Port internal self-tests out of libHalide into test/correctness

[alexreinking]

Halide has historically embedded self-test functions (e.g. Foo::test(), foo_test()) directly in src/*.cpp, invoked from a single test/internal.cpp binary. This moves every one of those test bodies (and their exclusive helper functions) out of libHalide entirely, into standalone test/correctness/*.cpp executables that exercise the same internal APIs via Halide::Internal::, following the same conventions as the rest of the correctness suite.

test/internal.cpp and the _test_internal test target are removed; a trivial test/pch_helper.cpp takes over _test_internal's role as the precompiled-header donor for the other test targets.

Two internal APIs needed small additions to support this:

• CodeGen_C.h gained three accessor functions (codegen_c_test_*) since the binary2cpp-generated blobs its test compares against aren't exported from the shared library.
• spirv_ir.cpp gets its own CMake wiring for the internal-only SpirvIR.h header and vendored SPIR-V headers.

Breaking changes

None

Checklist

• Tests added or updated (not required for docs, CI config, or typo fixes)
• Commits include AI attribution where applicable (see Code of Conduct)

[alexreinking]

Fidelity check: did the test bodies actually move unchanged?

Since each ported test's body still exists as pure deletions in the still-present src/*.cpp file, and each new test/correctness/*.cpp file is 100% new (so its content is pure additions), git diff between this commit (0cb496c08) and its parent (734f678e1) gives an exact before/after pair for every port. I stripped the mechanical porting scaffolding (#include "Halide.h", using namespace ..., the main() wrapper, printf("Success!\n")) from both sides and diffed the remainder with whitespace ignored.

Result: 94.7% of the ported content (1864 / 1968 lines) is identical, modulo whitespace.

src/*.cpp (old)	test/correctness/*.cpp (new)	old	new	equal	match %
IRPrinter.cpp	ir_printer.cpp	45	43	39	86.7%
CodeGen_C.cpp	codegen_c.cpp	158	188	147	78.2%
IREquality.cpp	ir_equality.cpp	66	33	25	37.9%
Bounds.cpp	bounds_internal.cpp	413	424	410	96.7%
IRMatch.cpp	expr_match.cpp	27	25	24	88.9%
Deinterleave.cpp	deinterleave_vector.cpp	30	28	28	93.3%
ModulusRemainder.cpp	modulus_remainder.cpp	25	23	23	92.0%
CSE.cpp	cse.cpp	136	133	128	94.1%
CPlusPlusMangle.cpp	cplusplus_mangle.cpp	404	403	402	99.5%
Monotonic.cpp	is_monotonic.cpp	68	66	66	97.1%
Reduction.cpp	split_predicate.cpp	52	52	50	96.2%
Associativity.cpp	associativity.cpp	260	261	257	98.5%
Generator.cpp	generator_internal.cpp	152	154	152	98.7%
AutoScheduleUtils.cpp	propagate_estimate.cpp	22	20	19	86.4%
UniquifyVariableNames.cpp	uniquify_variable_names.cpp	69	64	62	89.9%
SpirvIR.cpp	spirv_ir.cpp	41	38	32	78.0%
TOTAL		1968	1955	1864	94.7%

I inspected the residual (non-matching) lines for every file by hand. They fall into exactly two buckets:

1. Purely mechanical. The old .cpp files had file-scope using std::vector; / using std::string; / etc., so the moved code used bare vector/string/ostringstream. The standalone test files don't carry those using declarations over, so the same lines now read std::vector/std::string/std::ostringstream. This accounts for the bulk of the "residual" diff in ir_printer.cpp, expr_match.cpp, and uniquify_variable_names.cpp.

2. Deliberate, disclosed changes (each has an explanatory comment in the new file):

• ir_equality.cpp (37.9%, the lowest match) — the original used a private Comparer<N> template and Order enum that only existed in IREquality.cpp's own anonymous namespace, unreachable from outside that translation unit. Rewritten to use the public graph_equal/graph_less_than functions instead, preserving the same 3 test cases and the same intent (equality check, plus an antisymmetric-ordering check for the "hangs if comparison is exponential" case).
• codegen_c.cpp (78.2%) — the 3 binary2cpp-generated blob symbols (halide_c_template_CodeGen_C_prologue, etc.) it compares generated output against turned out not to be exported from the shared library (Halide's linker export script only allows halide_* runtime-API symbols and mangled Halide:: C++ symbols — not raw data blobs). Added 3 small Halide::Internal::codegen_c_test_* accessor functions to CodeGen_C.h/.cpp and call those instead.
• spirv_ir.cpp (78.0%) — dropped the dead #ifdef WITH_SPIRV / "SpirV IR test disabled" fallback branch, since the new test target always defines WITH_SPIRV; also swapped bare assert() for internal_assert() to match Halide's convention.
• bounds_internal.cpp (96.7%, only mentioning for the one real addition) — reimplemented a small private Interval simplify(const Interval&) overload that lived in Bounds.cpp's own anonymous namespace, since it isn't part of the public API surface.

Everything else — Deinterleave.cpp, ModulusRemainder.cpp, CSE.cpp, CPlusPlusMangle.cpp, Monotonic.cpp, Reduction.cpp, Associativity.cpp, Generator.cpp, AutoScheduleUtils.cpp — moved with zero behavioral change, only the mechanical main()/printf("Success!\n") wrapping.

Script and commands used to produce this (click to expand)

Run from the root of a checkout that has both commits in its history:

python3 compare_ports.py          # summary table (shown above)
python3 compare_ports.py --full   # + full unified diff per file, for spot-checking

compare_ports.py:

#!/usr/bin/env python3
"""Correlate deleted src/*.cpp test bodies with their new test/correctness/*.cpp
homes, and report how much of the diff is "real" content change vs mechanical
renaming/whitespace, by diffing the two blocks of text with whitespace ignored.

Run from the root of a Halide checkout that has both OLD_REF and NEW_REF in
its history (e.g. after `git fetch origin pull/9276/head`):

    python3 compare_ports.py          # summary table
    python3 compare_ports.py --full   # + full unified diff per file
"""
import difflib
import re
import subprocess
import sys

OLD_REF = "734f678e1"  # parent of the porting commit
NEW_REF = "0cb496c08"  # the porting commit itself

PAIRS = [
    ("src/IRPrinter.cpp", "test/correctness/ir_printer.cpp"),
    ("src/CodeGen_C.cpp", "test/correctness/codegen_c.cpp"),
    ("src/IREquality.cpp", "test/correctness/ir_equality.cpp"),
    ("src/Bounds.cpp", "test/correctness/bounds_internal.cpp"),
    ("src/IRMatch.cpp", "test/correctness/expr_match.cpp"),
    ("src/Deinterleave.cpp", "test/correctness/deinterleave_vector.cpp"),
    ("src/ModulusRemainder.cpp", "test/correctness/modulus_remainder.cpp"),
    ("src/CSE.cpp", "test/correctness/cse.cpp"),
    ("src/CPlusPlusMangle.cpp", "test/correctness/cplusplus_mangle.cpp"),
    ("src/Monotonic.cpp", "test/correctness/is_monotonic.cpp"),
    ("src/Reduction.cpp", "test/correctness/split_predicate.cpp"),
    ("src/Associativity.cpp", "test/correctness/associativity.cpp"),
    ("src/Generator.cpp", "test/correctness/generator_internal.cpp"),
    ("src/AutoScheduleUtils.cpp", "test/correctness/propagate_estimate.cpp"),
    ("src/UniquifyVariableNames.cpp", "test/correctness/uniquify_variable_names.cpp"),
    ("src/SpirvIR.cpp", "test/correctness/spirv_ir.cpp"),
]


def git_diff_lines(path, sign):
    """Return the lines added ('+') or removed ('-') for `path` between
    OLD_REF and NEW_REF, stripped of the diff marker, in original order."""
    out = subprocess.run(
        ["git", "diff", "-U0", "--no-color", OLD_REF, NEW_REF, "--", path],
        capture_output=True, text=True, check=True
    ).stdout
    lines = []
    for line in out.splitlines():
        if line.startswith("@@") or line.startswith("diff ") or line.startswith("index "):
            continue
        if line.startswith("--- ") or line.startswith("+++ "):
            continue
        if sign == "-" and line.startswith("-"):
            lines.append(line[1:])
        elif sign == "+" and line.startswith("+"):
            lines.append(line[1:])
    return lines


# Lines that are pure "porting scaffolding" and shouldn't count against fidelity.
BOILERPLATE = re.compile(
    r'^\s*('
    r'#include "Halide\.h"|'
    r'using namespace Halide;|'
    r'using namespace Halide::Internal;|'
    r'namespace Halide \{|'
    r'namespace Internal \{|'
    r'\}\s*//\s*namespace Internal|'
    r'\}\s*//\s*namespace Halide|'
    r'int main\(.*\)\s*\{|'
    r'printf\("Success!\\n"\);|'
    r'return 0;|'
    r'\}'
    r')\s*$'
)


def strip_boilerplate(lines):
    return [l for l in lines if not BOILERPLATE.match(l.strip())]


grand_old = grand_new = grand_equal = 0

for old_file, new_file in PAIRS:
    old_lines = strip_boilerplate(git_diff_lines(old_file, "-"))
    new_lines = strip_boilerplate(git_diff_lines(new_file, "+"))

    old_norm = [l.strip() for l in old_lines if l.strip()]
    new_norm = [l.strip() for l in new_lines if l.strip()]

    sm = difflib.SequenceMatcher(a=old_norm, b=new_norm, autojunk=False)
    equal = sum(block.size for block in sm.get_matching_blocks())
    total = max(len(old_norm), len(new_norm))
    pct = 100.0 * equal / total if total else 100.0

    grand_old += len(old_norm)
    grand_new += len(new_norm)
    grand_equal += equal

    print(f"{old_file:32s} -> {new_file:48s} "
          f"old={len(old_norm):4d} new={len(new_norm):4d} equal={equal:4d} ({pct:5.1f}%)")

    if "--full" in sys.argv:
        diff = difflib.unified_diff(old_norm, new_norm,
                                     fromfile=old_file, tofile=new_file, lineterm="")
        for line in diff:
            print("   ", line)

total = max(grand_old, grand_new)
pct = 100.0 * grand_equal / total if total else 100.0
print(f"\nTOTAL: old={grand_old} new={grand_new} equal={grand_equal} ({pct:.1f}% content-identical)")

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 69.91%. Comparing base (ceea694) to head (5c87484).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #9276      +/-   ##
==========================================
- Coverage   70.26%   69.91%   -0.35%     
==========================================
  Files         257      257              
  Lines       79106    77251    -1855     
  Branches    18954    18784     -170     
==========================================
- Hits        55583    54011    -1572     
+ Misses      17885    17676     -209     
+ Partials     5638     5564      -74     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.