Add mcounteren register

doc/01_overview/compliance.rst

@@ -43,6 +43,10 @@ In addition, the following instruction set extensions are available.
      - 2.0
      - always enabled
 
+   * - **"Zicntr" and "Zihpm": Extensions for Counters and Performance Monitoring**
+     - 2.0
+     - always enabled
+
    * - **Zifencei**: Instruction-Fetch Fence
      - 2.0
      - always enabled

doc/03_reference/performance_counters.rst

@@ -3,10 +3,13 @@
 Performance Counters
 ====================
 
-Ibex implements performance counters according to the RISC-V Privileged Specification, version 1.11 (see Hardware Performance Monitor, Section 3.1.11).
+Ibex implements performance counters according to the RISC-V Privileged Specification, version 1.11 (see Hardware Performance Monitor, Section 3.1.11) and supports the **Zihpm** (Hardware Performance Counters) extension.
 The performance counters are placed inside the Control and Status Registers (CSRs) and can be accessed with the ``CSRRW(I)`` and ``CSRRS/C(I)`` instructions.
 
-Ibex implements the clock cycle counter ``mcycle(h)``, the retired instruction counter ``minstret(h)``, as well as the 29 event counters ``mhpmcounter3(h)`` - ``mhpmcounter31(h)`` and the corresponding event selector CSRs ``mhpmevent3`` - ``mhpmevent31``, and the ``mcountinhibit`` CSR to individually enable/disable the counters.
+Ibex implements the machine-mode clock cycle counter ``mcycle(h)``, the retired instruction counter ``minstret(h)``, as well as the 29 event counters ``mhpmcounter3(h)`` - ``mhpmcounter31(h)`` and the corresponding event selector CSRs ``mhpmevent3`` - ``mhpmevent31``, and the ``mcountinhibit`` CSR to individually enable/disable the counters.
+
+Additionally, Ibex implements the Zicntr and Zihpm extensions which provide User-mode (U-mode) aliases for these performance counters: ``cycle(h)``, ``instret(h)``, and ``hpmcounter3(h)`` - ``hpmcounter31(h)``. These aliases provide read-only access to the exact same underlying hardware counters configured in M-mode.
+
 ``mcycle(h)`` and ``minstret(h)`` are always available and 64 bit wide.
 The ``mhpmcounter`` performance counters are optional (unavailable by default) and parametrizable in width.
 
@@ -60,6 +63,19 @@ In particular, to enable/disable ``mcycle(h)``, bit 0 must be written. For ``min
 The lower 32 bits of all counters can be accessed through the base register, whereas the upper 32 bits are accessed through the ``h``-register.
 Reads to all these registers are non-destructive.
 
+User-Mode Counter Access (mcounteren)
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Access to the U-mode counter aliases (``cycle(h)``, ``instret(h)``, and ``hpmcounterX(h)``) is controlled via the Machine Counter-Enable CSR (``mcounteren``). This register can gate access to the counters from less privileged modes to prevent benchmarking the core if desired.
+
+* **Bit 0** controls access to ``cycle(h)``.
+* **Bit 2** controls access to ``instret(h)``.
+* **Bit X** controls access to ``hpmcounterX(h)``.
+
+When a bit in ``mcounteren`` is clear (0), any attempt to read the corresponding counter alias from U-mode will trigger an illegal instruction exception.
+
+To secure this mechanism, the ``mcounteren`` register can be locked against software modifications using a MUBI input signal called ``mcounteren_writeable``. When this signal disables writes, any attempt by software to modify the contents of ``mcounteren`` is ignored.
+
 Parametrization at synthesis time
 ---------------------------------
 

dv/cosim/cosim.h

@@ -128,6 +128,17 @@ class Cosim {
   // A full 64-bit value is provided setting both the mcycle and mcycleh CSRs.
   virtual void set_mcycle(uint64_t mcycle) = 0;
 
+  // Set the value of minstret.
+  //
+  // The co-simulation model doesn't alter the value of minstret itself (other
+  // than instructions that do a direct CSR write). minstret should be set to
+  // the correct value before any `step` call that may execute an instruction
+  // that observes the value of minstret.
+  //
+  // A full 64-bit value is provided setting both the minstret and minstreth
+  // CSRs.
+  virtual void set_minstret(uint64_t minstret) = 0;
+
   // Set the value of a CSR. This is used when it is needed to have direct
   // communication between DUT and Spike (e.g. Performance counters).
   virtual void set_csr(const int csr_num, const uint32_t new_val) = 0;

dv/cosim/cosim_dpi.cc

@@ -52,6 +52,13 @@ void riscv_cosim_set_mcycle(Cosim *cosim, svBitVecVal *mcycle) {
   cosim->set_mcycle(mcycle_full);
 }
 
+void riscv_cosim_set_minstret(Cosim *cosim, svBitVecVal *minstret) {
+  assert(cosim);
+
+  uint64_t minstret_full = minstret[0] | (uint64_t)minstret[1] << 32;
+  cosim->set_minstret(minstret_full);
+}
+
 void riscv_cosim_set_csr(Cosim *cosim, const int csr_id,
                          const svBitVecVal *csr_val) {
   assert(cosim);

dv/cosim/cosim_dpi.h

@@ -23,6 +23,7 @@ void riscv_cosim_set_nmi(Cosim *cosim, svBit nmi);
 void riscv_cosim_set_nmi_int(Cosim *cosim, svBit nmi_int);
 void riscv_cosim_set_debug_req(Cosim *cosim, svBit debug_req);
 void riscv_cosim_set_mcycle(Cosim *cosim, svBitVecVal *mcycle);
+void riscv_cosim_set_minstret(Cosim *cosim, svBitVecVal *minstret);
 void riscv_cosim_set_csr(Cosim *cosim, const int csr_id,
                          const svBitVecVal *csr_val);
 void riscv_cosim_set_ic_scr_key_valid(Cosim *cosim, svBit valid);

dv/cosim/cosim_dpi.svh

@@ -18,6 +18,7 @@ import "DPI-C" function void riscv_cosim_set_nmi(chandle cosim_handle, bit nmi);
 import "DPI-C" function void riscv_cosim_set_nmi_int(chandle cosim_handle, bit nmi_int);
 import "DPI-C" function void riscv_cosim_set_debug_req(chandle cosim_handle, bit debug_req);
 import "DPI-C" function void riscv_cosim_set_mcycle(chandle cosim_handle, bit [63:0] mcycle);
+import "DPI-C" function void riscv_cosim_set_minstret(chandle cosim_handle, bit [63:0] minstret);
 import "DPI-C" function void riscv_cosim_set_csr(chandle cosim_handle, int csr_id,
   bit [31:0] csr_val);
 import "DPI-C" function void riscv_cosim_set_ic_scr_key_valid(chandle cosim_handle, bit valid);

dv/cosim/spike_cosim.cc

@@ -39,7 +39,10 @@ SpikeCosim::SpikeCosim(const std::string &isa_string, uint32_t start_pc,
                        uint32_t pmp_num_regions, uint32_t pmp_granularity,
                        uint32_t mhpm_counter_num, uint32_t dm_start_addr,
                        uint32_t dm_end_addr)
-    : nmi_mode(false), pending_iside_error(false), insn_cnt(0) {
+    : nmi_mode(false),
+      pending_iside_error(false),
+      insn_cnt(0),
+      mhpm_counter_num(mhpm_counter_num) {
   FILE *log_file = nullptr;
   if (trace_log_path.length() != 0) {
     log = std::make_unique<log_file_t>(trace_log_path.c_str());
@@ -747,6 +750,27 @@ void SpikeCosim::set_mcycle(uint64_t mcycle) {
   // to write all 64 bits at once at least.
 }
 
+void SpikeCosim::set_minstret(uint64_t minstret) {
+  uint32_t upper_minstret = minstret >> 32;
+  uint32_t lower_minstret = minstret & 0xffffffff;
+
+  // Spike decrements the MINSTRET CSR when you write to it. This is the same
+  // issue as with MCYCLE. See `set_mcycle` for more details.
+
+  // Write the lower half first, incremented twice due to the double decrement
+  processor->get_state()->csrmap[CSR_MINSTRET]->write(lower_minstret + 2);
+
+  if ((processor->get_state()->csrmap[CSR_MINSTRET]->read() & 0xffffffff) ==
+      0) {
+    // If the lower half is 0 at this point then the upper half will get
+    // decremented, so increment it first.
+    upper_minstret++;
+  }
+
+  // Set the upper half
+  processor->get_state()->csrmap[CSR_MINSTRETH]->write(upper_minstret);
+}
+
 void SpikeCosim::set_csr(const int csr_num, const uint32_t new_val) {
   // Note that this is tested with ibex-cosim-v0.3 version of Spike. 'set_csr'
   // method might have a hardwired zero for mhpmcounterX registers.
@@ -832,6 +856,19 @@ void SpikeCosim::fixup_csr(int csr_num, uint32_t csr_val) {
       processor->set_csr(csr_num, new_val);
 #else
       processor->put_csr(csr_num, new_val);
+#endif
+      break;
+    }
+    case CSR_MCOUNTEREN: {
+      // Bits 3..3+mhpm_counter_num-1 correspond to implemented HPM counters
+      reg_t hpm_mask = ((1 << mhpm_counter_num) - 1) << 3;
+      // Bit 0 and 2 are for mcycle and minstret which are always implemented
+      // Bit 1 is for time which is not implemented, hence the mask 0x5
+      reg_t new_val = csr_val & (0x5 | hpm_mask);
+#ifdef OLD_SPIKE
+      processor->set_csr(csr_num, new_val);
+#else
+      processor->put_csr(csr_num, new_val);
 #endif
       break;
     }

dv/cosim/spike_cosim.h

@@ -98,6 +98,7 @@ class SpikeCosim : public simif_t, public Cosim {
   void misaligned_pmp_fixup();
 
   unsigned int insn_cnt;
+  uint32_t mhpm_counter_num;
 
  public:
   SpikeCosim(const std::string &isa_string, uint32_t start_pc,
@@ -131,6 +132,7 @@ class SpikeCosim : public simif_t, public Cosim {
   void set_nmi_int(bool nmi_int) override;
   void set_debug_req(bool debug_req) override;
   void set_mcycle(uint64_t mcycle) override;
+  void set_minstret(uint64_t minstret) override;
   void set_csr(const int csr_num, const uint32_t new_val) override;
   void set_ic_scr_key_valid(bool valid) override;
   void notify_dside_access(const DSideAccessInfo &access_info) override;

dv/formal/check/top.sv

@@ -103,6 +103,7 @@ module top import ibex_pkg::*; #(
 
   // CPU Control Signals
   input  ibex_mubi_t                                                   fetch_enable_i,
+  input  ibex_mubi_t                                                   mcounteren_writable_i,
   output logic                                                         core_sleep_o,
   output logic                                                         alert_minor_o,
   output logic                                                         alert_major_internal_o,
@@ -162,7 +163,9 @@ NotDebug: assume property (!ibex_top_i.u_ibex_core.debug_mode & !debug_req_i);
 ConstantBoot: assume property (boot_addr_i == $past(boot_addr_i));
 // 3. Always fetch enable
 FetchEnable: assume property (fetch_enable_i == IbexMuBiOn);
-// 4. Never try to sleep if we couldn't ever wake up
+// 4. Always have mcounteren writable
+McounterenWritable: assume property (mcounteren_writable_i == IbexMuBiOn);
+// 5. Never try to sleep if we couldn't ever wake up
 WFIStart: assume property (`IDC.ctrl_fsm_cs == SLEEP |-> (
                                                           `CSR.mie_q.irq_software |
                                                           `CSR.mie_q.irq_timer |
@@ -200,6 +203,7 @@ WFIStart: assume property (`IDC.ctrl_fsm_cs == SLEEP |-> (
     logic [31:0] `X(mscratch); \
     logic [31:0] `X(mepc); \
     logic [63:0] `X(mcycle); \
+    logic [63:0] `X(cycle); \
     logic [31:0] `X(mshwmb); \
     logic [31:0] `X(mshwm); \
     logic [31:0] `X(mcounteren); \
@@ -441,13 +445,17 @@ logic ex_is_checkable_csr;
 assign ex_is_checkable_csr = ~(
   ((CSR_MHPMCOUNTER3H <= `CSR_ADDR) && (`CSR_ADDR <= CSR_MHPMCOUNTER31H)) |
   ((CSR_MHPMCOUNTER3 <= `CSR_ADDR) && (`CSR_ADDR <= CSR_MHPMCOUNTER31)) |
+  ((CSR_HPMCOUNTER3H <= `CSR_ADDR) && (`CSR_ADDR <= CSR_HPMCOUNTER31H)) |
+  ((CSR_HPMCOUNTER3 <= `CSR_ADDR) && (`CSR_ADDR <= CSR_HPMCOUNTER31)) |
   ((CSR_MHPMEVENT3 <= `CSR_ADDR) && (`CSR_ADDR <= CSR_MHPMEVENT31)) |
   (`CSR_ADDR == CSR_CPUCTRLSTS) | (`CSR_ADDR == CSR_SECURESEED) |
   (`CSR_ADDR == CSR_MIE) |
   (`CSR_ADDR == CSR_MCYCLE) | (`CSR_ADDR == CSR_MCYCLEH) |
+  (`CSR_ADDR == CSR_CYCLE) | (`CSR_ADDR == CSR_CYCLEH) |
 
   // TODO:
   (`CSR_ADDR == CSR_MINSTRET) | (`CSR_ADDR == CSR_MINSTRETH) |
+  (`CSR_ADDR == CSR_INSTRET) | (`CSR_ADDR == CSR_INSTRETH) |
   (`CSR_ADDR == CSR_MCOUNTINHIBIT)
 );
 

dv/riscv_compliance/rtl/ibex_riscv_compliance.sv

@@ -216,6 +216,7 @@ module ibex_riscv_compliance (
       .double_fault_seen_o       (                     ),
 
       .fetch_enable_i            (ibex_pkg::IbexMuBiOn ),
+      .mcounteren_writable_i     (ibex_pkg::IbexMuBiOn ),
       .alert_minor_o             (                     ),
       .alert_major_internal_o    (                     ),
       .alert_major_bus_o         (                     ),

dv/uvm/core_ibex/common/ibex_cosim_agent/ibex_cosim_scoreboard.sv

@@ -151,6 +151,7 @@ class ibex_cosim_scoreboard extends uvm_scoreboard;
       riscv_cosim_set_nmi_int(cosim_handle, rvfi_instr.nmi_int);
       riscv_cosim_set_mip(cosim_handle, rvfi_instr.pre_mip, rvfi_instr.post_mip);
       riscv_cosim_set_mcycle(cosim_handle, rvfi_instr.mcycle);
+      riscv_cosim_set_minstret(cosim_handle, rvfi_instr.minstret);
 
       // Set performance counters through a pseudo-backdoor write
       for (int i=0; i < 10; i++) begin

dv/uvm/core_ibex/common/ibex_cosim_agent/ibex_rvfi_monitor.sv

@@ -44,6 +44,7 @@ class ibex_rvfi_monitor extends uvm_monitor;
       trans_collected.debug_req        = vif.monitor_cb.ext_debug_req;
       trans_collected.rf_wr_suppress   = vif.monitor_cb.ext_rf_wr_suppress;
       trans_collected.mcycle           = vif.monitor_cb.ext_mcycle;
+      trans_collected.minstret         = vif.monitor_cb.ext_minstret;
       trans_collected.ic_scr_key_valid = vif.monitor_cb.ext_ic_scr_key_valid;
 
       for (int i=0; i < 10; i++) begin

dv/uvm/core_ibex/common/ibex_cosim_agent/ibex_rvfi_seq_item.sv

@@ -16,6 +16,7 @@ class ibex_rvfi_seq_item extends uvm_sequence_item;
   bit        debug_req;
   bit        rf_wr_suppress;
   bit [63:0] mcycle;
+  bit [63:0] minstret;
 
   bit [31:0] mhpmcounters  [10];
   bit [31:0] mhpmcountersh [10];
@@ -34,6 +35,7 @@ class ibex_rvfi_seq_item extends uvm_sequence_item;
     `uvm_field_int (debug_req, UVM_DEFAULT)
     `uvm_field_int (rf_wr_suppress, UVM_DEFAULT)
     `uvm_field_int (mcycle, UVM_DEFAULT)
+    `uvm_field_int (minstret, UVM_DEFAULT)
     `uvm_field_sarray_int (mhpmcounters, UVM_DEFAULT)
     `uvm_field_sarray_int (mhpmcountersh, UVM_DEFAULT)
     `uvm_field_int (ic_scr_key_valid, UVM_DEFAULT)