fix: make fast-math toggles work #140

omsherikar · 2025-11-22T16:56:55Z

Pull Request description

This PR fixes issue #136 by implementing the missing set_fast_math method and adding support for applying fast-math flags to floating-point instructions. Previously, calls to self.set_fast_math(True/False) in visit(BinaryOp) would crash because the method did not exist in LLVMLiteIRVisitor.

The fix includes:

Implementation of set_fast_math(enabled: bool) to track fast-math state
New helper _apply_fast_math(inst) to conditionally attach the "fast" flag to floating-point instructions
Integration of fast-math flag application into all floating-point operations (fadd, fsub, fmul, fdiv, fma)
Unit tests to verify fast-math flags are correctly applied when enabled

Solve #136

How to test these changes

Run makim tests.unit to execute the unit test suite
The new test test_set_fast_math_marks_float_ops verifies that fast-math flags are applied to floating-point operations when set_fast_math(True) is called
Verify that fast-math flags are not applied when set_fast_math(False) is called

Pull Request checklists

This PR is a:

bug-fix
new feature
maintenance

About this PR:

it includes tests.
the tests are executed on CI.
the tests generate log file(s) (path).
pre-commit hooks were executed locally.
this PR requires a project documentation update.

Author's checklist:

I have reviewed the changes and it contains no misspelling.
The code is well commented, especially in the parts that contain more
complexity.
New and old tests passed locally.

Additional information

The implementation tracks fast-math state via self._fast_math_enabled and applies flags only to floating-point instructions (both scalar and vector). The _apply_fast_math helper checks if the instruction type is a floating-point type before attaching the "fast" flag, ensuring that integer operations are not affected.

Reviewer's checklist

Copy and paste this template for your review's note:

## Reviewer's Checklist

- [ ] I managed to reproduce the problem locally from the `main` branch
- [ ] I managed to test the new changes locally
- [ ] I confirm that the issues mentioned were fixed/resolved .

github-actions · 2025-11-22T17:00:28Z

OSL ChatGPT Reviewer

NOTE: This is generated by an AI program, so some comments may not make sense.

src/irx/builders/llvmliteir.py

Risk: Applying fast-math via inst.flags on call instructions (e.g., llvm.fma) may be unsupported or generate invalid IR; many llvmlite call instructions don’t expose flags and will raise AttributeError at runtime. Make _apply_fast_math robust and only set flags when supported. Suggested change (L.441):

def _apply_fast_math(self, inst: ir.Instruction) -> None:
"""Attach fast-math flags when enabled and applicable."""
if not self._fast_math_enabled:
return
ty = inst.type
if isinstance(ty, ir.VectorType):
if not is_fp_type(ty.element):
return
elif not is_fp_type(ty):
return
# Prefer instruction-level fastmath API if present
if hasattr(inst, "fastmath"):
setattr(inst, "fastmath", "fast")
return
flags = getattr(inst, "flags", None)
if isinstance(flags, list) and "fast" not in flags:
flags.append("fast")
Maintainability/coverage: You’re manually touching many FP ops; consider also toggling the builder’s global fastmath to avoid misses and ensure future FP ops inherit it. Suggested change (L.429):

def set_fast_math(self, enabled: bool) -> None:
"""Enable/disable fast-math flags for subsequent FP instructions."""
self._fast_math_enabled = enabled
# Also set builder-level fastmath if available to cover future ops
fm = "fast" if enabled else None
ib = getattr(getattr(self, "_llvm", None), "ir_builder", None)
if ib is not None and hasattr(ib, "fastmath"):
setattr(ib, "fastmath", fm)

tests/test_llvmlite_helpers.py

The assertions rely on "fast" being present in inst.flags, which may not hold across llvmlite versions (some expose fast-math via inst.fastmath or expand the composite into granular flags like nnan/ninf/nsz/arcp/contract/afn/reassoc). Make the check version-tolerant and accept either the composite or granular flags.
Suggested change (L.102-L.105):
def _get_fp_flags(inst: object) -> tuple[str, ...]:
"""Return fast-math or flags in a version-tolerant way."""
fm = getattr(inst, "fastmath", None)
if fm:
return tuple(fm)
return tuple(getattr(inst, "flags", ()))
Update assertions to use the helper (L.110, L.115):
flags_fast = _get_fp_flags(inst_fast)
assert any(f in flags_fast for f in ("fast", "nnan", "ninf", "nsz", "arcp", "contract", "afn", "reassoc"))

flags_normal = _get_fp_flags(inst_normal)
assert all(f not in flags_normal for f in ("fast", "nnan", "ninf", "nsz", "arcp", "contract", "afn", "reassoc"))
This test reaches into private internals (visitor._llvm.ir_builder, visitor._apply_fast_math). That’s brittle and couples tests to implementation details. If possible, prefer asserting via a public API that emits the instruction with the expected flags, or add a small public helper to apply/check fast-math on an instruction.

omsherikar · 2025-11-22T17:40:24Z

@xmnlab @yuvimittal Please review this

xmnlab · 2025-11-25T03:52:58Z

hi @omsherikar , thanks for working on that.
I will review it this week. thanks for your patience.

omsherikar · 2025-11-25T12:31:12Z

hi @omsherikar , thanks for working on that. I will review it this week. thanks for your patience.

Sure @xmnlab, Thank you!

omsherikar · 2025-11-30T15:48:43Z

@xmnlab Please have a look into it

omsherikar · 2025-12-05T16:43:25Z

@yuvimittal Have a look at it

yuvimittal · 2025-12-11T15:28:24Z

src/irx/builders/llvmliteir.py

+        """Attach fast-math flags when enabled and applicable."""
+        if not self._fast_math_enabled:
+            return
+        ty = inst.type


inst.flags might be None, immutable or absent, so it can cause attributeError or TypeError.

Thanks for the feedback @yuvimittal. and sorry for the delay .Should I handle it with a try-except block?

I've added defensive handling using getattr and try/except so it won't crash if flags is None or immutable.

github-actions · 2025-12-11T16:45:03Z

OSL ChatGPT Reviewer

NOTE: This is generated by an AI program, so some comments may not make sense.

src/irx/builders/llvmliteir.py

High-risk bug: using inst.flags to attach "fast" is incorrect for FP ops in llvmlite and may either be ignored or break IR verification. Floating-point fast-math must use FastMathFlags via the inst.fastmath attribute. Replace the tail of _apply_fast_math accordingly. (L.439)

def _apply_fast_math(self, inst: ir.Instruction) -> None:
"""Attach fast-math flags when enabled and applicable."""
if not self._fast_math_enabled:
return
ty = inst.type
if isinstance(ty, ir.VectorType):
if not is_fp_type(ty.element):
return
elif not is_fp_type(ty):
return
fm = ir.FastMathFlags()
fm.set_fast()
inst.fastmath = fm
Optional simplification with broader coverage: set/clear builder-level fastmath once in set_fast_math to automatically apply to subsequent FP instructions (including calls to intrinsics), reducing the need to remember _apply_fast_math at each site. (L.428)

def set_fast_math(self, enabled: bool) -> None:
"""Enable/disable fast-math flags for subsequent FP instructions."""
self._fast_math_enabled = enabled
if enabled:
fm = ir.FastMathFlags()
fm.set_fast()
self._llvm.ir_builder.fastmath = fm
else:
self._llvm.ir_builder.fastmath = None

tests/test_llvmlite_helpers.py

Tests rely on private internals (_llvm, FLOAT_TYPE, ir_builder, _apply_fast_math), which makes them brittle to refactors. At minimum, avoid the private type constant. Use llvmlite’s public type instead (L.104):
float_ty: ir.types.FloatType = ir.FloatType()
Asserting on inst.flags is likely to be fragile across llvmlite versions. Assert on the emitted IR instead to reduce coupling (L.111, L.117):
assert "fast" in str(inst_fast)
assert "fast" not in str(inst_normal)

Copilot

Pull request overview

This PR fixes issue #136 by implementing fast-math flag support for floating-point operations in the LLVM IR builder. The implementation adds a set_fast_math method to track fast-math state and a helper to conditionally apply the "fast" flag to floating-point instructions.

Key Changes:

Added set_fast_math(enabled: bool) method and _fast_math_enabled instance variable to track fast-math state
Implemented _apply_fast_math(inst) helper that checks instruction types and applies flags appropriately
Integrated fast-math flag application into all floating-point operations (fadd, fsub, fmul, fdiv, fma)

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/irx/builders/llvmliteir.py	Adds fast-math state tracking, helper method, and integrates flag application into FP operations
tests/test_llvmlite_helpers.py	Adds unit test verifying fast-math flags are correctly applied when enabled and absent when disabled

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

src/irx/builders/llvmliteir.py

tests/test_llvmlite_helpers.py

github-actions · 2025-12-16T15:50:53Z

OSL ChatGPT Reviewer

NOTE: This is generated by an AI program, so some comments may not make sense.

src/irx/builders/llvmliteir.py

High risk: _apply_fast_math mutates inst.flags with "fast". In llvmlite, FP fast-math is controlled via the instruction’s fastmath attribute, not generic flags; appending "fast" to flags may be ignored by FP ops and calls, so fast-math likely never takes effect. Use inst.fastmath = "fast" (and only fall back to flags if you must for legacy). Suggested replacement (L.446):

def _apply_fast_math(self, inst: ir.Instruction) -> None:
"""Attach fast-math flags when enabled and applicable."""
if not self._fast_math_enabled:
return
ty = inst.type
if isinstance(ty, ir.VectorType):
if not is_fp_type(ty.element):
return
elif not is_fp_type(ty):
return
if hasattr(inst, "fastmath"):
inst.fastmath = "fast"
else:
flags = getattr(inst, "flags", None)
if isinstance(flags, list) and "fast" not in flags:
flags.append("fast")

tests/test_llvmlite_helpers.py

LGTM!

…mall try/except

github-actions · 2025-12-16T15:57:43Z

OSL ChatGPT Reviewer

NOTE: This is generated by an AI program, so some comments may not make sense.

src/irx/builders/llvmliteir.py

Bug: fast-math flags are applied to inst.flags, but in llvmlite FP ops/calls use the fastmath attribute, not flags. As written, this likely no-ops for fadd/fsub/fmul/fdiv/call. Please set inst.fastmath (and assign, not just mutate) and handle vector element attr robustly. Suggested fix:

(L.435)
def _apply_fast_math(self, inst: ir.Instruction) -> None:
"""Attach fast-math flags when enabled and applicable."""
if not self._fast_math_enabled:
return
ty = inst.type
if isinstance(ty, ir.VectorType):
elem = getattr(ty, "element", getattr(ty, "element_type", None))
if elem is None or not is_fp_type(elem):
return
elif not is_fp_type(ty):
return
fm = getattr(inst, "fastmath", None)
if fm is None:
return
if isinstance(fm, (list, tuple)):
if "fast" not in fm:
inst.fastmath = list(fm) + ["fast"]
else:
inst.fastmath = ["fast"]

tests/test_llvmlite_helpers.py

The test is brittle by asserting the umbrella flag string "fast". If the implementation switches to specific FMF subsets (nnan/ninf/nsz/arcp/contract/reassoc), this will fail despite correct behavior. Suggest relaxing to “any FMF present” and guarding for llvmlite variants that expose flags differently. (L.110, L.115)
Replace:
assert "fast" in inst_fast.flags
assert "fast" not in inst_normal.flags
With:
fmf: set[str] = {"fast", "nnan", "ninf", "nsz", "arcp", "contract", "reassoc"} # type: ignore[var-annotated]
assert bool(fmf & set(getattr(inst_fast, "flags", getattr(inst_fast, "fastmath", []))))
assert not bool(fmf & set(getattr(inst_normal, "flags", getattr(inst_normal, "fastmath", []))))
The test couples to private internals (visitor._llvm.FLOAT_TYPE, visitor._llvm.ir_builder, visitor._apply_fast_math). This increases fragility and can break with internal refactors. If possible, prefer public APIs, e.g., use ir.FloatType() over the private constant. (L.103)
Replace:
float_ty = visitor._llvm.FLOAT_TYPE
With:
float_ty: ir.Type = ir.FloatType() # """Use public llvmlite type to avoid private coupling."""

omsherikar · 2025-12-16T16:03:19Z

@yuvimittal Please review the changes

yuvimittal reviewed Dec 11, 2025

View reviewed changes

Copilot AI review requested due to automatic review settings December 16, 2025 15:47

Copilot AI reviewed Dec 16, 2025

View reviewed changes

src/irx/builders/llvmliteir.py Show resolved Hide resolved

src/irx/builders/llvmliteir.py Show resolved Hide resolved

tests/test_llvmlite_helpers.py Show resolved Hide resolved

omsherikar added 2 commits December 16, 2025 21:23

fix: make fast-math toggles work

7880ec7

Added a getattr guard around inst.flags and wrapped the append in a s…

843fe25

…mall try/except

omsherikar force-pushed the fix/fast-math-toggle branch from 66644bd to 843fe25 Compare December 16, 2025 15:53

omsherikar requested a review from yuvimittal December 16, 2025 15:56

fix: make fast-math toggles work #140

Are you sure you want to change the base?

fix: make fast-math toggles work #140

Uh oh!

Conversation

omsherikar commented Nov 22, 2025

Pull Request description

How to test these changes

Pull Request checklists

Additional information

Reviewer's checklist

Uh oh!

github-actions bot commented Nov 22, 2025

OSL ChatGPT Reviewer

src/irx/builders/llvmliteir.py

tests/test_llvmlite_helpers.py

Uh oh!

omsherikar commented Nov 22, 2025

Uh oh!

xmnlab commented Nov 25, 2025

Uh oh!

omsherikar commented Nov 25, 2025

Uh oh!

omsherikar commented Nov 30, 2025

Uh oh!

omsherikar commented Dec 5, 2025

Uh oh!

yuvimittal Dec 11, 2025

Choose a reason for hiding this comment

Uh oh!

omsherikar Dec 15, 2025

Choose a reason for hiding this comment

Uh oh!

omsherikar Dec 16, 2025

Choose a reason for hiding this comment

Uh oh!

github-actions bot commented Dec 11, 2025

OSL ChatGPT Reviewer

src/irx/builders/llvmliteir.py

tests/test_llvmlite_helpers.py

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Dec 16, 2025

OSL ChatGPT Reviewer

src/irx/builders/llvmliteir.py

tests/test_llvmlite_helpers.py

Uh oh!

github-actions bot commented Dec 16, 2025

OSL ChatGPT Reviewer

src/irx/builders/llvmliteir.py

tests/test_llvmlite_helpers.py

Uh oh!

omsherikar commented Dec 16, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants