12 @@ begin
v.do_update := '0'; v.do_update := '0';
v.extra_cycle := '0'; v.extra_cycle := '0';


if HAS_FPU and l_in.is_32bit = '1' then
v.store_data := x"00000000" & store_sp_data;
else
v.store_data := l_in.data;
end if;

addr := lsu_sum; addr := lsu_sum;
if l_in.second = '1' then if l_in.second = '1' then
-- for the second half of a 16-byte transfer, use next_addr -- for the second half of a 16-byte transfer, use next_addr
@ -609,12 +583,7 @@ begin


case l_in.op is case l_in.op is
when OP_STORE => when OP_STORE =>
if HAS_FPU and l_in.is_32bit = '1' then req := '1';
v.state := FPR_CONV;
fp_reg_conv := '1';
else
req := '1';
end if;
when OP_LOAD => when OP_LOAD =>
req := '1'; req := '1';
v.load := '1'; v.load := '1';
@ -684,7 +653,20 @@ begin
end if; end if;
end if; end if;


v.busy := req or mmureq or mmu_mtspr or fp_reg_conv; v.busy := req or mmureq or mmu_mtspr;
end if;

-- Work out controls for store formatting
if l_in.valid = '1' then
byte_offset := unsigned(lsu_sum(2 downto 0));
byte_rev := l_in.byte_reverse;
length := l_in.length;
brev_lenm1 := "000";
if byte_rev = '1' then
brev_lenm1 := unsigned(length(2 downto 0)) - 1;
end if;
v.byte_offset := byte_offset;
v.brev_mask := brev_lenm1;
end if; end if;


-- Work out load formatter controls for next cycle -- Work out load formatter controls for next cycle

Loading…
Cancel
Save