Merge pull request #440 from paulusmack/compliance

paulusmack · web-flow · commit c5abe3c0a9d3 · 2025-02-24T08:59:22.000+11:00
More compliance improvements - [H]DEXCR, no-op SPRs, writable TB
diff --git a/common.vhdl b/common.vhdl
@@ -39,6 +39,8 @@ package common is
     constant SPR_DAR    : spr_num_t := 19;
     constant SPR_TB     : spr_num_t := 268;
     constant SPR_TBU    : spr_num_t := 269;
+    constant SPR_TBLW   : spr_num_t := 284;
+    constant SPR_TBUW   : spr_num_t := 285;
     constant SPR_DEC    : spr_num_t := 22;
     constant SPR_SRR0   : spr_num_t := 26;
     constant SPR_SRR1   : spr_num_t := 27;
@@ -71,6 +73,14 @@ package common is
     constant SPR_DAWRX1 : spr_num_t := 189;
     constant SPR_HASHKEYR  : spr_num_t := 468;
     constant SPR_HASHPKEYR : spr_num_t := 469;
+    constant SPR_DEXCR  : spr_num_t := 828;
+    constant SPR_DEXCRU : spr_num_t := 812;
+    constant SPR_HDEXCR : spr_num_t := 471;
+    constant SPR_HDEXCU : spr_num_t := 455;
+    constant SPR_NOOP0  : spr_num_t := 808;
+    constant SPR_NOOP1  : spr_num_t := 809;
+    constant SPR_NOOP2  : spr_num_t := 810;
+    constant SPR_NOOP3  : spr_num_t := 811;
 
     -- PMU registers
     constant SPR_UPMC1  : spr_num_t := 771;
@@ -167,6 +177,7 @@ package common is
         ispmu : std_ulogic;
         ronly : std_ulogic;
         wonly : std_ulogic;
+        noop  : std_ulogic;
     end record;
     constant spr_id_init : spr_id := (sel => "0000", others => '0');
 
@@ -184,6 +195,7 @@ package common is
     constant SPRSEL_DSCR  : spr_selector := 4x"b";
     constant SPRSEL_PIR   : spr_selector := 4x"c";
     constant SPRSEL_CIABR : spr_selector := 4x"d";
+    constant SPRSEL_DEXCR : spr_selector := 4x"e";
     constant SPRSEL_XER   : spr_selector := 4x"f";
 
     -- FSCR and HFSCR bit numbers
@@ -266,6 +278,16 @@ package common is
         pri : std_ulogic_vector(31 downto 0);   -- 8 bits each for 4 cpus
     end record;
 
+    -- Bits in each half of DEXCR and HDEXCR
+    subtype aspect_bits_t is std_ulogic_vector(4 downto 0);
+    constant aspect_bits_init : aspect_bits_t := (others => '1');
+    -- Bit numbers in aspect_bits_t
+    constant DEXCR_SBHE   : integer := 4;       -- speculative branch hint enable
+    constant DEXCR_IBRTPD : integer := 3;       -- indirect branch recurrent target prediction disable
+    constant DEXCR_SRAPD  : integer := 2;       -- subroutine return address prediction disable
+    constant DEXCR_NPHIE  : integer := 1;       -- non-privileged hash instruction enable
+    constant DEXCR_PHIE   : integer := 0;       -- privileged hash instruction enable
+
     -- This needs to die...
     type ctrl_t is record
         wait_state: std_ulogic;
@@ -287,14 +309,26 @@ package common is
         heir: std_ulogic_vector(63 downto 0);
         dscr: std_ulogic_vector(24 downto 0);
         ciabr: std_ulogic_vector(63 downto 0);
+        dexcr_pnh: aspect_bits_t;
+        dexcr_pro: aspect_bits_t;
+        hdexcr_hyp: aspect_bits_t;
+        hdexcr_enf: aspect_bits_t;
     end record;
     constant ctrl_t_init : ctrl_t :=
         (wait_state => '0', run => '1', xer_low => 18x"0",
          fscr_ic => x"0", fscr_pref => '1', fscr_scv => '1', fscr_tar => '1', fscr_dscr => '1',
          hfscr_ic => x"0", hfscr_pref => '1', hfscr_tar => '1', hfscr_dscr => '1', hfscr_fp => '1',
          dscr => (others => '0'),
+         dexcr_pnh => aspect_bits_init, dexcr_pro => aspect_bits_init,
+         hdexcr_hyp => aspect_bits_init, hdexcr_enf => aspect_bits_init,
          others => (others => '0'));
 
+    type timebase_ctrl is record
+        reset   : std_ulogic;
+        rd_prot : std_ulogic;           -- read-protect => userspace can't read TB
+        freeze  : std_ulogic;
+    end record;
+
     type Fetch1ToIcacheType is record
 	req: std_ulogic;
         fetch_fail : std_ulogic;
@@ -604,6 +638,7 @@ package common is
         e2stall : std_ulogic;
         msr : std_ulogic_vector(63 downto 0);
         hashkey : std_ulogic_vector(63 downto 0);
+        hash_enable : std_ulogic;
     end record;
     constant Execute1ToLoadstore1Init : Execute1ToLoadstore1Type :=
         (valid => '0', op => OP_ILLEGAL, ci => '0', byte_reverse => '0',
@@ -616,7 +651,7 @@ package common is
          length => (others => '0'),
          mode_32bit => '0', is_32bit => '0', prefixed => '0',
          repeat => '0', second => '0', e2stall => '0',
-         msr => (others => '0'), hashkey => (others => '0'));
+         msr => (others => '0'), hashkey => (others => '0'), hash_enable => '0');
 
     type Loadstore1ToExecute1Type is record
         busy : std_ulogic;
diff --git a/core.vhdl b/core.vhdl
@@ -31,8 +31,8 @@ entity core is
 	-- Alternate reset (0xffff0000) for use by DRAM init fw
 	alt_reset    : in std_ulogic;
 
-        -- Global timebase
-        timebase     : in std_ulogic_vector(63 downto 0);
+        -- Global timebase control
+        tb_ctrl      : in timebase_ctrl;
 
 	-- Wishbone interface
         wishbone_insn_in  : in wishbone_slave_out;
@@ -309,6 +309,7 @@ begin
 	    busy_in => decode2_busy_in,
             stall_out => decode2_stall_out,
             flush_in => flush,
+            tb_ctrl => tb_ctrl,
             complete_in => complete,
 	    stopped_out => dbg_core_is_stopped,
             d_in => decode1_to_decode2,
@@ -376,7 +377,7 @@ begin
         port map (
             clk => clk,
             rst => rst_ex1,
-            timebase => timebase,
+            tb_ctrl => tb_ctrl,
             flush_in => flush,
 	    busy_out => ex1_busy_out,
             e_in => decode2_to_execute1,
diff --git a/decode1.vhdl b/decode1.vhdl
@@ -457,11 +457,20 @@ architecture behaviour of decode1 is
         i.ispmu := '0';
         i.ronly := '0';
         i.wonly := '0';
+        i.noop  := '0';
         case sprn is
             when SPR_TB =>
                 i.sel := SPRSEL_TB;
+                i.ronly := '1';
             when SPR_TBU =>
                 i.sel := SPRSEL_TBU;
+                i.ronly := '1';
+            when SPR_TBLW =>
+                i.sel := SPRSEL_TB;
+                i.wonly := '1';
+            when SPR_TBUW =>
+                i.sel := SPRSEL_TB;
+                i.wonly := '1';
             when SPR_DEC =>
                 i.sel := SPRSEL_DEC;
             when SPR_PVR =>
@@ -499,6 +508,13 @@ architecture behaviour of decode1 is
                 i.sel := SPRSEL_PIR;
             when SPR_CIABR =>
                 i.sel := SPRSEL_CIABR;
+            when SPR_DEXCR | SPR_HDEXCR =>
+                i.sel := SPRSEL_DEXCR;
+            when SPR_DEXCRU | SPR_HDEXCU =>
+                i.sel := SPRSEL_DEXCR;
+                i.ronly := '1';
+            when SPR_NOOP0 | SPR_NOOP1 | SPR_NOOP2 | SPR_NOOP3 =>
+                i.noop := '1';
             when others =>
                 i.valid := '0';
         end case;
diff --git a/decode2.vhdl b/decode2.vhdl
@@ -27,6 +27,8 @@ entity decode2 is
 
         flush_in: in std_ulogic;
 
+        tb_ctrl : timebase_ctrl;
+
         d_in  : in Decode1ToDecode2Type;
 
         e_out : out Decode2ToExecute1Type;
@@ -696,16 +698,23 @@ begin
             if op = OP_MFSPR then
                 if d_in.ram_spr.valid = '1' then
                     v.e.result_sel := "101";        -- ramspr_result
-                elsif d_in.spr_info.valid = '0' or d_in.spr_info.wonly = '1' then
+                elsif d_in.spr_info.valid = '0' or d_in.spr_info.wonly = '1' or
+                    d_in.spr_info.noop = '1' then
                     -- Privileged mfspr to invalid/unimplemented SPR numbers
                     -- writes the contents of RT back to RT (i.e. it's a no-op)
+                    -- as does any mfspr from the reserved/noop SPR numbers
                     v.e.result_sel := "001";        -- logical_result
                 end if;
             end if;
             v.e.privileged := d_in.decode.privileged;
             if (op = OP_MFSPR or op = OP_MTSPR) and d_in.insn(20) = '1' then
                 v.e.privileged := '1';
             end if;
+            -- Reading TB is privileged if syscon_tb_ctrl.rd_protect is 1
+            if tb_ctrl.rd_prot = '1' and op = OP_MFSPR and d_in.spr_info.valid = '1' and
+                 (d_in.spr_info.sel = SPRSEL_TB or d_in.spr_info.sel = SPRSEL_TBU) then
+                v.e.privileged := '1';
+            end if;
             v.e.prefixed := d_in.prefixed;
             v.e.prefix := d_in.prefix;
             v.e.illegal_suffix := d_in.illegal_suffix;
diff --git a/execute1.vhdl b/execute1.vhdl
@@ -34,7 +34,7 @@ entity execute1 is
 	ext_irq_in : std_ulogic;
         interrupt_in : WritebackToExecute1Type;
 
-        timebase : std_ulogic_vector(63 downto 0);
+        tb_ctrl : timebase_ctrl;
 
 	-- asynchronous
         l_out : out Execute1ToLoadstore1Type;
@@ -101,6 +101,8 @@ architecture behaviour of execute1 is
         write_ciabr : std_ulogic;
         enter_wait : std_ulogic;
         scv_trap : std_ulogic;
+        write_tbl : std_ulogic;
+        write_tbu : std_ulogic;
     end record;
     constant side_effect_init : side_effect_type := (others => '0');
 
@@ -279,6 +281,10 @@ architecture behaviour of execute1 is
 
     signal stage2_stall : std_ulogic;
 
+    signal timebase : std_ulogic_vector(63 downto 0);
+    signal tb_next  : std_ulogic_vector(63 downto 0);
+    signal tb_carry : std_ulogic;
+
     type privilege_level is (USER, SUPER);
     type op_privilege_array is array(insn_type_t) of privilege_level;
     constant op_privilege: op_privilege_array := (
@@ -425,6 +431,32 @@ architecture behaviour of execute1 is
         return ret;
     end;
 
+    -- return contents of DEXCR or HDEXCR
+    -- top 32 bits are zeroed for access via non-privileged number
+    function assemble_dexcr(c: ctrl_t; insn: std_ulogic_vector(31 downto 0)) return std_ulogic_vector is
+        variable ret : std_ulogic_vector(63 downto 0);
+        variable spr : std_ulogic_vector(9 downto 0);
+        variable dexh, dexl : aspect_bits_t;
+    begin
+        ret := (others => '0');
+        spr := insn(15 downto 11) & insn(20 downto 16);
+        if spr(9) = '1' then
+            dexh := c.dexcr_pnh;
+            dexl := c.dexcr_pro;
+        else
+            dexh := c.hdexcr_hyp;
+            dexl := c.hdexcr_enf;
+        end if;
+        if spr(4) = '0' then
+            dexl := (others => '0');
+        end if;
+        ret := dexh(DEXCR_SBHE) & "00" & dexh(DEXCR_IBRTPD) & dexh(DEXCR_SRAPD) &
+               dexh(DEXCR_NPHIE) & dexh(DEXCR_PHIE) & 25x"0" &
+               dexl(DEXCR_SBHE) & "00" & dexl(DEXCR_IBRTPD) & dexl(DEXCR_SRAPD) &
+               dexl(DEXCR_NPHIE) & dexl(DEXCR_PHIE) & 25x"0";
+        return ret;
+    end;
+
     -- Tell vivado to keep the hierarchy for the random module so that the
     -- net names in the xdc file match.
     attribute keep_hierarchy : string;
@@ -527,6 +559,43 @@ begin
             p_out => pmu_to_x
             );
 
+    -- Timebase just increments at the system clock frequency.
+    -- Ideally it would (appear to) run at 512MHz like IBM POWER systems,
+    -- but Linux seems to cope OK with it being 100MHz or whatever.
+    tbase: process(clk)
+    begin
+        if rising_edge(clk) then
+            if tb_ctrl.reset = '1' then
+                timebase <= (others => '0');
+                tb_carry <= '0';
+            else
+                timebase <= tb_next;
+                tb_carry <= and(tb_next(31 downto 0));
+            end if;
+        end if;
+    end process;
+
+    tbase_comb: process(all)
+        variable thi, tlo : std_ulogic_vector(31 downto 0);
+        variable carry : std_ulogic;
+    begin
+        tlo := timebase(31 downto 0);
+        thi := timebase(63 downto 32);
+        carry := '0';
+        if stage2_stall = '0' and ex1.se.write_tbl = '1' then
+            tlo := ex1.e.write_data(31 downto 0);
+        elsif tb_ctrl.freeze = '0' then
+            tlo := std_ulogic_vector(unsigned(tlo) + 1);
+            carry := tb_carry;
+        end if;
+        if stage2_stall = '0' and ex1.se.write_tbu = '1' then
+            thi := ex1.e.write_data(31 downto 0);
+        else
+            thi := std_ulogic_vector(unsigned(thi) + carry);
+        end if;
+        tb_next <= thi & tlo;
+    end process;
+
     dbg_ctrl_out <= ctrl;
     log_rd_addr <= ex2.log_addr_spr;
 
@@ -1306,7 +1375,7 @@ begin
             when OP_DARN =>
 	    when OP_MFMSR =>
 	    when OP_MFSPR =>
-		if e_in.spr_is_ram = '1' then
+		if e_in.spr_is_ram = '1' or e_in.spr_select.noop = '1' then
                     if e_in.valid = '1' and not is_X(e_in.insn) then
                         report "MFSPR to SPR " & integer'image(decode_spr_num(e_in.insn)) &
                             "=" & to_hstring(alu_result);
@@ -1398,6 +1467,10 @@ begin
                             v.se.write_dscr := '1';
                         when SPRSEL_CIABR =>
                             v.se.write_ciabr := '1';
+                        when SPRSEL_TB =>
+                            v.se.write_tbl := '1';
+                        when SPRSEL_TBU =>
+                            v.se.write_tbu := '1';
                         when others =>
                     end case;
                 end if;
@@ -1600,6 +1673,7 @@ begin
         variable go : std_ulogic;
         variable bypass_valid : std_ulogic;
         variable is_scv : std_ulogic;
+        variable dex : aspect_bits_t;
     begin
 	v := ex1;
         if busy_out = '0' then
@@ -1735,6 +1809,13 @@ begin
         bperm_start <= go and actions.start_bperm;
         pmu_trace <= go and actions.do_trace;
 
+        -- evaluate DEXCR/HDEXCR bits that apply at present
+        if ex1.msr(MSR_PR) = '0' then
+            dex := ctrl.hdexcr_hyp;
+        else
+            dex := ctrl.dexcr_pro or ctrl.hdexcr_enf;
+        end if;
+
         if not HAS_FPU and ex1.div_in_progress = '1' then
             v.div_in_progress := not divider_to_x.valid;
             v.busy := not divider_to_x.valid;
@@ -1850,6 +1931,11 @@ begin
         lv.second := e_in.second;
         lv.e2stall := fp_in.f2stall;
         lv.hashkey := ramspr_odd;
+        if e_in.insn(7) = '0' then
+            lv.hash_enable := dex(DEXCR_PHIE);
+        else
+            lv.hash_enable := dex(DEXCR_NPHIE);
+        end if;
 
         -- Outputs to FPU
         fv.op := e_in.insn_type;
@@ -1897,6 +1983,7 @@ begin
         39x"0" & ctrl.dscr when SPRSEL_DSCR,
         56x"0" & std_ulogic_vector(to_unsigned(CPU_INDEX, 8)) when SPRSEL_PIR,
         ctrl.ciabr when SPRSEL_CIABR,
+        assemble_dexcr(ctrl, ex1.insn) when SPRSEL_DEXCR,
         assemble_xer(ex1.e.xerc, ctrl.xer_low) when others;
 
     stage2_stall <= l_in.l2stall or fp_in.f2stall;
diff --git a/include/microwatt_soc.h b/include/microwatt_soc.h
@@ -67,6 +67,9 @@
 #define   SYS_REG_GIT_IS_DIRTY			(1ull << 63)
 #define SYS_REG_CPU_CTRL		0x58
 #define   SYS_REG_CPU_CTRL_ENABLE		0xff
+#define SYS_REG_TB_CTRL			0x60
+#define   SYS_REG_TB_CTRL_FREEZE		0x01
+#define   SYS_REG_TB_CTRL_RD_PROTECT		0x02
 
 /*
  * Register definitions for the potato UART
diff --git a/loadstore1.vhdl b/loadstore1.vhdl
diff --git a/soc.vhdl b/soc.vhdl
diff --git a/syscon.vhdl b/syscon.vhdl
