pu3 poc5

1 files changed, 31 insertions, 23 deletions

diff --git a/pu3/pu3.cpu.c b/pu3/pu3.cpu.c
index 312b5a8..a10ca3f 100644
--- a/pu3/pu3.cpu.c
+++ b/pu3/pu3.cpu.c
@@ -86,6 +86,7 @@ typedef struct {
 //     uint16_t data;
 //     uint32_t dest;
 // } pu3_inst_t;
+
 // instructions
 // 0:  nop
 // 1:  hlt
@@ -109,14 +110,14 @@ typedef struct {
 #define prep_raw()                                                                                 \
     ({                                                                                             \
         rhs = (uint32_t) get_data_u16(cur);                                                        \
-        if (get_spare(cur) & 0x40) rhs = ctx.memory.memory_u32[ rhs ];                             \
+        if (get_spare(cur) & 0x40) rhs = mem32[ rhs ];                                             \
     })
 #define prep_vals()                                                                                \
     ({                                                                                             \
         spare = get_spare(cur);                                                                    \
         lhs   = get_dest_u32(cur);                                                                 \
         rhs   = (uint32_t) get_data_u16(cur);                                                      \
-        if (spare & 0x40) rhs = ctx.memory.memory_u32[ rhs ];                                      \
+        if (spare & 0x40) rhs = mem32[ rhs ];                                                      \
         if (unlikely(spare & 0x80)) {                                                              \
             out = lhs;                                                                             \
             lhs = rhs;                                                                             \
@@ -129,7 +130,10 @@ pu3(pu3_ctx ctx) {
     uint64_t   i = 0;
     pu3_inst_t cur;
 
-    pu3_inst_t *pc = ctx.memory.memory_u64 + *ctx.pc;
+    uint32_t   *mem32 = ctx.memory.memory_u32;
+    pu3_inst_t *mem64 = ctx.memory.memory_u64;
+
+    pu3_inst_t *pc = mem64 + *ctx.pc;
 
     byte op;
     byte spare;
@@ -162,82 +166,86 @@ pu3_inst_0: // nop
 
 pu3_inst_2: // mov
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] = rhs;
+    mem32[ lhs ] = rhs;
     goto pu3_start;
 
 pu3_inst_3: // add
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] += rhs;
+    mem32[ lhs ] += rhs;
     goto pu3_start;
 
 pu3_inst_4: // sub
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] -= rhs;
+    mem32[ lhs ] -= rhs;
     goto pu3_start;
 
 pu3_inst_5: // inc
     prep_only_left();
-    ctx.memory.memory_u32[ lhs ]++;
+    mem32[ lhs ]++;
     goto pu3_start;
 
 pu3_inst_6: // dec
     prep_only_left();
-    ctx.memory.memory_u32[ lhs ]--;
+    mem32[ lhs ]--;
     goto pu3_start;
 
 pu3_inst_7: // mul
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] *= rhs;
+    mem32[ lhs ] *= rhs;
     goto pu3_start;
 
 pu3_inst_8: // lsh
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] <<= rhs;
+    mem32[ lhs ] <<= rhs;
     goto pu3_start;
 
 pu3_inst_9: // rsh
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] >>= rhs;
+    mem32[ lhs ] >>= rhs;
     goto pu3_start;
 
 pu3_inst_10: // and
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] &= rhs;
+    mem32[ lhs ] &= rhs;
     goto pu3_start;
 
 pu3_inst_11: // not
     prep_vals();
-    ctx.memory.memory_u32[ lhs ] = !rhs;
+    mem32[ lhs ] = !rhs;
     goto pu3_start;
 
 pu3_inst_12: // jmp
     prep_only_left();
-    pc = ctx.memory.memory_u64 + lhs;
+    pc = mem64 + lhs;
     goto pu3_start;
 
 pu3_inst_13: // jnz
     prep_vals();
-    if (rhs) pc = ctx.memory.memory_u64 + lhs;
+    if (rhs) pc = mem64 + lhs;
     goto pu3_start;
 
-pu3_inst_14: // unused
+pu3_inst_14: // dmv
+    prep_vals();
+    mem32[ lhs ] = mem32[ mem32[ lhs ] ];
+    goto pu3_start;
 pu3_inst_15: // unused
+    goto pu3_start;
 
 pu3_inst_1: // hlt
-    *ctx.pc = pc - ctx.memory.memory_u64;
+    *ctx.pc = pc - mem64;
     return *ctx.time_slice - i;
 }
 
 const string op_names[] = { "nop", "hlt", "mov", "add", "sub", "inc", "dec", "mul",
-                            "lsh", "rsh", "and", "not", "jmp", "jnz", "???", "???" };
+                            "lsh", "rsh", "and", "not", "jmp", "jnz", "dmv", "???" };
 
 // bench: measure using clock()
 double bench_cycle(byte instr, pu3_ctx ctx) {
     int i;
     for (i = 16; i < ctx.mem_size - 16; i++) {
         set_scroll_byte(ctx.memory.memory_u64[ i ], instr & 0b1111);
-        set_data_u16(ctx.memory.memory_u64[ i ], 2);
-        set_dest_u32(ctx.memory.memory_u64[ i ], 4);
+        set_data_u16(ctx.memory.memory_u64[ i ], 4);
+        set_dest_u32(ctx.memory.memory_u64[ i ], i);
 
         if (i == ctx.mem_size - 20) {
             let j = 0;
@@ -270,7 +278,7 @@ double bench_cycle(byte instr, pu3_ctx ctx) {
 }
 
 int main() {
-    const uint32_t SIZE = 1 << 24;
+    const uint32_t SIZE = 1 << 23;
 
     let mem_arr = calloc(SIZE, sizeof(pu3_inst_t));
     if (!mem_arr) {
@@ -286,8 +294,8 @@ int main() {
     int    i;
     int    total  = 0;
     double benchs = 0;
-    for (i = 0; i < 12; i++) {
-        if (i == 1) continue;
+    for (i = 0; i < 15; i++) {
+        if (i == 1 || i == 12 || i == 13) continue;
         total++;
         benchs += bench_cycle(i, ctx);
     }