core: Crack update-form loads into two internal ops

This uses the instruction-doubling machinery to send load with update
instructions down to loadstore1 as two separate ops, rather than
one op with two destinations.  This will help to simplify the value
tracking mechanisms.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>

diff --git a/common.vhdl b/common.vhdl
index 7bf827703dfc6ea9009eda9cbd3caea4a3c8d52b..893127f78d5d26462f9414ddf413c22669fa2e86 100644 (file)
--- a/common.vhdl
+++ b/common.vhdl
@@ -333,7 +333,6 @@ package common is
        byte_reverse : std_ulogic;
        sign_extend : std_ulogic;                       -- do we need to sign extend?
        update : std_ulogic;                            -- is this an update instruction?
-       update_reg : gpr_index_t;                       -- if so, the register to update
        xerc : xer_common_t;
         reserve : std_ulogic;                           -- set for larx/stcx.
         rc : std_ulogic;                                -- set for stcx.
@@ -351,7 +350,7 @@ package common is
                                                                      addr1 => (others => '0'), addr2 => (others => '0'), data => (others => '0'),
                                                                      write_reg => (others => '0'), length => (others => '0'),
                                                                      mode_32bit => '0', is_32bit => '0',
-                                                                     repeat => '0', second => '0', others => (others => '0'));
+                                                                     repeat => '0', second => '0');
 
     type Loadstore1ToExecute1Type is record
         busy : std_ulogic;

diff --git a/decode1.vhdl b/decode1.vhdl
index ebe59befcbc9c5a37aa129a4801f1cc6dcf13d2e..0f3410d544f9dcbd925787a794fb7d4907a7e682 100644 (file)
--- a/decode1.vhdl
+++ b/decode1.vhdl
@@ -84,18 +84,18 @@ architecture behaviour of decode1 is
         11 =>       (ALU,  NONE, OP_CMP,       RA,         CONST_SI,    NONE, NONE, '0', '1', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '1', NONE, '0', '0', NONE), -- cmpi
         10 =>       (ALU,  NONE, OP_CMP,       RA,         CONST_UI,    NONE, NONE, '0', '1', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- cmpli
         34 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lbz
-        35 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lbzu
+        35 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lbzu
         50 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lfd
-        51 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lfdu
+        51 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lfdu
         48 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0', NONE), -- lfs
-        49 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- lfsu
+        49 =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', DUPD), -- lfsu
         42 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lha
-        43 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lhau
+        43 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lhau
         40 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lhz
-        41 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lhzu
+        41 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lhzu
         56 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_DQ,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', DRTE), -- lq
         32 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwz
-        33 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lwzu
+        33 =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lwzu
          7 =>       (ALU,  NONE, OP_MUL_L64,   RA,         CONST_SI,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '1', NONE, '0', '0', NONE), -- mulli
         24 =>       (ALU,  NONE, OP_OR,        NONE,       CONST_UI,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- ori
         25 =>       (ALU,  NONE, OP_OR,        NONE,       CONST_UI_HI, RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- oris
@@ -104,15 +104,15 @@ architecture behaviour of decode1 is
         23 =>       (ALU,  NONE, OP_RLC,       NONE,       RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '0', RC,   '0', '0', NONE), -- rlwnm
         17 =>       (ALU,  NONE, OP_SC,        NONE,       NONE,        NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- sc
         38 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stb
-        39 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stbu
+        39 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stbu
         54 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stfd
-        55 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stfdu
+        55 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  RA,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stfdu
         52 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0', NONE), -- stfs
-        53 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- stfsu
+        53 =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, CONST_SI,    FRS,  RA,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- stfsu
         44 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- sth
-        45 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- sthu
+        45 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- sthu
         36 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stw
-        37 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stwu
+        37 =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_SI,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stwu
          8 =>       (ALU,  NONE, OP_ADD,       RA,         CONST_SI,    NONE, RT,   '0', '0', '1', '0', ONE,  '1', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- subfic
          2 =>       (ALU,  NONE, OP_TRAP,      RA,         CONST_SI,    NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '1', NONE), -- tdi
          3 =>       (ALU,  NONE, OP_TRAP,      RA,         CONST_SI,    NONE, NONE, '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '0', NONE, '0', '1', NONE), -- twi
@@ -290,33 +290,33 @@ architecture behaviour of decode1 is
         2#1111101111#  =>       (ALU,  NONE, OP_ISEL,      RA_OR_ZERO, RB,          NONE, RT,   '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- isel
         2#0000110100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '1', '0', '0', NONE, '0', '0', NONE), -- lbarx
         2#1101010101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lbzcix
-        2#0001110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lbzux
+        2#0001110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lbzux
         2#0001010111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lbzx
         2#0001010100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '1', '0', '0', NONE, '0', '0', NONE), -- ldarx
         2#1000010100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- ldbrx
         2#1101110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- ldcix
-        2#0000110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- ldux
+        2#0000110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- ldux
         2#0000010101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- ldx
         2#1001010111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lfdx
-        2#1001110111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lfdux
+        2#1001110111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lfdux
         2#1101010111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lfiwax
         2#1101110111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lfiwzx
         2#1000010111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0', NONE), -- lfsx
-        2#1000110111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- lfsux
+        2#1000110111#  =>       (LDST, FPU,  OP_LOAD,      RA_OR_ZERO, RB,          NONE, FRT,  '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', DUPD), -- lfsux
         2#0001110100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '1', '0', '0', NONE, '0', '0', NONE), -- lharx
-        2#0101110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lhaux
+        2#0101110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lhaux
         2#0101010111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '1', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lhax
         2#1100010110#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lhbrx
         2#1100110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lhzcix
-        2#0100110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lhzux
+        2#0100110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lhzux
         2#0100010111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lhzx
         2#0100010100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '1', '0', '0', NONE, '0', '0', DRTE), -- lqarx
         2#0000010100#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '1', '0', '0', NONE, '0', '0', NONE), -- lwarx
-        2#0101110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lwaux
+        2#0101110101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lwaux
         2#0101010101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwax
         2#1000010110#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwbrx
         2#1100010101#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwzcix
-        2#0000110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- lwzux
+        2#0000110111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- lwzux
         2#0000010111#  =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, RB,          NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwzx
         2#1001000000#  =>       (ALU,  NONE, OP_MCRXRX,    NONE,       NONE,        NONE, NONE, '0', '1', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- mcrxrx
         2#0000010011#  =>       (ALU,  NONE, OP_MFCR,      NONE,       NONE,        NONE, RT,   '1', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- mfcr/mfocrf
@@ -376,28 +376,28 @@ architecture behaviour of decode1 is
         2#1000011000#  =>       (ALU,  NONE, OP_SHR,       NONE,       RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', NONE, '0', '0', '0', '0', '1', '0', RC,   '0', '0', NONE), -- srw
         2#1111010101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stbcix
         2#1010110110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0', NONE), -- stbcx
-        2#0011110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stbux
+        2#0011110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stbux
         2#0011010111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is1B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stbx
         2#1010010100#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stdbrx
         2#1111110101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stdcix
         2#0011010110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0', NONE), -- stdcx
-        2#0010110101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stdux
+        2#0010110101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stdux
         2#0010010101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stdx
         2#1011010111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stfdx
-        2#1011110111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stfdux
+        2#1011110111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  RA,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stfdux
         2#1111010111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stfiwx
         2#1010010111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '1', '0', NONE, '0', '0', NONE), -- stfsx
-        2#1010110111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- stfsux
+        2#1010110111#  =>       (LDST, FPU,  OP_STORE,     RA_OR_ZERO, RB,          FRS,  RA,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '1', '0', NONE, '0', '0', NONE), -- stfsux
         2#1110010110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- sthbrx
         2#1110110101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- sthcix
         2#1011010110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0', NONE), -- sthcx
-        2#0110110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- sthux
+        2#0110110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- sthux
         2#0110010111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is2B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- sthx
         2#0010110110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0', DRSE), -- stqcx
         2#1010010110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '1', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stwbrx
         2#1110010101#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stwcix
         2#0010010110#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '1', '0', '0', ONE,  '0', '0', NONE), -- stwcx
-        2#0010110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stwux
+        2#0010110111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   RA,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stwux
         2#0010010111#  =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, RB,          RS,   NONE, '0', '0', '0', '0', ZERO, '0', is4B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- stwx
         2#0000101000#  =>       (ALU,  NONE, OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0', NONE), -- subf
         2#1000101000#  =>       (ALU,  NONE, OP_ADD,       RA,         RB,          NONE, RT,   '0', '0', '1', '0', ONE,  '0', NONE, '0', '0', '0', '0', '0', '0', RC,   '0', '0', NONE), -- subfo
@@ -424,7 +424,7 @@ architecture behaviour of decode1 is
         --              unit  fac   internal      in1         in2          in3   out   CR   CR   inv  inv  cry   cry  ldst  BR   sgn  upd  rsrv 32b  sgn  rc    lk   sgl  rpt
         --                               op                                            in   out   A   out  in    out  len        ext                                 pipe
         0     =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_DS,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- ld
-        1     =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_DS,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- ldu
+        1     =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_DS,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', DUPD), -- ldu
         2     =>       (LDST, NONE, OP_LOAD,      RA_OR_ZERO, CONST_DS,    NONE, RT,   '0', '0', '0', '0', ZERO, '0', is4B, '0', '1', '0', '0', '0', '0', NONE, '0', '0', NONE), -- lwa
         others   => decode_rom_init
         );
@@ -451,7 +451,7 @@ architecture behaviour of decode1 is
         --              unit  fac   internal      in1         in2          in3   out   CR   CR   inv  inv  cry   cry  ldst  BR   sgn  upd  rsrv 32b  sgn  rc    lk   sgl  rpt
         --                                op                                           in   out   A   out  in    out  len        ext                                 pipe
         0     =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_DS,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', NONE), -- std
-        1     =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_DS,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stdu
+        1     =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_DS,    RS,   RA,   '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '1', '0', '0', '0', NONE, '0', '0', NONE), -- stdu
         2     =>       (LDST, NONE, OP_STORE,     RA_OR_ZERO, CONST_DS,    RS,   NONE, '0', '0', '0', '0', ZERO, '0', is8B, '0', '0', '0', '0', '0', '0', NONE, '0', '0', DRSE), -- stq
         others   => decode_rom_init
         );

diff --git a/decode2.vhdl b/decode2.vhdl
index e00a05de91ab273aa8b0ea84af098ccc0f23758e..03360575324de4a77e21e22021e243e0c1d5e411 100644 (file)
--- a/decode2.vhdl
+++ b/decode2.vhdl
@@ -407,6 +407,11 @@ begin
                     if r.repeat = d_in.big_endian then
                         decoded_reg_o.reg(0) := '1';
                     end if;
+                when DUPD =>
+                    -- update-form loads, 2nd instruction writes RA
+                    if r.repeat = '1' then
+                        decoded_reg_o.reg := decoded_reg_a.reg;
+                    end if;
                 when others =>
             end case;
         end if;
@@ -492,13 +497,9 @@ begin
         if EX1_BYPASS and d_in.decode.unit = ALU then
             gpr_bypassable <= '1';
         end if;
-        update_gpr_write_valid <= d_in.decode.update;
-        update_gpr_write_reg <= decoded_reg_a.reg;
-        if v.e.lr = '1' then
-            -- there are no instructions that have both update=1 and lr=1
-            update_gpr_write_valid <= '1';
-            update_gpr_write_reg <= fast_spr_num(SPR_LR);
-        end if;
+
+        update_gpr_write_valid <= v.e.lr;
+        update_gpr_write_reg <= fast_spr_num(SPR_LR);
 
         gpr_a_read_valid <= decoded_reg_a.reg_valid;
         gpr_a_read <= decoded_reg_a.reg;

diff --git a/decode_types.vhdl b/decode_types.vhdl
index 02790a61fe560d8b73422357ea7899dd42f277d7..885cc91d79569b79b97c7123dd486c7481a05bca 100644 (file)
--- a/decode_types.vhdl
+++ b/decode_types.vhdl
@@ -54,7 +54,8 @@ package decode_types is
 
     type repeat_t is (NONE,      -- instruction is not repeated
                       DRSE,      -- double RS, endian twist
-                      DRTE);     -- double RT, endian twist
+                      DRTE,      -- double RT, endian twist
+                      DUPD);     -- update-form load
 
     type decode_rom_t is record
        unit         : unit_t;

diff --git a/execute1.vhdl b/execute1.vhdl
index 25b1dc7dcc3f2f2081850fd6c33870f6e7d70101..559f34f1923dc1f262d7a21301a225005b631d5e 100644 (file)
--- a/execute1.vhdl
+++ b/execute1.vhdl
@@ -1344,7 +1344,6 @@ begin
         lv.byte_reverse := e_in.byte_reverse xnor ctrl.msr(MSR_LE);
         lv.sign_extend := e_in.sign_extend;
         lv.update := e_in.update;
-        lv.update_reg := gspr_to_gpr(e_in.read_reg1);
         lv.xerc := v.e.xerc;
         lv.reserve := e_in.reserve;
         lv.rc := e_in.rc;

diff --git a/loadstore1.vhdl b/loadstore1.vhdl
index b83eed62a0d826d3adead55fde4e4a0a363b5ab1..fc5bcf9cfca85a5caf579fd0533cd7a7112183f4 100644 (file)
--- a/loadstore1.vhdl
+++ b/loadstore1.vhdl
@@ -72,7 +72,6 @@ architecture behave of loadstore1 is
         brev_mask    : unsigned(2 downto 0);
        sign_extend  : std_ulogic;
        update       : std_ulogic;
-       update_reg   : gpr_index_t;
        xerc         : xer_common_t;
         reserve      : std_ulogic;
         atomic       : std_ulogic;
@@ -415,7 +414,7 @@ begin
         when ACK_WAIT =>
             -- r.wr_sel gets set one cycle after we come into ACK_WAIT state,
             -- which is OK because the dcache always takes at least two cycles.
-            if r.update = '1' and (r.load = '0' or (HAS_FPU and r.load_sp = '1')) then
+            if r.update = '1' and r.load = '0' then
                 v.wr_sel := "01";
             end if;
             if d_in.error = '1' then
@@ -446,16 +445,9 @@ begin
                     write_enable := r.load and not r.load_sp;
                     if HAS_FPU and r.load_sp = '1' then
                         -- SP to DP conversion takes a cycle
-                        -- Write back rA update in this cycle if needed
-                        do_update := r.update;
                         v.wr_sel := "10";
                         v.state := FINISH_LFS;
-                    elsif r.extra_cycle = '1' then
-                        -- loads with rA update need an extra cycle
-                        v.wr_sel := "01";
-                        v.state := COMPLETE;
-                        v.do_update := r.update;
-                    else
+                    elsif r.load = '0' then
                         -- stores write back rA update in this cycle
                         do_update := r.update;
                     end if;
@@ -516,7 +508,6 @@ begin
             v.byte_reverse := l_in.byte_reverse;
             v.sign_extend := l_in.sign_extend;
             v.update := l_in.update;
-            v.update_reg := l_in.update_reg;
             v.xerc := l_in.xerc;
             v.reserve := l_in.reserve;
             v.rc := l_in.rc;
@@ -526,7 +517,6 @@ begin
             v.load_sp := '0';
             v.wait_dcache := '0';
             v.wait_mmu := '0';
-            v.do_update := '0';
             v.extra_cycle := '0';
 
             if HAS_FPU and l_in.is_32bit = '1' then
@@ -537,13 +527,21 @@ begin
 
             addr := lsu_sum;
             if l_in.second = '1' then
-                -- for the second half of a 16-byte transfer, use next_addr
-                addr := next_addr;
+                -- second half of load with update does the update
+                if l_in.op = OP_LOAD and l_in.update = '1' then
+                    v.do_update := '1';
+                else
+                    -- for the second half of a 16-byte transfer, use next_addr
+                    addr := next_addr;
+                end if;
             end if;
             if l_in.mode_32bit = '1' then
                 addr(63 downto 32) := (others => '0');
             end if;
-            v.addr := addr;
+            if v.do_update = '0' then
+                -- preserve previous r.addr for load with update
+                v.addr := addr;
+            end if;
             maddr := l_in.addr2;    -- address from RB for tlbie
 
             -- XXX Temporary hack. Mark the op as non-cachable if the address
@@ -566,7 +564,7 @@ begin
             -- check alignment for larx/stcx
             misaligned := or (std_ulogic_vector(unsigned(l_in.length(2 downto 0)) - 1) and addr(2 downto 0));
             v.align_intr := l_in.reserve and misaligned;
-            if l_in.repeat = '1' and l_in.second = '0' and addr(3) = '1' then
+            if l_in.repeat = '1' and l_in.second = '0' and l_in.update = '0' and addr(3) = '1' then
                 -- length is really 16 not 8
                 -- Make misaligned lq cause an alignment interrupt in LE mode,
                 -- in order to avoid the case with RA = RT + 1 where the second half
@@ -585,14 +583,17 @@ begin
                 when OP_STORE =>
                     req := '1';
                 when OP_LOAD =>
-                    req := '1';
                     v.load := '1';
-                    -- Allow an extra cycle for RA update on loads
-                    v.extra_cycle := l_in.update;
-                    if HAS_FPU and l_in.is_32bit = '1' then
-                        -- Allow an extra cycle for SP->DP precision conversion
-                        v.load_sp := '1';
-                        v.extra_cycle := '1';
+                    if l_in.second = '1' and l_in.update = '1' then
+                        v.wr_sel := "01";
+                        v.state := COMPLETE;
+                    else
+                        req := '1';
+                        if HAS_FPU and l_in.is_32bit = '1' then
+                            -- Allow an extra cycle for SP->DP precision conversion
+                            v.load_sp := '1';
+                            v.extra_cycle := '1';
+                        end if;
                     end if;
                 when OP_DCBZ =>
                     v.align_intr := v.nc;
@@ -724,22 +725,19 @@ begin
         -- Multiplex either cache data to the destination GPR or
         -- the address for the rA update.
         l_out.valid <= done;
+        l_out.write_reg <= r.write_reg;
         case r.wr_sel is
         when "00" =>
             l_out.write_enable <= '1';
-            l_out.write_reg <= r.write_reg;
             l_out.write_data <= r.sprval;
         when "01" =>
             l_out.write_enable <= do_update;
-            l_out.write_reg <= gpr_to_gspr(r.update_reg);
             l_out.write_data <= r.addr;
         when "10" =>
             l_out.write_enable <= '1';
-            l_out.write_reg <= r.write_reg;
             l_out.write_data <= load_dp_data;
         when others =>
             l_out.write_enable <= write_enable;
-            l_out.write_reg <= r.write_reg;
             l_out.write_data <= data_trimmed;
         end case;
         l_out.xerc <= r.xerc;