Merge pull request #109 from sysprog21/mmu-caching

jserv · web-flow · commit ec83f4fb8aa8 · 2025-11-01T10:44:38.000+08:00
Upgrade MMU cache to 8×2 set-associative
diff --git a/main.c b/main.c
@@ -910,6 +910,18 @@ static int semu_step(emu_state_t *emu)
 }
 
 #ifdef MMU_CACHE_STATS
+static vm_t *global_vm_for_signal = NULL;
+static volatile sig_atomic_t signal_received = 0;
+
+/* Forward declaration */
+static void print_mmu_cache_stats(vm_t *vm);
+
+/* Async-signal-safe handler: only set flag, defer printing */
+static void signal_handler_stats(int sig UNUSED)
+{
+    signal_received = 1;
+}
+
 static void print_mmu_cache_stats(vm_t *vm)
 {
     fprintf(stderr, "\n=== MMU Cache Statistics ===\n");
@@ -918,15 +930,25 @@ static void print_mmu_cache_stats(vm_t *vm)
         uint64_t fetch_total =
             hart->cache_fetch.hits + hart->cache_fetch.misses;
 
-        /* Combine 2-way load cache statistics */
-        uint64_t load_hits =
-            hart->cache_load[0].hits + hart->cache_load[1].hits;
-        uint64_t load_misses =
-            hart->cache_load[0].misses + hart->cache_load[1].misses;
+        /* Combine 8-set × 2-way load cache statistics */
+        uint64_t load_hits = 0, load_misses = 0;
+        for (int set = 0; set < 8; set++) {
+            for (int way = 0; way < 2; way++) {
+                load_hits += hart->cache_load[set].ways[way].hits;
+                load_misses += hart->cache_load[set].ways[way].misses;
+            }
+        }
         uint64_t load_total = load_hits + load_misses;
 
-        uint64_t store_total =
-            hart->cache_store.hits + hart->cache_store.misses;
+        /* Combine 8-set × 2-way store cache statistics */
+        uint64_t store_hits = 0, store_misses = 0;
+        for (int set = 0; set < 8; set++) {
+            for (int way = 0; way < 2; way++) {
+                store_hits += hart->cache_store[set].ways[way].hits;
+                store_misses += hart->cache_store[set].ways[way].misses;
+            }
+        }
+        uint64_t store_total = store_hits + store_misses;
 
         fprintf(stderr, "\nHart %u:\n", i);
         fprintf(stderr, "  Fetch: %12llu hits, %12llu misses",
@@ -936,18 +958,18 @@ static void print_mmu_cache_stats(vm_t *vm)
                     100.0 * hart->cache_fetch.hits / fetch_total);
         fprintf(stderr, "\n");
 
-        fprintf(stderr, "  Load:  %12llu hits, %12llu misses (2-way)",
-                load_hits, load_misses);
+        fprintf(stderr, "  Load:  %12llu hits, %12llu misses (8x2)", load_hits,
+                load_misses);
         if (load_total > 0)
             fprintf(stderr, " (%.2f%% hit rate)",
                     100.0 * load_hits / load_total);
         fprintf(stderr, "\n");
 
-        fprintf(stderr, "  Store: %12llu hits, %12llu misses",
-                hart->cache_store.hits, hart->cache_store.misses);
+        fprintf(stderr, "  Store: %12llu hits, %12llu misses (8x2)", store_hits,
+                store_misses);
         if (store_total > 0)
             fprintf(stderr, " (%.2f%% hit rate)",
-                    100.0 * hart->cache_store.hits / store_total);
+                    100.0 * store_hits / store_total);
         fprintf(stderr, "\n");
     }
 }
@@ -1007,6 +1029,13 @@ static int semu_run(emu_state_t *emu)
 #endif
 
         while (!emu->stopped) {
+#ifdef MMU_CACHE_STATS
+            /* Check if signal received (SIGINT/SIGTERM) */
+            if (signal_received) {
+                print_mmu_cache_stats(&emu->vm);
+                return 0;
+            }
+#endif
             /* Resume each hart's coroutine in round-robin fashion */
             for (uint32_t i = 0; i < vm->n_hart; i++) {
                 coro_resume_hart(i);
@@ -1100,6 +1129,11 @@ static int semu_run(emu_state_t *emu)
             if (ret)
                 return ret;
 #ifdef MMU_CACHE_STATS
+            /* Check if signal received (SIGINT/SIGTERM) */
+            if (signal_received) {
+                print_mmu_cache_stats(&emu->vm);
+                return 0;
+            }
             /* Exit after running for 15 seconds to collect statistics */
             gettimeofday(&current_time, NULL);
             long elapsed_sec = current_time.tv_sec - start_time.tv_sec;
@@ -1246,6 +1280,12 @@ int main(int argc, char **argv)
     if (ret)
         return ret;
 
+#ifdef MMU_CACHE_STATS
+    global_vm_for_signal = &emu.vm;
+    signal(SIGINT, signal_handler_stats);
+    signal(SIGTERM, signal_handler_stats);
+#endif
+
     if (emu.debug)
         ret = semu_run_debug(&emu);
     else
diff --git a/riscv.c b/riscv.c
@@ -185,9 +185,18 @@ static inline uint32_t read_rs2(const hart_t *vm, uint32_t insn)
 void mmu_invalidate(hart_t *vm)
 {
     vm->cache_fetch.n_pages = 0xFFFFFFFF;
-    vm->cache_load[0].n_pages = 0xFFFFFFFF;
-    vm->cache_load[1].n_pages = 0xFFFFFFFF;
-    vm->cache_store.n_pages = 0xFFFFFFFF;
+    /* Invalidate all 8 sets × 2 ways for load cache */
+    for (int set = 0; set < 8; set++) {
+        for (int way = 0; way < 2; way++)
+            vm->cache_load[set].ways[way].n_pages = 0xFFFFFFFF;
+        vm->cache_load[set].lru = 0; /* Reset LRU to way 0 */
+    }
+    /* Invalidate all 8 sets × 2 ways for store cache */
+    for (int set = 0; set < 8; set++) {
+        for (int way = 0; way < 2; way++)
+            vm->cache_store[set].ways[way].n_pages = 0xFFFFFFFF;
+        vm->cache_store[set].lru = 0; /* Reset LRU to way 0 */
+    }
 }
 
 /* Pre-verify the root page table to minimize page table access during
@@ -333,13 +342,36 @@ static void mmu_load(hart_t *vm,
 {
     uint32_t vpn = addr >> RV_PAGE_SHIFT;
     uint32_t phys_addr;
-    /* 2-entry direct-mapped cache: use parity hash to select entry */
-    uint32_t index = __builtin_parity(vpn) & 0x1;
+    /* 8-set × 2-way set-associative cache: use 3-bit parity hash */
+    uint32_t set_idx = (__builtin_parity(vpn & 0xAAAAAAAA) << 2) |
+                       (__builtin_parity(vpn & 0x55555555) << 1) |
+                       __builtin_parity(vpn & 0xCCCCCCCC);
+
+    mmu_cache_set_t *set = &vm->cache_load[set_idx];
+
+    /* Check both ways in the set */
+    int hit_way = -1;
+    for (int way = 0; way < 2; way++) {
+        if (likely(set->ways[way].n_pages == vpn)) {
+            hit_way = way;
+            break;
+        }
+    }
 
-    if (unlikely(vpn != vm->cache_load[index].n_pages)) {
+    if (likely(hit_way >= 0)) {
+        /* Cache hit: reconstruct physical address from cached PPN */
+#ifdef MMU_CACHE_STATS
+        set->ways[hit_way].hits++;
+#endif
+        phys_addr = (set->ways[hit_way].phys_ppn << RV_PAGE_SHIFT) |
+                    (addr & MASK(RV_PAGE_SHIFT));
+        /* Update LRU: mark the other way as replacement candidate */
+        set->lru = 1 - hit_way;
+    } else {
         /* Cache miss: do full translation */
+        int victim_way = set->lru; /* Use LRU bit to select victim */
 #ifdef MMU_CACHE_STATS
-        vm->cache_load[index].misses++;
+        set->ways[victim_way].misses++;
 #endif
         phys_addr = addr;
         mmu_translate(vm, &phys_addr,
@@ -348,16 +380,11 @@ static void mmu_load(hart_t *vm,
                       RV_EXC_LOAD_PFAULT);
         if (vm->error)
             return;
-        /* Cache physical page number (not a pointer) */
-        vm->cache_load[index].n_pages = vpn;
-        vm->cache_load[index].phys_ppn = phys_addr >> RV_PAGE_SHIFT;
-    } else {
-        /* Cache hit: reconstruct physical address from cached PPN */
-#ifdef MMU_CACHE_STATS
-        vm->cache_load[index].hits++;
-#endif
-        phys_addr = (vm->cache_load[index].phys_ppn << RV_PAGE_SHIFT) |
-                    (addr & MASK(RV_PAGE_SHIFT));
+        /* Replace victim way with new translation */
+        set->ways[victim_way].n_pages = vpn;
+        set->ways[victim_way].phys_ppn = phys_addr >> RV_PAGE_SHIFT;
+        /* Update LRU: mark the other way for next eviction */
+        set->lru = 1 - victim_way;
     }
 
     vm->mem_load(vm, phys_addr, width, value);
@@ -376,28 +403,48 @@ static bool mmu_store(hart_t *vm,
 {
     uint32_t vpn = addr >> RV_PAGE_SHIFT;
     uint32_t phys_addr;
+    /* 8-set × 2-way set-associative cache: use 3-bit parity hash */
+    uint32_t set_idx = (__builtin_parity(vpn & 0xAAAAAAAA) << 2) |
+                       (__builtin_parity(vpn & 0x55555555) << 1) |
+                       __builtin_parity(vpn & 0xCCCCCCCC);
+
+    mmu_cache_set_t *set = &vm->cache_store[set_idx];
+
+    /* Check both ways in the set */
+    int hit_way = -1;
+    for (int way = 0; way < 2; way++) {
+        if (likely(set->ways[way].n_pages == vpn)) {
+            hit_way = way;
+            break;
+        }
+    }
 
-    if (unlikely(vpn != vm->cache_store.n_pages)) {
+    if (likely(hit_way >= 0)) {
+        /* Cache hit: reconstruct physical address from cached PPN */
+#ifdef MMU_CACHE_STATS
+        set->ways[hit_way].hits++;
+#endif
+        phys_addr = (set->ways[hit_way].phys_ppn << RV_PAGE_SHIFT) |
+                    (addr & MASK(RV_PAGE_SHIFT));
+        /* Update LRU: mark the other way as replacement candidate */
+        set->lru = 1 - hit_way;
+    } else {
         /* Cache miss: do full translation */
+        int victim_way = set->lru; /* Use LRU bit to select victim */
 #ifdef MMU_CACHE_STATS
-        vm->cache_store.misses++;
+        set->ways[victim_way].misses++;
 #endif
         phys_addr = addr;
         mmu_translate(vm, &phys_addr, (1 << 2), (1 << 6) | (1 << 7),
                       vm->sstatus_sum && vm->s_mode, RV_EXC_STORE_FAULT,
                       RV_EXC_STORE_PFAULT);
         if (vm->error)
             return false;
-        /* Cache physical page number (not a pointer) */
-        vm->cache_store.n_pages = vpn;
-        vm->cache_store.phys_ppn = phys_addr >> RV_PAGE_SHIFT;
-    } else {
-        /* Cache hit: reconstruct physical address from cached PPN */
-#ifdef MMU_CACHE_STATS
-        vm->cache_store.hits++;
-#endif
-        phys_addr = (vm->cache_store.phys_ppn << RV_PAGE_SHIFT) |
-                    (addr & MASK(RV_PAGE_SHIFT));
+        /* Replace victim way with new translation */
+        set->ways[victim_way].n_pages = vpn;
+        set->ways[victim_way].phys_ppn = phys_addr >> RV_PAGE_SHIFT;
+        /* Update LRU: mark the other way for next eviction */
+        set->lru = 1 - victim_way;
     }
 
     if (unlikely(cond)) {
diff --git a/riscv.h b/riscv.h
@@ -51,6 +51,12 @@ typedef struct {
 #endif
 } mmu_addr_cache_t;
 
+/* Set-associative cache structure for load operations */
+typedef struct {
+    mmu_addr_cache_t ways[2]; /* 2-way associative */
+    uint8_t lru;              /* LRU bit: 0 or 1 (which way to replace) */
+} mmu_cache_set_t;
+
 /* To use the emulator, start by initializing a hart_t object with zero values,
  * invoke vm_init(), and set the required environment-supplied callbacks. You
  * may also set other necessary fields such as argument registers and s_mode,
@@ -101,9 +107,10 @@ struct __hart_internal {
     uint32_t exc_cause, exc_val;
 
     mmu_fetch_cache_t cache_fetch;
-    /* 2-entry direct-mapped with hash-based indexing */
-    mmu_addr_cache_t cache_load[2];
-    mmu_addr_cache_t cache_store;
+    /* 8-set × 2-way set-associative cache with 3-bit parity hash indexing */
+    mmu_cache_set_t cache_load[8];
+    /* 8-set × 2-way set-associative cache for store operations */
+    mmu_cache_set_t cache_store[8];
 
     /* Supervisor state */
     bool s_mode;
diff --git a/utils.c b/utils.c
@@ -1,4 +1,6 @@
+#include <math.h>
 #include <stdbool.h>
+#include <stdio.h>
 #include <time.h>
 
 #include "utils.h"
@@ -24,6 +26,10 @@ bool boot_complete = false;
 static double ticks_increment;
 static double boot_ticks;
 
+/* Timer calibration statistics */
+static uint64_t timer_call_count = 0;
+static int timer_n_harts = 1;
+
 /* Calculate "x * n / d" without unnecessary overflow or loss of precision.
  *
  * Reference:
@@ -88,6 +94,7 @@ static uint64_t semu_timer_clocksource(semu_timer_t *timer)
     static bool first_switch = true;
 
     if (!boot_complete) {
+        timer_call_count++;
         boot_ticks += ticks_increment;
         return (uint64_t) boot_ticks;
     }
@@ -98,6 +105,34 @@ static uint64_t semu_timer_clocksource(semu_timer_t *timer)
 
         /* Calculate the offset between the real time and the emulator time */
         offset = (int64_t) (real_ticks - boot_ticks);
+
+#ifdef SEMU_TIMER_STATS
+        /* Output timer calibration statistics (only when SEMU_TIMER_STATS is
+         * defined) */
+        double actual_coefficient = (double) timer_call_count / timer_n_harts;
+        double current_coefficient = 1.744e8;
+        double recommended_coefficient = actual_coefficient;
+
+        fprintf(stderr, "\n[Timer Calibration Statistics]\n");
+        fprintf(stderr, "  Boot completed after %llu timer calls\n",
+                (unsigned long long) timer_call_count);
+        fprintf(stderr, "  Number of harts: %d\n", timer_n_harts);
+        fprintf(stderr, "  Actual coefficient: %.3e (%.2f calls per hart)\n",
+                actual_coefficient, actual_coefficient);
+        fprintf(stderr, "  Current coefficient: %.3e\n", current_coefficient);
+        fprintf(stderr, "  Difference: %.2f%% %s\n",
+                fabs(actual_coefficient - current_coefficient) /
+                    current_coefficient * 100.0,
+                actual_coefficient > current_coefficient ? "(more calls)"
+                                                         : "(fewer calls)");
+        fprintf(stderr, "\n[Recommendation]\n");
+        fprintf(stderr, "  Update utils.c line 121 to:\n");
+        fprintf(stderr,
+                "  ticks_increment = (SEMU_BOOT_TARGET_TIME * CLOCK_FREQ) / "
+                "(%.3e * n_harts);\n",
+                recommended_coefficient);
+        fprintf(stderr, "\n");
+#endif
     }
     return (uint64_t) ((int64_t) real_ticks - offset);
 }
@@ -108,14 +143,28 @@ void semu_timer_init(semu_timer_t *timer, uint64_t freq, int n_harts)
     timer->begin = mult_frac(host_time_ns(), timer->freq, 1e9);
     boot_ticks = timer->begin; /* Initialize the fake ticks for boot process */
 
+    /* Store n_harts for calibration statistics */
+    timer_n_harts = n_harts;
+
     /* According to statistics, the number of times 'semu_timer_clocksource'
-     * called is approximately 'SMP count * 2.15 * 1e8'. By the time the boot
+     * called is approximately 'SMP count * 1.744 * 1e8'. By the time the boot
      * process is completed, the emulator will have a total of 'boot seconds *
-     * frequency' ticks. Therefore, each time, '(boot seconds * frequency) /
-     * (2.15 * 1e8 * SMP count)' ticks need to be added.
+     * frequency' ticks. Therefore, each time, (boot seconds * frequency) /
+     * (1.744 * 1e8 * SMP count) ticks need to be added.
+     *
+     * Note: This coefficient was recalibrated after MMU cache optimization
+     * (8×2 set-associative with 99%+ hit rate). The original coefficient
+     * (2.15 * 1e8) was based on measurements before the optimization. With
+     * faster CPU execution, fewer timer calls are needed to complete boot.
+     *
+     * Calibration history:
+     * - Original (pre-MMU cache): 2.15 × 10^8
+     * - After MMU cache (measured): 1.696 × 10^8 (-21.1%)
+     * - Verification measurement: 1.744 × 10^8 (error: 2.85%)
+     * - Final coefficient: 1.744 × 10^8 (based on verification)
      */
     ticks_increment =
-        (SEMU_BOOT_TARGET_TIME * CLOCK_FREQ) / (2.15 * 1e8 * n_harts);
+        (SEMU_BOOT_TARGET_TIME * CLOCK_FREQ) / (1.744 * 1e8 * n_harts);
 }
 
 uint64_t semu_timer_get(semu_timer_t *timer)
