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.

Author: markshannon

Include/cpython/code.h

@@ -7,9 +7,11 @@ typedef uint16_t _Py_CODEUNIT;
 #ifdef WORDS_BIGENDIAN
 #  define _Py_OPCODE(word) ((word) >> 8)
 #  define _Py_OPARG(word) ((word) & 255)
+#  define _Py_MAKECODEUNIT(opcode, oparg) (((opcode)<<8)|(oparg))
 #else
 #  define _Py_OPCODE(word) ((word) & 255)
 #  define _Py_OPARG(word) ((word) >> 8)
+#  define _Py_MAKECODEUNIT(opcode, oparg) ((opcode)|((oparg)<<8))
 #endif
 
 typedef struct _PyOpcache _PyOpcache;
@@ -43,24 +45,27 @@ struct PyCodeObject {
     /* These fields are set with provided values on new code objects. */
 
     // The hottest fields (in the eval loop) are grouped here at the top.
-    PyObject *co_code;          /* instruction opcodes */
     PyObject *co_consts;        /* list (constants used) */
     PyObject *co_names;         /* list of strings (names used) */
+    _Py_CODEUNIT *co_firstinstr; /* Pointer to first instruction, used for quickening */
+    PyObject *co_exceptiontable; /* Byte string encoding exception handling table */
     int co_flags;               /* CO_..., see below */
+    int co_warmup;              /* Warmup counter for quickening */
+
     // The rest are not so impactful on performance.
     int co_argcount;            /* #arguments, except *args */
     int co_posonlyargcount;     /* #positional only arguments */
     int co_kwonlyargcount;      /* #keyword only arguments */
     int co_stacksize;           /* #entries needed for evaluation stack */
     int co_firstlineno;         /* first source line number */
+    PyObject *co_code;          /* instruction opcodes */
     PyObject *co_varnames;      /* tuple of strings (local variable names) */
     PyObject *co_cellvars;      /* tuple of strings (cell variable names) */
     PyObject *co_freevars;      /* tuple of strings (free variable names) */
     PyObject *co_filename;      /* unicode (where it was loaded from) */
     PyObject *co_name;          /* unicode (name, for reference) */
     PyObject *co_linetable;     /* string (encoding addr<->lineno mapping) See
                                    Objects/lnotab_notes.txt for details. */
-    PyObject *co_exceptiontable; /* Byte string encoding exception handling table */
 
     /* These fields are set with computed values on new code objects. */
 
@@ -78,6 +83,10 @@ struct PyCodeObject {
        Type is a void* to keep the format private in codeobject.c to force
        people to go through the proper APIs. */
     void *co_extra;
+    /* Quickened instructions and cache, or NULL
+     This should be treated as opaque by all code except the specializer and
+     interpreter. */
+    union _cache_or_instruction *co_quickened;
 
     /* Per opcodes just-in-time cache
      *

Include/internal/pycore_code.h

@@ -4,6 +4,7 @@
 extern "C" {
 #endif
 
+/* Legacy Opcache */
 
 typedef struct {
     PyObject *ptr;  /* Cached pointer (borrowed reference) */
@@ -26,6 +27,129 @@ struct _PyOpcache {
 };
 
 
+/* PEP 659
+ * Specialization and quickening structs and helper functions
+ */
+
+typedef struct {
+    int32_t cache_count;
+    int32_t _; /* Force 8 byte size */
+} _PyEntryZero;
+
+typedef struct {
+    uint8_t original_oparg;
+    uint8_t counter;
+    uint16_t index;
+} _PyAdaptiveEntry;
+
+/* Add specialized versions of entries to this union.
+ *
+ * Do not break the invariant: sizeof(SpecializedCacheEntry) == 8
+ * Preserving this invariant is necessary because:
+    - If any one form uses more space, then all must and on 64 bit machines
+      this is likely to double the memory consumption of caches
+    - The function for calculating the offset of caches assumes a 4:1
+      cache:instruction size ratio. Changing that would need careful
+      analysis to choose a new function.
+ */
+typedef union {
+    _PyEntryZero zero;
+    _PyAdaptiveEntry adaptive;
+} SpecializedCacheEntry;
+
+#define INSTRUCTIONS_PER_ENTRY (sizeof(SpecializedCacheEntry)/sizeof(_Py_CODEUNIT))
+
+/* Maximum size of code to quicken, in code units. */
+#define MAX_SIZE_TO_QUICKEN 5000
+
+typedef union _cache_or_instruction {
+    _Py_CODEUNIT code[1];
+    SpecializedCacheEntry entry;
+} SpecializedCacheOrInstruction;
+
+/* Get pointer to the nth cache entry, from the first instruction and n.
+ * Cache entries are indexed backwards, with [count-1] first in memory, and [0] last.
+ * The zeroth entry immediately precedes the instructions.
+ */
+static inline SpecializedCacheEntry *
+_GetSpecializedCacheEntry(_Py_CODEUNIT *first_instr, Py_ssize_t n)
+{
+    SpecializedCacheOrInstruction *last_cache_plus_one = (SpecializedCacheOrInstruction *)first_instr;
+    assert(&last_cache_plus_one->code[0] == first_instr);
+    return &last_cache_plus_one[-1-n].entry;
+}
+
+/* Following two functions form a pair.
+ *
+ * oparg_from_offset_and_index() is used to compute the oparg
+ * when quickening, so that offset_from_oparg_and_nexti()
+ * can be used at runtime to compute the offset.
+ *
+ * The relationship between the three values is currently
+ *     offset == (index>>1) + oparg
+ * This relation is chosen based on the following observations:
+ * 1. typically 1 in 4 instructions need a cache
+ * 2. instructions that need a cache typically use 2 entries
+ *  These observations imply:  offset ≈ index/2
+ *  We use the oparg to fine tune the relation to avoid wasting space
+ * and allow consecutive instructions to use caches.
+ *
+ * If the number of cache entries < number of instructions/2 we will waste
+ * some small amoount of space.
+ * If the number of cache entries > (number of instructions/2) + 255, then
+ * some instructions will not be able to use a cache.
+ * In practice, we expect some small amount of wasted space in a shorter functions
+ * and only functions exceeding a 1000 lines or more not to have enugh cache space.
+ *
+ */
+static inline int
+oparg_from_offset_and_nexti(int offset, int nexti)
+{
+    return offset-(nexti>>1);
+}
+
+static inline int
+offset_from_oparg_and_nexti(int oparg, int nexti)
+{
+    return (nexti>>1)+oparg;
+}
+
+/* Get pointer to the cache entry associated with an instruction.
+ * nexti is the index of the instruction plus one.
+ * nexti is used as it corresponds to the instruction pointer in the interpreter.
+ * This doesn't check that an entry has been allocated for that instruction. */
+static inline SpecializedCacheEntry *
+_GetSpecializedCacheEntryForInstruction(_Py_CODEUNIT *first_instr, int nexti, int oparg)
+{
+    return _GetSpecializedCacheEntry(
+        first_instr,
+        offset_from_oparg_and_nexti(oparg, nexti)
+    );
+}
+
+#define QUICKENING_WARMUP_DELAY 8
+
+/* We want to compare to zero for efficiency, so we offset values accordingly */
+#define QUICKENING_INITIAL_WARMUP_VALUE (-QUICKENING_WARMUP_DELAY)
+#define QUICKENING_WARMUP_COLDEST 1
+
+static inline void
+PyCodeObject_IncrementWarmup(PyCodeObject * co)
+{
+    co->co_warmup++;
+}
+
+/* Used by the interpreter to determine when a code object should be quickened */
+static inline int
+PyCodeObject_IsWarmedUp(PyCodeObject * co)
+{
+    return (co->co_warmup == 0);
+}
+
+int _Py_Quicken(PyCodeObject *code);
+
+extern Py_ssize_t _Py_QuickenedCount;
+
 struct _PyCodeConstructor {
     /* metadata */
     PyObject *filename;

Lib/test/libregrtest/refleak.py

@@ -73,9 +73,10 @@ def get_pooled_int(value):
     alloc_deltas = [0] * repcount
     fd_deltas = [0] * repcount
     getallocatedblocks = sys.getallocatedblocks
+    getallocatedblocks = sys.getallocatedblocks
     gettotalrefcount = sys.gettotalrefcount
+    _getquickenedcount = sys._getquickenedcount
     fd_count = os_helper.fd_count
-
     # initialize variables to make pyflakes quiet
     rc_before = alloc_before = fd_before = 0
 
@@ -92,7 +93,7 @@ def get_pooled_int(value):
 
         # dash_R_cleanup() ends with collecting cyclic trash:
         # read memory statistics immediately after.
-        alloc_after = getallocatedblocks()
+        alloc_after = getallocatedblocks() - _getquickenedcount()
         rc_after = gettotalrefcount()
         fd_after = fd_count()
 

Makefile.pre.in

@@ -378,6 +378,7 @@ PYTHON_OBJS=	\
 		Python/pythonrun.o \
 		Python/pytime.o \
 		Python/bootstrap_hash.o \
+		Python/specialize.o \
 		Python/structmember.o \
 		Python/symtable.o \
 		Python/sysmodule.o \

Misc/NEWS.d/next/Core and Builtins/2021-05-20-12-43-04.bpo-44187.3lk0L1.rst

@@ -0,0 +1,3 @@
+Implement quickening in the interpreter. This offers no advantages as
+yet, but is an enabler of future optimizations. See PEP 659 for full
+explanation.

Objects/codeobject.c

@@ -211,6 +211,7 @@ init_code(PyCodeObject *co, struct _PyCodeConstructor *con)
 
     Py_INCREF(con->code);
     co->co_code = con->code;
+    co->co_firstinstr = (_Py_CODEUNIT *)PyBytes_AS_STRING(con->code);
     co->co_firstlineno = con->firstlineno;
     Py_INCREF(con->linetable);
     co->co_linetable = con->linetable;
@@ -250,6 +251,8 @@ init_code(PyCodeObject *co, struct _PyCodeConstructor *con)
     co->co_opcache = NULL;
     co->co_opcache_flag = 0;
     co->co_opcache_size = 0;
+    co->co_warmup = QUICKENING_INITIAL_WARMUP_VALUE;
+    co->co_quickened = NULL;
 }
 
 /* The caller is responsible for ensuring that the given data is valid. */
@@ -376,7 +379,8 @@ PyCode_NewWithPosOnlyArgs(int argcount, int posonlyargcount, int kwonlyargcount,
     if (_PyCode_Validate(&con) < 0) {
         return NULL;
     }
-
+    assert(PyBytes_GET_SIZE(code) % sizeof(_Py_CODEUNIT) == 0);
+    assert(_Py_IS_ALIGNED(PyBytes_AS_STRING(code), sizeof(_Py_CODEUNIT)));
     if (nlocals != PyTuple_GET_SIZE(varnames)) {
         PyErr_SetString(PyExc_ValueError,
                         "code: co_nlocals != len(co_varnames)");
@@ -1039,6 +1043,10 @@ code_dealloc(PyCodeObject *co)
         PyMem_Free(co->co_cell2arg);
     if (co->co_weakreflist != NULL)
         PyObject_ClearWeakRefs((PyObject*)co);
+    if (co->co_quickened) {
+        PyMem_Free(co->co_quickened);
+        _Py_QuickenedCount--;
+    }
     PyObject_Free(co);
 }
 

PCbuild/pythoncore.vcxproj

@@ -487,6 +487,7 @@
     <ClCompile Include="..\Python\dtoa.c" />
     <ClCompile Include="..\Python\Python-ast.c" />
     <ClCompile Include="..\Python\pythonrun.c" />
+    <ClCompile Include="..\Python\specialize.c" />
     <ClCompile Include="..\Python\suggestions.c" />
     <ClCompile Include="..\Python\structmember.c" />
     <ClCompile Include="..\Python\symtable.c" />

PCbuild/pythoncore.vcxproj.filters

@@ -1103,6 +1103,9 @@
     <ClCompile Include="..\Python\pythonrun.c">
       <Filter>Python</Filter>
     </ClCompile>
+    <ClCompile Include="..\Python\specialize.c">
+      <Filter>Python</Filter>
+    </ClCompile>
     <ClCompile Include="..\Python\structmember.c">
       <Filter>Python</Filter>
     </ClCompile>

Python/ceval.c

@@ -1343,6 +1343,14 @@ eval_frame_handle_pending(PyThreadState *tstate)
 #define JUMPTO(x)       (next_instr = first_instr + (x))
 #define JUMPBY(x)       (next_instr += (x))
 
+/* Get opcode and oparg from original instructions, not quickened form. */
+#define TRACING_NEXTOPARG() do { \
+        _Py_CODEUNIT word = ((_Py_CODEUNIT *)PyBytes_AS_STRING(co->co_code))[INSTR_OFFSET()]; \
+        opcode = _Py_OPCODE(word); \
+        oparg = _Py_OPARG(word); \
+        next_instr++; \
+    } while (0)
+
 /* OpCode prediction macros
     Some opcodes tend to come in pairs thus making it possible to
     predict the second code when the first is run.  For example,
@@ -1644,15 +1652,23 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, PyFrameObject *f, int throwflag)
     if (PyDTrace_FUNCTION_ENTRY_ENABLED())
         dtrace_function_entry(f);
 
+    /* Increment the warmup counter and quicken if warm enough
+     * _Py_Quicken is idempotent so we don't worry about overflow */
+    if (!PyCodeObject_IsWarmedUp(co)) {
+        PyCodeObject_IncrementWarmup(co);
+        if (PyCodeObject_IsWarmedUp(co)) {
+            if (_Py_Quicken(co)) {
+                goto exit_eval_frame;
+            }
+        }
+    }
+
+
     names = co->co_names;
     consts = co->co_consts;
     fastlocals = f->f_localsptr;
+    first_instr = co->co_firstinstr;
     freevars = f->f_localsptr + co->co_nlocals;
-    assert(PyBytes_Check(co->co_code));
-    assert(PyBytes_GET_SIZE(co->co_code) <= INT_MAX);
-    assert(PyBytes_GET_SIZE(co->co_code) % sizeof(_Py_CODEUNIT) == 0);
-    assert(_Py_IS_ALIGNED(PyBytes_AS_STRING(co->co_code), sizeof(_Py_CODEUNIT)));
-    first_instr = (_Py_CODEUNIT *) PyBytes_AS_STRING(co->co_code);
     /*
        f->f_lasti refers to the index of the last instruction,
        unless it's -1 in which case next_instr should be first_instr.
@@ -1757,7 +1773,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, PyFrameObject *f, int throwflag)
 
     tracing_dispatch:
         f->f_lasti = INSTR_OFFSET();
-        NEXTOPARG();
+        TRACING_NEXTOPARG();
 
         if (PyDTrace_LINE_ENABLED())
             maybe_dtrace_line(f, &trace_info);