Add MAX_GENERATED_CACHED_REGISTER in analyzer

Author: aisk

Tools/cases_generator/analyzer.py

@@ -7,7 +7,7 @@
 
 from parser import Stmt, SimpleStmt, BlockStmt, IfStmt, WhileStmt, ForStmt, MacroIfStmt
 
-MAX_CACHED_REGISTER = 3
+MAX_GENERATED_CACHED_REGISTER = 5
 
 @dataclass
 class EscapingCall:
@@ -1369,13 +1369,13 @@ def is_large(uop: Uop) -> bool:
 
 def get_uop_cache_depths(uop: Uop) -> Iterator[tuple[int, int, int]]:
     if uop.name == "_SPILL_OR_RELOAD":
-        for inputs in range(MAX_CACHED_REGISTER+1):
-            for outputs in range(MAX_CACHED_REGISTER+1):
+        for inputs in range(MAX_GENERATED_CACHED_REGISTER + 1):
+            for outputs in range(MAX_GENERATED_CACHED_REGISTER + 1):
                 if inputs != outputs:
                     yield inputs, outputs, inputs
         return
     if uop.name in ("_DEOPT", "_HANDLE_PENDING_AND_DEOPT", "_EXIT_TRACE", "_DYNAMIC_EXIT"):
-        for i in range(MAX_CACHED_REGISTER+1):
+        for i in range(MAX_GENERATED_CACHED_REGISTER + 1):
             yield i, 0, 0
         return
     if uop.name in ("_START_EXECUTOR", "_JUMP_TO_TOP", "_COLD_EXIT"):
@@ -1397,20 +1397,20 @@ def get_uop_cache_depths(uop: Uop) -> Iterator[tuple[int, int, int]]:
             has_array = True
             break
         ideal_outputs += 1
-    if ideal_inputs > MAX_CACHED_REGISTER:
-        ideal_inputs = MAX_CACHED_REGISTER
-    if ideal_outputs > MAX_CACHED_REGISTER:
-        ideal_outputs = MAX_CACHED_REGISTER
+    if ideal_inputs > MAX_GENERATED_CACHED_REGISTER:
+        ideal_inputs = MAX_GENERATED_CACHED_REGISTER
+    if ideal_outputs > MAX_GENERATED_CACHED_REGISTER:
+        ideal_outputs = MAX_GENERATED_CACHED_REGISTER
     at_end = uop.properties.sync_sp or uop.properties.side_exit_at_end
     exit_depth = ideal_outputs if at_end else ideal_inputs
     if uop.properties.escapes or uop.properties.sync_sp or has_array or is_large(uop):
         yield ideal_inputs, ideal_outputs, exit_depth
         return
-    for inputs in range(MAX_CACHED_REGISTER + 1):
+    for inputs in range(MAX_GENERATED_CACHED_REGISTER + 1):
         outputs = ideal_outputs - ideal_inputs + inputs
         if outputs < ideal_outputs:
             outputs = ideal_outputs
-        elif outputs > MAX_CACHED_REGISTER:
+        elif outputs > MAX_GENERATED_CACHED_REGISTER:
             continue
         yield inputs, outputs, outputs if at_end else inputs
 

Tools/cases_generator/tier2_generator.py

@@ -16,7 +16,7 @@
     analysis_error,
     get_uop_cache_depths,
     is_large,
-    MAX_CACHED_REGISTER,
+    MAX_GENERATED_CACHED_REGISTER,
 )
 
 from generators_common import (
@@ -202,8 +202,10 @@ def cache_items(self, stack: Stack, cached_items: int, zero_regs: bool) -> None:
             # replace this with a "clobber" to tell
             # the compiler that these values are unused
             # without having to emit any code.
-            for i in range(cached_items, MAX_CACHED_REGISTER):
+            for i in range(cached_items, MAX_GENERATED_CACHED_REGISTER):
+                self.out.emit(f"#if MAX_CACHED_REGISTER >= {i + 1}\n")
                 self.out.emit(f"_tos_cache{i} = PyStackRef_ZERO_BITS;\n")
+                self.out.emit("#endif\n")
         self.emit(f"SET_CURRENT_CACHED_VALUES({cached_items});\n")
 
 
@@ -277,6 +279,9 @@ def generate_tier2(
             continue
         for inputs, outputs, exit_depth in get_uop_cache_depths(uop):
             emitter = Tier2Emitter(out, analysis.labels, exit_depth)
+            needed_cached_registers = max(inputs, outputs)
+            if needed_cached_registers:
+                out.emit(f"#if MAX_CACHED_REGISTER >= {needed_cached_registers}\n")
             out.emit(f"case {uop.name}_r{inputs}{outputs}: {{\n")
             out.emit(f"CHECK_CURRENT_CACHED_VALUES({inputs});\n")
             out.emit("assert(WITHIN_STACK_BOUNDS_IGNORING_CACHE());\n")
@@ -293,6 +298,8 @@ def generate_tier2(
             out.start_line()
             out.emit("}")
             out.emit("\n\n")
+            if needed_cached_registers:
+                out.emit("#endif\n\n")
     out.emit("\n")
     outfile.write("#undef TIER_TWO\n")
 

Tools/cases_generator/uop_metadata_generator.py

@@ -10,7 +10,6 @@
     analyze_files,
     get_uop_cache_depths,
     Uop,
-    MAX_CACHED_REGISTER,
 )
 from generators_common import (
     DEFAULT_INPUT,
@@ -23,6 +22,7 @@
 from typing import TextIO
 
 DEFAULT_OUTPUT = ROOT / "Include/internal/pycore_uop_metadata.h"
+MAX_CACHED_REGISTER = 3  # Specify this by different platform
 
 
 def uop_cache_info(uop: Uop) -> list[str] | None:
@@ -37,6 +37,8 @@ def uop_cache_info(uop: Uop) -> list[str] | None:
     high = -1
     defined = [ False ] * 4
     for inputs, outputs, exit_depth in get_uop_cache_depths(uop):
+        if max(inputs, outputs) > MAX_CACHED_REGISTER:
+            continue
         entries[inputs] = f"{{ {outputs}, {exit_depth}, {uop.name}_r{inputs}{outputs} }},\n"
         if inputs < low:
             low = inputs