fix instruction decode

gen_input/templates/interp/CORENAME.cpp.gtl

@@ -85,6 +85,7 @@ protected:
 
     inline const char *name(size_t index){return traits::reg_aliases.at(index);}
 
+    compile_func decode_inst(code_word_t instr) ;
     virt_addr_t execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit) override;
 
     // some compile time constants
@@ -98,46 +99,13 @@ protected:
     std::array<compile_func, LUT_SIZE_C> lut_00, lut_01, lut_10;
     std::array<compile_func, LUT_SIZE> lut_11;
 
-    std::array<compile_func *, 4> qlut;
+    struct instruction_pattern {
+        uint32_t value;
+        uint32_t mask;
+        compile_func opc;
+    };
 
-    std::array<const uint32_t, 4> lutmasks = {{EXTR_MASK16, EXTR_MASK16, EXTR_MASK16, EXTR_MASK32}};
-
-    void expand_bit_mask(int pos, uint32_t mask, uint32_t value, uint32_t valid, uint32_t idx, compile_func lut[],
-                         compile_func f) {
-        if (pos < 0) {
-            lut[idx] = f;
-        } else {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                expand_bit_mask(pos - 1, mask, value, valid, idx, lut, f);
-            } else {
-                if ((valid & bitmask) == 0) {
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1), lut, f);
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1) + 1, lut, f);
-                } else {
-                    auto new_val = idx << 1;
-                    if ((value & bitmask) != 0) new_val++;
-                    expand_bit_mask(pos - 1, mask, value, valid, new_val, lut, f);
-                }
-            }
-        }
-    }
-
-    inline uint32_t extract_fields(uint32_t val) { return extract_fields(29, val >> 2, lutmasks[val & 0x3], 0); }
-
-    uint32_t extract_fields(int pos, uint32_t val, uint32_t mask, uint32_t lut_val) {
-        if (pos >= 0) {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                lut_val = extract_fields(pos - 1, val, mask, lut_val);
-            } else {
-                auto new_val = lut_val << 1;
-                if ((val & bitmask) != 0) new_val++;
-                lut_val = extract_fields(pos - 1, val, mask, new_val);
-            }
-        }
-        return lut_val;
-    }
+    std::array<std::vector<instruction_pattern>, 4> qlut;
 
     inline void raise(uint16_t trap_id, uint16_t cause){
         auto trap_val =  0x80ULL << 24 | (cause << 16) | trap_id;
@@ -296,22 +264,46 @@ template <typename CODE_WORD> void debug_fn(CODE_WORD insn) {
 
 template <typename ARCH> vm_impl<ARCH>::vm_impl() { this(new ARCH()); }
 
+// according to
+// https://stackoverflow.com/questions/8871204/count-number-of-1s-in-binary-representation
+#ifdef __GCC__
+constexpr size_t bit_count(uint32_t u) { return __builtin_popcount(u); }
+#elif __cplusplus < 201402L
+constexpr size_t uCount(uint32_t u) { return u - ((u >> 1) & 033333333333) - ((u >> 2) & 011111111111); }
+constexpr size_t bit_count(uint32_t u) { return ((uCount(u) + (uCount(u) >> 3)) & 030707070707) % 63; }
+#else
+constexpr size_t bit_count(uint32_t u) {
+    size_t uCount = u - ((u >> 1) & 033333333333) - ((u >> 2) & 011111111111);
+    return ((uCount + (uCount >> 3)) & 030707070707) % 63;
+}
+#endif
+
 template <typename ARCH>
 vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
 : vm_base<ARCH>(core, core_id, cluster_id) {
-    qlut[0] = lut_00.data();
-    qlut[1] = lut_01.data();
-    qlut[2] = lut_10.data();
-    qlut[3] = lut_11.data();
     for (auto instr : instr_descr) {
-        auto quantrant = instr.value & 0x3;
-        expand_bit_mask(29, lutmasks[quantrant], instr.value >> 2, instr.mask >> 2, 0, qlut[quantrant], instr.op);
+        auto quadrant = instr.value & 0x3;
+        qlut[quadrant].push_back(instruction_pattern{instr.value, instr.mask, instr.op});
+    }
+    for(auto& lut: qlut){
+        std::sort(std::begin(lut), std::end(lut), [](instruction_pattern const& a, instruction_pattern const& b){
+            return bit_count(a.mask) > bit_count(b.mask);
+        });
     }
 }
 
 inline bool is_count_limit_enabled(finish_cond_e cond){
     return (cond & finish_cond_e::COUNT_LIMIT) == finish_cond_e::COUNT_LIMIT;
 }
+
+template <typename ARCH>
+typename vm_impl<ARCH>::compile_func vm_impl<ARCH>::decode_inst(code_word_t instr){
+    for(auto& e: qlut[instr&0x3]){
+        if(!((instr&e.mask) ^ e.value )) return e.opc;
+    }
+    return &this_class::illegal_intruction;
+}
+
 template <typename ARCH>
 typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit){
     // we fetch at max 4 byte, alignment is 2
@@ -329,10 +321,7 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
         }
         if ((cond & finish_cond_e::JUMP_TO_SELF) == finish_cond_e::JUMP_TO_SELF &&
                 (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
-        auto lut_val = extract_fields(insn);
-        auto f = qlut[insn & 0x3][lut_val];
-        if (!f)
-            f = &this_class::illegal_intruction;
+        auto f = decode_inst(insn);
         pc = (this->*f)(pc, insn);
     }
     return pc;

incl/iss/arch/riscv_hart_m_p.h

@@ -144,10 +144,10 @@ public:
 
         mstatus_t mstatus;
 
-        static const reg_t mstatus_reset_val = 0;
+        static const reg_t mstatus_reset_val = 0x1800;
 
         void write_mstatus(T val) {
-            auto mask = get_mask();
+            auto mask = get_mask() &0xff; // MPP is hardcode as 0x3
             auto new_val = (mstatus.backing.val & ~mask) | (val & mask);
             mstatus = new_val;
         }

src/vm/interp/vm_tgc_c.cpp

@@ -85,6 +85,7 @@ protected:
 
     inline const char *name(size_t index){return traits::reg_aliases.at(index);}
 
+    compile_func decode_inst(code_word_t instr) ;
     virt_addr_t execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit) override;
 
     // some compile time constants
@@ -98,46 +99,13 @@ protected:
     std::array<compile_func, LUT_SIZE_C> lut_00, lut_01, lut_10;
     std::array<compile_func, LUT_SIZE> lut_11;
 
-    std::array<compile_func *, 4> qlut;
+    struct instruction_pattern {
+        uint32_t value;
+        uint32_t mask;
+        compile_func opc;
+    };
 
-    std::array<const uint32_t, 4> lutmasks = {{EXTR_MASK16, EXTR_MASK16, EXTR_MASK16, EXTR_MASK32}};
-
-    void expand_bit_mask(int pos, uint32_t mask, uint32_t value, uint32_t valid, uint32_t idx, compile_func lut[],
-                         compile_func f) {
-        if (pos < 0) {
-            lut[idx] = f;
-        } else {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                expand_bit_mask(pos - 1, mask, value, valid, idx, lut, f);
-            } else {
-                if ((valid & bitmask) == 0) {
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1), lut, f);
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1) + 1, lut, f);
-                } else {
-                    auto new_val = idx << 1;
-                    if ((value & bitmask) != 0) new_val++;
-                    expand_bit_mask(pos - 1, mask, value, valid, new_val, lut, f);
-                }
-            }
-        }
-    }
-
-    inline uint32_t extract_fields(uint32_t val) { return extract_fields(29, val >> 2, lutmasks[val & 0x3], 0); }
-
-    uint32_t extract_fields(int pos, uint32_t val, uint32_t mask, uint32_t lut_val) {
-        if (pos >= 0) {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                lut_val = extract_fields(pos - 1, val, mask, lut_val);
-            } else {
-                auto new_val = lut_val << 1;
-                if ((val & bitmask) != 0) new_val++;
-                lut_val = extract_fields(pos - 1, val, mask, new_val);
-            }
-        }
-        return lut_val;
-    }
+    std::array<std::vector<instruction_pattern>, 4> qlut;
 
     inline void raise(uint16_t trap_id, uint16_t cause){
         auto trap_val =  0x80ULL << 24 | (cause << 16) | trap_id;
@@ -4113,22 +4081,46 @@ template <typename CODE_WORD> void debug_fn(CODE_WORD insn) {
 
 template <typename ARCH> vm_impl<ARCH>::vm_impl() { this(new ARCH()); }
 
+// according to
+// https://stackoverflow.com/questions/8871204/count-number-of-1s-in-binary-representation
+#ifdef __GCC__
+constexpr size_t bit_count(uint32_t u) { return __builtin_popcount(u); }
+#elif __cplusplus < 201402L
+constexpr size_t uCount(uint32_t u) { return u - ((u >> 1) & 033333333333) - ((u >> 2) & 011111111111); }
+constexpr size_t bit_count(uint32_t u) { return ((uCount(u) + (uCount(u) >> 3)) & 030707070707) % 63; }
+#else
+constexpr size_t bit_count(uint32_t u) {
+    size_t uCount = u - ((u >> 1) & 033333333333) - ((u >> 2) & 011111111111);
+    return ((uCount + (uCount >> 3)) & 030707070707) % 63;
+}
+#endif
+
 template <typename ARCH>
 vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
 : vm_base<ARCH>(core, core_id, cluster_id) {
-    qlut[0] = lut_00.data();
-    qlut[1] = lut_01.data();
-    qlut[2] = lut_10.data();
-    qlut[3] = lut_11.data();
     for (auto instr : instr_descr) {
-        auto quantrant = instr.value & 0x3;
-        expand_bit_mask(29, lutmasks[quantrant], instr.value >> 2, instr.mask >> 2, 0, qlut[quantrant], instr.op);
+        auto quadrant = instr.value & 0x3;
+        qlut[quadrant].push_back(instruction_pattern{instr.value, instr.mask, instr.op});
+    }
+    for(auto& lut: qlut){
+        std::sort(std::begin(lut), std::end(lut), [](instruction_pattern const& a, instruction_pattern const& b){
+            return bit_count(a.mask) > bit_count(b.mask);
+        });
     }
 }
 
 inline bool is_count_limit_enabled(finish_cond_e cond){
     return (cond & finish_cond_e::COUNT_LIMIT) == finish_cond_e::COUNT_LIMIT;
 }
+
+template <typename ARCH>
+typename vm_impl<ARCH>::compile_func vm_impl<ARCH>::decode_inst(code_word_t instr){
+    for(auto& e: qlut[instr&0x3]){
+        if(!((instr&e.mask) ^ e.value )) return e.opc;
+    }
+    return &this_class::illegal_intruction;
+}
+
 template <typename ARCH>
 typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit){
     // we fetch at max 4 byte, alignment is 2
@@ -4146,10 +4138,7 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
         }
         if ((cond & finish_cond_e::JUMP_TO_SELF) == finish_cond_e::JUMP_TO_SELF &&
                 (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
-        auto lut_val = extract_fields(insn);
-        auto f = qlut[insn & 0x3][lut_val];
-        if (!f)
-            f = &this_class::illegal_intruction;
+        auto f = decode_inst(insn);
         pc = (this->*f)(pc, insn);
     }
     return pc;