You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
microwatt/writeback.vhdl

233 lines
8.0 KiB
VHDL

library ieee;
use ieee.std_logic_1164.all;
use ieee.numeric_std.all;
library work;
use work.common.all;
use work.crhelpers.all;
entity writeback is
port (
clk : in std_ulogic;
e_in : in Execute1ToWritebackType;
l_in : in DcacheToWritebackType;
m_in : in MultiplyToWritebackType;
d_in : in DividerToWritebackType;
w_out : out WritebackToRegisterFileType;
c_out : out WritebackToCrFileType;
complete_out : out std_ulogic
);
end entity writeback;
architecture behaviour of writeback is
subtype byte_index_t is unsigned(2 downto 0);
type permutation_t is array(0 to 7) of byte_index_t;
subtype byte_trim_t is std_ulogic_vector(1 downto 0);
type trim_ctl_t is array(0 to 7) of byte_trim_t;
type byte_sel_t is array(0 to 7) of std_ulogic;
signal data_len : unsigned(3 downto 0);
signal data_in : std_ulogic_vector(63 downto 0);
signal data_permuted : std_ulogic_vector(63 downto 0);
signal data_trimmed : std_ulogic_vector(63 downto 0);
signal data_latched : std_ulogic_vector(63 downto 0);
signal perm : permutation_t;
signal use_second : byte_sel_t;
signal byte_offset : unsigned(2 downto 0);
signal brev_lenm1 : unsigned(2 downto 0);
signal trim_ctl : trim_ctl_t;
signal rc : std_ulogic;
signal partial_write : std_ulogic;
signal sign_extend : std_ulogic;
signal negative : std_ulogic;
signal second_word : std_ulogic;
signal zero : std_ulogic;
begin
writeback_0: process(clk)
begin
if rising_edge(clk) then
if partial_write = '1' then
data_latched <= data_permuted;
end if;
end if;
end process;
writeback_1: process(all)
variable x : std_ulogic_vector(0 downto 0);
variable y : std_ulogic_vector(0 downto 0);
variable z : std_ulogic_vector(0 downto 0);
variable w : std_ulogic_vector(0 downto 0);
variable j : integer;
variable k : unsigned(3 downto 0);
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
variable cf: std_ulogic_vector(3 downto 0);
variable xe: xer_common_t;
begin
x := "" & e_in.valid;
y := "" & l_in.valid;
z := "" & m_in.valid;
w := "" & d_in.valid;
assert (to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z)) + to_integer(unsigned(w))) <= 1 severity failure;
x := "" & e_in.write_enable;
y := "" & l_in.write_enable;
z := "" & m_in.write_reg_enable;
w := "" & d_in.write_reg_enable;
assert (to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z)) + to_integer(unsigned(w))) <= 1 severity failure;
w := "" & e_in.write_cr_enable;
x := "" & (e_in.write_enable and e_in.rc);
y := "" & (m_in.valid and m_in.rc);
z := "" & (d_in.valid and d_in.rc);
assert (to_integer(unsigned(w)) + to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z))) <= 1 severity failure;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
x := "" & e_in.write_xerc_enable;
y := "" & m_in.write_xerc_enable;
z := "" & D_in.write_xerc_enable;
assert (to_integer(unsigned(x)) + to_integer(unsigned(y)) + to_integer(unsigned(z))) <= 1 severity failure;
w_out <= WritebackToRegisterFileInit;
c_out <= WritebackToCrFileInit;
complete_out <= '0';
if e_in.valid = '1' or l_in.valid = '1' or m_in.valid = '1' or d_in.valid = '1' then
complete_out <= '1';
end if;
rc <= '0';
brev_lenm1 <= "000";
byte_offset <= "000";
data_len <= x"8";
partial_write <= '0';
sign_extend <= '0';
second_word <= '0';
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
xe := e_in.xerc;
if e_in.write_enable = '1' then
w_out.write_reg <= e_in.write_reg;
w_out.write_enable <= '1';
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
data_in <= e_in.write_data;
data_len <= unsigned(e_in.write_len);
sign_extend <= e_in.sign_extend;
rc <= e_in.rc;
end if;
if e_in.write_cr_enable = '1' then
c_out.write_cr_enable <= '1';
c_out.write_cr_mask <= e_in.write_cr_mask;
c_out.write_cr_data <= e_in.write_cr_data;
end if;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
if e_in.write_xerc_enable = '1' then
c_out.write_xerc_enable <= '1';
c_out.write_xerc_data <= e_in.xerc;
end if;
if l_in.write_enable = '1' then
w_out.write_reg <= gpr_to_gspr(l_in.write_reg);
data_in <= l_in.write_data;
data_len <= unsigned(l_in.write_len);
byte_offset <= unsigned(l_in.write_shift);
sign_extend <= l_in.sign_extend;
if l_in.byte_reverse = '1' then
brev_lenm1 <= unsigned(l_in.write_len(2 downto 0)) - 1;
end if;
w_out.write_enable <= '1';
second_word <= l_in.second_word;
if l_in.valid = '0' and (data_len + byte_offset > 8) then
partial_write <= '1';
end if;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
xe := l_in.xerc;
end if;
if m_in.write_reg_enable = '1' then
w_out.write_enable <= '1';
w_out.write_reg <= gpr_to_gspr(m_in.write_reg_nr);
data_in <= m_in.write_reg_data;
rc <= m_in.rc;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
xe := m_in.xerc;
end if;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
if m_in.write_xerc_enable = '1' then
c_out.write_xerc_enable <= '1';
c_out.write_xerc_data <= m_in.xerc;
end if;
if d_in.write_reg_enable = '1' then
w_out.write_enable <= '1';
w_out.write_reg <= gpr_to_gspr(d_in.write_reg_nr);
data_in <= d_in.write_reg_data;
rc <= d_in.rc;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
xe := d_in.xerc;
end if;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
if d_in.write_xerc_enable = '1' then
c_out.write_xerc_enable <= '1';
c_out.write_xerc_data <= d_in.xerc;
end if;
-- shift and byte-reverse data bytes
for i in 0 to 7 loop
k := ('0' & (to_unsigned(i, 3) xor brev_lenm1)) + ('0' & byte_offset);
perm(i) <= k(2 downto 0);
use_second(i) <= k(3);
end loop;
for i in 0 to 7 loop
j := to_integer(perm(i)) * 8;
data_permuted(i * 8 + 7 downto i * 8) <= data_in(j + 7 downto j);
end loop;
-- If the data can arrive split over two cycles, this will be correct
-- provided we don't have both sign extension and byte reversal.
negative <= (data_len(3) and data_permuted(63)) or
(data_len(2) and data_permuted(31)) or
(data_len(1) and data_permuted(15)) or
(data_len(0) and data_permuted(7));
-- trim and sign-extend
for i in 0 to 7 loop
if i < to_integer(data_len) then
if second_word = '1' then
trim_ctl(i) <= '1' & not use_second(i);
else
trim_ctl(i) <= not use_second(i) & '0';
end if;
else
trim_ctl(i) <= '0' & (negative and sign_extend);
end if;
end loop;
zero <= not negative;
for i in 0 to 7 loop
case trim_ctl(i) is
when "11" =>
data_trimmed(i * 8 + 7 downto i * 8) <= data_latched(i * 8 + 7 downto i * 8);
when "10" =>
data_trimmed(i * 8 + 7 downto i * 8) <= data_permuted(i * 8 + 7 downto i * 8);
if or data_permuted(i * 8 + 7 downto i * 8) /= '0' then
zero <= '0';
end if;
when "01" =>
data_trimmed(i * 8 + 7 downto i * 8) <= x"FF";
when others =>
data_trimmed(i * 8 + 7 downto i * 8) <= x"00";
end case;
end loop;
-- deliver to regfile
w_out.write_data <= data_trimmed;
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
-- Perform CR0 update for RC forms
if rc = '1' then
c_out.write_cr_enable <= '1';
c_out.write_cr_mask <= num_to_fxm(0);
Add basic XER support The carry is currently internal to execute1. We don't handle any of the other XER fields. This creates type called "xer_common_t" that contains the commonly used XER bits (CA, CA32, SO, OV, OV32). The value is stored in the CR file (though it could be a separate module). The rest of the bits will be implemented as a separate SPR and the two parts reconciled in mfspr/mtspr in latter commits. We always read XER in decode2 (there is little point not to) and send it down all pipeline branches as it will be needed in writeback for all type of instructions when CR0:SO needs to be updated (such forms exist for all pipeline branches even if we don't yet implement them). To avoid having to track XER hazards, we forward it back in EX1. This assumes that other pipeline branches that can modify it (mult and div) are running single issue for now. One additional hazard to beware of is an XER:SO modifying instruction in EX1 followed immediately by a store conditional. Due to our writeback latency, the store will go down the LSU with the previous XER value, thus the stcx. will set CR0:SO using an obsolete SO value. I doubt there exist any code relying on this behaviour being correct but we should account for it regardless, possibly by ensuring that stcx. remain single issue initially, or later by adding some minimal tracking or moving the LSU into the same pipeline as execute. Missing some obscure XER affecting instructions like addex or mcrxrx. [paulus@ozlabs.org - fix CA32 and OV32 for OP_ADD, fix order of arguments to set_ov] Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
4 years ago
cf(3) := negative;
cf(2) := not negative and not zero;
cf(1) := zero;
cf(0) := xe.so;
c_out.write_cr_data(31 downto 28) <= cf;
end if;
end process;
end;