JIT: Remove redundant branches to jump in the assembly optimizer

* Refactor JIT assembly optimizer making instructions instances not just strings
* Remove redundant jumps and branches where legal to do so
* Modifies _BINARY_OP_SUBSCR_STR_INT to avoid excessive inlining depth

TEST_TEMPLATE: projects/python/templates/default
PYTHON_OPTIONS: --configure-flags "--enable-experimental-jit"

Jira: ENTLLT-8874
Change-Id: I605618c92814059bc63a1d16825a66d1824f3df2

Author: markshannon

Python/bytecodes.c

@@ -982,9 +982,10 @@ dummy_func(
             DEOPT_IF(!_PyLong_IsNonNegativeCompact((PyLongObject *)sub));
             Py_ssize_t index = ((PyLongObject*)sub)->long_value.ob_digit[0];
             DEOPT_IF(PyUnicode_GET_LENGTH(str) <= index);
-            // Specialize for reading an ASCII character from any string:
-            Py_UCS4 c = PyUnicode_READ_CHAR(str, index);
-            DEOPT_IF(Py_ARRAY_LENGTH(_Py_SINGLETON(strings).ascii) <= c);
+            // Specialize for reading an ASCII character from an ASCII string:
+            DEOPT_IF(!PyUnicode_IS_COMPACT_ASCII(str));
+            uint8_t c = PyUnicode_1BYTE_DATA(str)[index];
+            assert(c < 128);
             STAT_INC(BINARY_OP, hit);
             PyObject *res_o = (PyObject*)&_Py_SINGLETON(strings).ascii[c];
             PyStackRef_CLOSE_SPECIALIZED(sub_st, _PyLong_ExactDealloc);

Python/executor_cases.c.h

@@ -181,6 +181,7 @@ set_bits(uint32_t *loc, uint8_t loc_start, uint64_t value, uint8_t value_start,
 #define IS_AARCH64_ADRP(I)        (((I) & 0x9F000000) == 0x90000000)
 #define IS_AARCH64_BRANCH(I)      (((I) & 0x7C000000) == 0x14000000)
 #define IS_AARCH64_BRANCH_COND(I) (((I) & 0x7C000000) == 0x54000000)
+#define IS_AARCH64_BRANCH_ZERO(I) (((I) & 0x7E000000) == 0x34000000)
 #define IS_AARCH64_TEST_AND_BRANCH(I) (((I) & 0x7E000000) == 0x36000000)
 #define IS_AARCH64_LDR_OR_STR(I)  (((I) & 0x3B000000) == 0x39000000)
 #define IS_AARCH64_MOV(I)         (((I) & 0x9F800000) == 0x92800000)
@@ -348,7 +349,7 @@ void
 patch_aarch64_19r(unsigned char *location, uint64_t value)
 {
     uint32_t *loc32 = (uint32_t *)location;
-    assert(IS_AARCH64_BRANCH_COND(*loc32));
+    assert(IS_AARCH64_BRANCH_COND(*loc32) || IS_AARCH64_BRANCH_ZERO(*loc32));
     value -= (uintptr_t)location;
     // Check that we're not out of range of 21 signed bits:
     assert((int64_t)value >= -(1 << 20));

Python/generated_cases.c.h

@@ -614,6 +614,8 @@ def has_error_without_pop(op: parser.CodeDef) -> bool:
     "PyUnicode_Concat",
     "PyUnicode_GET_LENGTH",
     "PyUnicode_READ_CHAR",
+    "PyUnicode_IS_COMPACT_ASCII",
+    "PyUnicode_1BYTE_DATA",
     "Py_ARRAY_LENGTH",
     "Py_FatalError",
     "Py_INCREF",