Thanks to visit codestin.com
Credit goes to github.com

Skip to content

gh-106581: Support multiple uops pushing new values #108895

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Sep 5, 2023
Merged

gh-106581: Support multiple uops pushing new values #108895

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
85e214a
Test for two uops that both push a new value
gvanrossum Aug 28, 2023
ab56dba
Refactor to use write_all_pokes() in two places
gvanrossum Sep 4, 2023
ddfda0e
Be strict about _{PUSH,POP}_FRAME, SAVE_CURRENT_IP
gvanrossum Sep 4, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 30 additions & 0 deletions Lib/test/test_generated_cases.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -532,6 +532,36 @@ def test_macro_cond_effect(self):
"""
self.run_cases_test(input, output)

def test_macro_push_push(self):
input = """
op(A, (-- val1)) {
val1 = spam();
}
op(B, (-- val2)) {
val2 = spam();
}
macro(M) = A + B;
"""
output = """
TARGET(M) {
PyObject *val1;
PyObject *val2;
// A
{
val1 = spam();
}
// B
{
val2 = spam();
}
STACK_GROW(2);
stack_pointer[-2] = val1;
stack_pointer[-1] = val2;
DISPATCH();
}
"""
self.run_cases_test(input, output)


if __name__ == "__main__":
unittest.main()
70 changes: 53 additions & 17 deletions Tools/cases_generator/stacking.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,15 +261,19 @@ def adjust_higher(self, eff: StackEffect) -> None:
self.final_offset.higher(eff)

def adjust(self, offset: StackOffset) -> None:
for down in offset.deep:
deep = list(offset.deep)
high = list(offset.high)
for down in deep:
self.adjust_deeper(down)
for up in offset.high:
for up in high:
self.adjust_higher(up)
Comment on lines 263 to 269
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just curious, why? I assume it has something to do with the comment you removed below (# Use clone() since adjust_inverse() mutates final_offset), but I'm not sure I understand.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Indeed. That .clone() call was needed because adjust_higher() and adjust_deeper() change self.final_offset (in place), and the argument here may be self.final_offset. (And ditto for min_offset.) So previously my "fix" for this was to pass in a clone of the argument. The new fix avoids the need to clone in the caller but effectively clones the deep and high lists of the argument.


def adjust_inverse(self, offset: StackOffset) -> None:
for down in offset.deep:
deep = list(offset.deep)
high = list(offset.high)
for down in deep:
self.adjust_higher(down)
for up in offset.high:
for up in high:
self.adjust_deeper(up)

def collect_vars(self) -> dict[str, StackEffect]:
Expand Down Expand Up @@ -426,15 +430,26 @@ def write_components(
)

if mgr.instr.name in ("_PUSH_FRAME", "_POP_FRAME"):
# Adjust stack to min_offset (input effects materialized)
# Adjust stack to min_offset.
# This means that all input effects of this instruction
# are materialized, but not its output effects.
# That's as intended, since these two are so special.
out.stack_adjust(mgr.min_offset.deep, mgr.min_offset.high)
# Use clone() since adjust_inverse() mutates final_offset
mgr.adjust_inverse(mgr.final_offset.clone())
# However, for tier 2, pretend the stack is at final offset.
mgr.adjust_inverse(mgr.final_offset)
if tier == TIER_ONE:
# TODO: Check in analyzer that _{PUSH,POP}_FRAME is last.
assert (
mgr is managers[-1]
), f"Expected {mgr.instr.name!r} to be the last uop"
assert_no_pokes(managers)

if mgr.instr.name == "SAVE_CURRENT_IP":
next_instr_is_set = True
if cache_offset:
out.emit(f"next_instr += {cache_offset};")
if tier == TIER_ONE:
assert_no_pokes(managers)
Comment on lines +445 to +452
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why are pokes bad for these?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because I don't want to think about what to do with loose pokes in macros containing SAVE_CURRENT_IP (i.e. some specializations of CALL), since next_instr_is_set causes the normal handling of loose pokes below to be skipped.

If and when we need to support that case, this assert will remind us (or at least me :-) that some additional effort is needed here. (Kind of like how the LOAD_FAST_PUSH_NULL use case reminded me to fix this issue here.)


if len(parts) == 1:
mgr.instr.write_body(out, 0, mgr.active_caches, tier)
Expand All @@ -443,19 +458,41 @@ def write_components(
mgr.instr.write_body(out, -4, mgr.active_caches, tier)

if mgr is managers[-1] and not next_instr_is_set:
# TODO: Explain why this adjustment is needed.
# Adjust the stack to its final depth, *then* write the
# pokes for all preceding uops.
# Note that for array output effects we may still write
# past the stack top.
Comment on lines +463 to +464
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems surprising to me. When could this happen?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

E.g. in UNPACK_SEQUENCE and its specializations. For array outputs, the variable is just a pointer, and it is initialized before the body of the opcode is executed. The body is supposed to write its output directly into the stack using this pointer. But the SP is not adjusted until after the body is executed. This is iffy only if you believe that something else might write over the portion of the stack above the SP while that code is running. But nothing ever does that.

out.stack_adjust(mgr.final_offset.deep, mgr.final_offset.high)
# Use clone() since adjust_inverse() mutates final_offset
mgr.adjust_inverse(mgr.final_offset.clone())
write_all_pokes(mgr.final_offset, managers, out)

return next_instr_is_set


def assert_no_pokes(managers: list[EffectManager]) -> None:
for mgr in managers:
for poke in mgr.pokes:
if not poke.effect.size and poke.effect.name not in mgr.instr.unmoved_names:
out.assign(
poke.as_stack_effect(),
poke.effect,
)
assert (
poke.effect.name == UNUSED
), f"Unexpected poke of {poke.effect.name} in {mgr.instr.name!r}"

return next_instr_is_set

def write_all_pokes(
offset: StackOffset, managers: list[EffectManager], out: Formatter
) -> None:
# Emit all remaining pushes (pokes)
for m in managers:
m.adjust_inverse(offset)
write_pokes(m, out)


def write_pokes(mgr: EffectManager, out: Formatter) -> None:
for poke in mgr.pokes:
if not poke.effect.size and poke.effect.name not in mgr.instr.unmoved_names:
out.assign(
poke.as_stack_effect(),
poke.effect,
)


def write_single_instr_for_abstract_interp(instr: Instruction, out: Formatter) -> None:
Expand All @@ -478,8 +515,7 @@ def _write_components_for_abstract_interp(
for mgr in managers:
if mgr is managers[-1]:
out.stack_adjust(mgr.final_offset.deep, mgr.final_offset.high)
# Use clone() since adjust_inverse() mutates final_offset
mgr.adjust_inverse(mgr.final_offset.clone())
mgr.adjust_inverse(mgr.final_offset)
# NULL out the output stack effects
for poke in mgr.pokes:
if not poke.effect.size and poke.effect.name not in mgr.instr.unmoved_names:
Expand Down