bpo-44187: Quickening infrastructure (GH-26264)

* Add co_firstinstr field to code object.

* Implement barebones quickening.

* Use non-quickened bytecode when tracing.

* Add NEWS item

* Add new file to Windows build.

* Don't specialize instructions with EXTENDED_ARG.

1 files changed, 197 insertions, 0 deletions

diff --git a/Python/specialize.c b/Python/specialize.c
new file mode 100644
index 00000000000..07152d80538
--- /dev/null
+++ b/Python/specialize.c
@@ -0,0 +1,197 @@
+
+#include "Python.h"
+#include "pycore_code.h"
+#include "opcode.h"
+
+
+/* We layout the quickened data as a bi-directional array:
+ * Instructions upwards, cache entries downwards.
+ * first_instr is aligned to a SpecializedCacheEntry.
+ * The nth instruction is located at first_instr[n]
+ * The nth cache is located at ((SpecializedCacheEntry *)first_instr)[-1-n]
+ * The first (index 0) cache entry is reserved for the count, to enable finding
+ * the first instruction from the base pointer.
+ * The cache_count argument must include space for the count.
+ * We use the SpecializedCacheOrInstruction union to refer to the data
+ * to avoid type punning.
+
+ Layout of quickened data, each line 8 bytes for M cache entries and N instructions:
+
+ <cache_count>                              <---- co->co_quickened
+ <cache M-1>
+ <cache M-2>
+ ...
+ <cache 0>
+ <instr 0> <instr 1> <instr 2> <instr 3>    <--- co->co_first_instr
+ <instr 4> <instr 5> <instr 6> <instr 7>
+ ...
+ <instr N-1>
+*/
+
+Py_ssize_t _Py_QuickenedCount = 0;
+
+static SpecializedCacheOrInstruction *
+allocate(int cache_count, int instruction_count)
+{
+    assert(sizeof(SpecializedCacheOrInstruction) == 2*sizeof(int32_t));
+    assert(sizeof(SpecializedCacheEntry) == 2*sizeof(int32_t));
+    assert(cache_count > 0);
+    assert(instruction_count > 0);
+    int count = cache_count + (instruction_count + INSTRUCTIONS_PER_ENTRY -1)/INSTRUCTIONS_PER_ENTRY;
+    SpecializedCacheOrInstruction *array = (SpecializedCacheOrInstruction *)
+        PyMem_Malloc(sizeof(SpecializedCacheOrInstruction) * count);
+    if (array == NULL) {
+        PyErr_NoMemory();
+        return NULL;
+    }
+    _Py_QuickenedCount++;
+    array[0].entry.zero.cache_count = cache_count;
+    return array;
+}
+
+static int
+get_cache_count(SpecializedCacheOrInstruction *quickened) {
+    return quickened[0].entry.zero.cache_count;
+}
+
+/* Map from opcode to adaptive opcode.
+  Values of zero are ignored. */
+static uint8_t adaptive_opcodes[256] = { 0 };
+
+/* The number of cache entries required for a "family" of instructions. */
+static uint8_t cache_requirements[256] = { 0 };
+
+/* Return the oparg for the cache_offset and instruction index.
+ *
+ * If no cache is needed then return the original oparg.
+ * If a cache is needed, but cannot be accessed because
+ * oparg would be too large, then return -1.
+ *
+ * Also updates the cache_offset, as it may need to be incremented by
+ * more than the cache requirements, if many instructions do not need caches.
+ *
+ * See pycore_code.h for details of how the cache offset,
+ * instruction index and oparg are related */
+static int
+oparg_from_instruction_and_update_offset(int index, int opcode, int original_oparg, int *cache_offset) {
+    /* The instruction pointer in the interpreter points to the next
+     * instruction, so we compute the offset using nexti (index + 1) */
+    int nexti = index + 1;
+    uint8_t need = cache_requirements[opcode];
+    if (need == 0) {
+        return original_oparg;
+    }
+    assert(adaptive_opcodes[opcode] != 0);
+    int oparg = oparg_from_offset_and_nexti(*cache_offset, nexti);
+    assert(*cache_offset == offset_from_oparg_and_nexti(oparg, nexti));
+    /* Some cache space is wasted here as the minimum possible offset is (nexti>>1) */
+    if (oparg < 0) {
+        oparg = 0;
+        *cache_offset = offset_from_oparg_and_nexti(oparg, nexti);
+    }
+    else if (oparg > 255) {
+        return -1;
+    }
+    *cache_offset += need;
+    return oparg;
+}
+
+static int
+entries_needed(_Py_CODEUNIT *code, int len)
+{
+    int cache_offset = 0;
+    int previous_opcode = -1;
+    for (int i = 0; i < len; i++) {
+        uint8_t opcode = _Py_OPCODE(code[i]);
+        if (previous_opcode != EXTENDED_ARG) {
+            oparg_from_instruction_and_update_offset(i, opcode, 0, &cache_offset);
+        }
+        previous_opcode = opcode;
+    }
+    return cache_offset + 1;   // One extra for the count entry
+}
+
+static inline _Py_CODEUNIT *
+first_instruction(SpecializedCacheOrInstruction *quickened)
+{
+    return &quickened[get_cache_count(quickened)].code[0];
+}
+
+/** Insert adaptive instructions and superinstructions.
+ *
+ * Skip instruction preceded by EXTENDED_ARG for adaptive
+ * instructions as those are both very rare and tricky
+ * to handle.
+ */
+static void
+optimize(SpecializedCacheOrInstruction *quickened, int len)
+{
+    _Py_CODEUNIT *instructions = first_instruction(quickened);
+    int cache_offset = 0;
+    int previous_opcode = -1;
+    for(int i = 0; i < len; i++) {
+        int opcode = _Py_OPCODE(instructions[i]);
+        int oparg = _Py_OPARG(instructions[i]);
+        uint8_t adaptive_opcode = adaptive_opcodes[opcode];
+        if (adaptive_opcode && previous_opcode != EXTENDED_ARG) {
+            int new_oparg = oparg_from_instruction_and_update_offset(
+                i, opcode, oparg, &cache_offset
+            );
+            if (new_oparg < 0) {
+                /* Not possible to allocate a cache for this instruction */
+                previous_opcode = opcode;
+                continue;
+            }
+            instructions[i] = _Py_MAKECODEUNIT(adaptive_opcode, new_oparg);
+            previous_opcode = adaptive_opcode;
+            int entries_needed = cache_requirements[opcode];
+            if (entries_needed) {
+                /* Initialize the adpative cache entry */
+                int cache0_offset = cache_offset-entries_needed;
+                SpecializedCacheEntry *cache =
+                    _GetSpecializedCacheEntry(instructions, cache0_offset);
+                cache->adaptive.original_oparg = oparg;
+                cache->adaptive.counter = 0;
+            }
+        }
+        else {
+            /* Super instructions don't use the cache,
+             * so no need to update the offset. */
+            switch (opcode) {
+                /* Insert superinstructions here
+                 E.g.
+                case LOAD_FAST:
+                    if (previous_opcode == LOAD_FAST)
+                        instructions[i-1] = _Py_MAKECODEUNIT(LOAD_FAST__LOAD_FAST, oparg);
+                 */
+            }
+            previous_opcode = opcode;
+        }
+    }
+    assert(cache_offset+1 == get_cache_count(quickened));
+}
+
+int
+_Py_Quicken(PyCodeObject *code) {
+    if (code->co_quickened) {
+        return 0;
+    }
+    Py_ssize_t size = PyBytes_GET_SIZE(code->co_code);
+    int instr_count = (int)(size/sizeof(_Py_CODEUNIT));
+    if (instr_count > MAX_SIZE_TO_QUICKEN) {
+        code->co_warmup = QUICKENING_WARMUP_COLDEST;
+        return 0;
+    }
+    int entry_count = entries_needed(code->co_firstinstr, instr_count);
+    SpecializedCacheOrInstruction *quickened = allocate(entry_count, instr_count);
+    if (quickened == NULL) {
+        return -1;
+    }
+    _Py_CODEUNIT *new_instructions = first_instruction(quickened);
+    memcpy(new_instructions, code->co_firstinstr, size);
+    optimize(quickened, instr_count);
+    code->co_quickened = quickened;
+    code->co_firstinstr = new_instructions;
+    return 0;
+}
+