Optimize _POP_CALL_TWO_LOAD_CONST_INLINE_BORROW

Author: tomasr8

Include/internal/pycore_uop_ids.h

@@ -5287,18 +5287,39 @@ dummy_func(
             value = PyStackRef_FromPyObjectBorrow(ptr);
         }
 
-        tier2 pure op (_POP_TOP_LOAD_CONST_INLINE_BORROW, (ptr/4, pop -- value)) {
+        tier2 op(_POP_CALL, (callable, null --)) {
+            (void)null; // Silence compiler warnings about unused variables
+            DEAD(null);
+            PyStackRef_CLOSE(callable);
+        }
+
+        tier2 op(_POP_CALL_ONE, (callable, null, pop --)) {
+            PyStackRef_CLOSE(pop);
+            (void)null; // Silence compiler warnings about unused variables
+            DEAD(null);
+            PyStackRef_CLOSE(callable);
+        }
+
+        tier2 op(_POP_CALL_TWO, (callable, null, pop1, pop2 --)) {
+            PyStackRef_CLOSE(pop2);
+            PyStackRef_CLOSE(pop1);
+            (void)null; // Silence compiler warnings about unused variables
+            DEAD(null);
+            PyStackRef_CLOSE(callable);
+        }
+
+        tier2 op(_POP_TOP_LOAD_CONST_INLINE_BORROW, (ptr/4, pop -- value)) {
             PyStackRef_CLOSE(pop);
             value = PyStackRef_FromPyObjectBorrow(ptr);
         }
 
-        tier2 pure op(_POP_TWO_LOAD_CONST_INLINE_BORROW, (ptr/4, pop1, pop2 -- value)) {
+        tier2 op(_POP_TWO_LOAD_CONST_INLINE_BORROW, (ptr/4, pop1, pop2 -- value)) {
             PyStackRef_CLOSE(pop2);
             PyStackRef_CLOSE(pop1);
             value = PyStackRef_FromPyObjectBorrow(ptr);
         }
 
-        tier2 pure op(_POP_CALL_TWO_LOAD_CONST_INLINE_BORROW, (ptr/4, callable, null, pop1, pop2 -- value)) {
+        tier2 op(_POP_CALL_TWO_LOAD_CONST_INLINE_BORROW, (ptr/4, callable, null, pop1, pop2 -- value)) {
             PyStackRef_CLOSE(pop2);
             PyStackRef_CLOSE(pop1);
             (void)null; // Silence compiler warnings about unused variables

Include/internal/pycore_uop_metadata.h

@@ -533,13 +533,17 @@ const uint16_t op_without_push[MAX_UOP_ID + 1] = {
     [_POP_TOP_LOAD_CONST_INLINE] = _POP_TOP,
     [_POP_TOP_LOAD_CONST_INLINE_BORROW] = _POP_TOP,
     [_POP_TWO_LOAD_CONST_INLINE_BORROW] = _POP_TWO,
+    [_POP_CALL_TWO_LOAD_CONST_INLINE_BORROW] = _POP_CALL_TWO,
 };
 
 const uint16_t op_without_pop[MAX_UOP_ID + 1] = {
     [_POP_TOP] = _NOP,
     [_POP_TOP_LOAD_CONST_INLINE] = _LOAD_CONST_INLINE,
     [_POP_TOP_LOAD_CONST_INLINE_BORROW] = _LOAD_CONST_INLINE_BORROW,
     [_POP_TWO_LOAD_CONST_INLINE_BORROW] = _POP_TOP_LOAD_CONST_INLINE_BORROW,
+    [_POP_CALL_TWO] = _POP_CALL_ONE,
+    [_POP_CALL_ONE] = _POP_CALL,
+    // [_POP_CALL] = _POP_TOP,  # TOS is NULL, this is handled separately in remove_unneeded_uops
 };
 
 
@@ -587,6 +591,16 @@ remove_unneeded_uops(_PyUOpInstruction *buffer, int buffer_size)
                     last->opcode = op_without_push[last->opcode];
                     opcode = buffer[pc].opcode = op_without_pop[opcode];
                 }
+                if (opcode == _POP_CALL) {
+                    _PyUOpInstruction *last = &buffer[pc - 1];
+                    while (last->opcode == _NOP) {
+                        last--;
+                    }
+                    if (last->opcode == _PUSH_NULL) {
+                        last->opcode = _NOP;
+                        opcode = buffer[pc].opcode = _POP_TOP;
+                    }
+                }
                 /* _PUSH_FRAME doesn't escape or error, but it
                  * does need the IP for the return address */
                 bool needs_ip = opcode == _PUSH_FRAME;