gh-133672: Allow LOAD_FAST to be optimized to LOAD_FAST_BORROW

Lib/test/test_peepholer.py

@@ -2430,14 +2430,14 @@ def test_optimized(self):
         ]
         self.check(insts, expected)
 
-    def test_unoptimized_if_unconsumed(self):
+    def test_optimize_unconsumed_when_is_safe(self):
         insts = [
             ("LOAD_FAST", 0, 1),
             ("LOAD_FAST", 1, 2),
             ("POP_TOP", None, 3),
         ]
         expected = [
-            ("LOAD_FAST", 0, 1),
+            ("LOAD_FAST_BORROW", 0, 1),
             ("LOAD_FAST_BORROW", 1, 2),
             ("POP_TOP", None, 3),
         ]
@@ -2449,7 +2449,7 @@ def test_unoptimized_if_unconsumed(self):
             ("POP_TOP", None, 3),
         ]
         expected = [
-            ("LOAD_FAST", 0, 1),
+            ("LOAD_FAST_BORROW", 0, 1),
             ("NOP", None, 2),
             ("NOP", None, 3),
         ]
@@ -2509,7 +2509,12 @@ def test_consume_some_inputs_no_outputs(self):
             ("GET_LEN", None, 2),
             ("LIST_APPEND", 0, 3),
         ]
-        self.check(insts, insts)
+        expected = [
+            ("LOAD_FAST_BORROW", 0, 1),
+            ("GET_LEN", None, 2),
+            ("LIST_APPEND", 0, 3),
+        ]
+        self.check(insts, expected)
 
     def test_check_exc_match(self):
         insts = [
@@ -2518,7 +2523,7 @@ def test_check_exc_match(self):
             ("CHECK_EXC_MATCH", None, 3)
         ]
         expected = [
-            ("LOAD_FAST", 0, 1),
+            ("LOAD_FAST_BORROW", 0, 1),
             ("LOAD_FAST_BORROW", 1, 2),
             ("CHECK_EXC_MATCH", None, 3)
         ]
@@ -2537,7 +2542,19 @@ def test_for_iter(self):
             ("LOAD_CONST", 0, 7),
             ("RETURN_VALUE", None, 8),
         ]
-        self.cfg_optimization_test(insts, insts, consts=[None])
+        expected = [
+            ("LOAD_FAST_BORROW", 0, 1),
+            top := self.Label(),
+            ("FOR_ITER", end := self.Label(), 2),
+            ("STORE_FAST", 2, 3),
+            ("JUMP", top, 4),
+            end,
+            ("END_FOR", None, 5),
+            ("POP_TOP", None, 6),
+            ("LOAD_CONST", 0, 7),
+            ("RETURN_VALUE", None, 8),
+        ]
+        self.cfg_optimization_test(insts, expected, consts=[None])
 
     def test_load_attr(self):
         insts = [
@@ -2556,7 +2573,7 @@ def test_load_attr(self):
             ("LOAD_ATTR", 1, 2),
         ]
         expected = [
-            ("LOAD_FAST", 0, 1),
+            ("LOAD_FAST_BORROW", 0, 1),
             ("LOAD_ATTR", 1, 2),
         ]
         self.check(insts, expected)
@@ -2586,7 +2603,7 @@ def test_super_attr(self):
         expected = [
             ("LOAD_FAST_BORROW", 0, 1),
             ("LOAD_FAST_BORROW", 1, 2),
-            ("LOAD_FAST", 2, 3),
+            ("LOAD_FAST_BORROW", 2, 3),
             ("LOAD_SUPER_ATTR", 1, 4),
         ]
         self.check(insts, expected)
@@ -2603,7 +2620,7 @@ def test_send(self):
             ("RETURN_VALUE", None, 7)
         ]
         expected = [
-            ("LOAD_FAST", 0, 1),
+            ("LOAD_FAST_BORROW", 0, 1),
             ("LOAD_FAST_BORROW", 1, 2),
             ("SEND", end := self.Label(), 3),
             ("LOAD_CONST", 0, 4),

Misc/NEWS.d/next/Core_and_Builtins/2025-05-09-00-06-52.gh-issue-133672.y6QwwH.rst

@@ -0,0 +1 @@
+The peephole optimizer now converts LOAD_FAST to LOAD_FAST_BORROW more aggressively. This optimization is applied even if the loaded variable remains live on the stack at the end of a basic block, as long as the borrow is otherwise determined to be safe. This can reduce reference counting overhead in certain code patterns, such as with iterables in loops.

Python/flowgraph.c

@@ -2976,7 +2976,7 @@ optimize_load_fast(cfg_builder *g)
 
         // Optimize instructions
         for (int i = 0; i < block->b_iused; i++) {
-            if (!instr_flags[i]) {
+            if (!(instr_flags[i] & (SUPPORT_KILLED | STORED_AS_LOCAL))) {
                 cfg_instr *instr = &block->b_instr[i];
                 switch (instr->i_opcode) {
                     case LOAD_FAST: