Fix Mono LLVM lowering of MinNumber/MaxNumber to preserve the sign of zero

[tannergooding]

Fixes #131130.

double.MinNumber/MaxNumber (and the float/Half variants) implement IEEE 754-2019 minimumNumber/maximumNumber, which are NaN-suppressing and treat -0 as less than +0. The Mono LLVM scalar lowering used llvm.minnum/maxnum (IEEE 754-2008 minNum/maxNum), which leave the sign of zero unspecified -- e.g. minnum(+0, -0) returns +0 on x86, so double.MinNumber(+0.0, -0.0) returned +0.0 instead of -0.0 on the LLVM AOT configuration.

Lower to llvm.minimumnum/maximumnum instead (available since LLVM 20; Mono is on LLVM 23). These are the exact 2019 minimumNumber/maximumNumber operations -- NaN-suppressing and sign-of-zero aware -- and on AArch64 still lower to a single fminnm/fmaxnm. Min/Max already use llvm.minimum/maximum and are correct, so they are unchanged.

I confirmed the divergence and the fix by JIT-executing each intrinsic on x86-64: minnum(+0, -0) returns +0 (wrong), while minimumnum(+0, -0) returns -0 (correct) and still suppresses NaN.

---

Un-skips the Runtime_130831 regression test (added in #130832) on Mono, which covers the Min/MinNumber signed-zero cases directly and runs on the AllSubsets_Mono_LLVMAot_RuntimeTests leg.

Note

This change was authored with the assistance of GitHub Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 4 pipeline(s).
12 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @steveisok, @vitek-karas
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR updates Mono’s LLVM backend to correctly implement IEEE 754-2019 minimumNumber / maximumNumber semantics for MinNumber / MaxNumber, specifically preserving the sign of zero (e.g., min(+0, -0) => -0) on LLVM AOT. It also re-enables a previously skipped regression test on Mono now that the behavior should match the BCL contract.

Changes:

• Switch Mono LLVM scalar lowering of MinNumber/MaxNumber from llvm.minnum/llvm.maxnum to llvm.minimumnum/llvm.maximumnum.
• Update the intrinsic ID list to use the new LLVM intrinsic names.
• Unskip Runtime_130831 on Mono to validate signed-zero behavior.

Show a summary per file

File	Description
src/mono/mono/mini/mini-llvm.c	Routes OP_*MINNUM/OP_*MAXNUM to the new MINIMUMNUM/MAXIMUMNUM intrinsics during LLVM emission.
src/mono/mono/mini/llvm-intrinsics.h	Replaces MINNUM/MAXNUM intrinsic definitions with MINIMUMNUM/MAXIMUMNUM to match the intended semantics.
src/mono/mono/mini/intrinsics.c	Updates commentary/documentation for MinNumber/MaxNumber lowering to reflect the new LLVM intrinsics and semantics.
src/tests/JIT/Regression/JitBlue/Runtime_130831/Runtime_130831.cs	Removes [SkipOnMono] so the signed-zero regression test runs on Mono again.

Copilot's findings

Comments suppressed due to low confidence (1)

src/tests/JIT/Regression/JitBlue/Runtime_130831/Runtime_130831.cs:22

• The test now runs on Mono and validates the signed-zero behavior for Min/MinNumber, but this PR also changes the LLVM lowering for MaxNumber. It would be good to add Max/MaxNumber signed-zero assertions here as well so the new MAXIMUMNUM lowering is covered and a future regression is caught.

    [Fact]
    public static void TestEntryPoint()
    {
        Assert.Equal(BitConverter.DoubleToInt64Bits(-0.0), BitConverter.DoubleToInt64Bits(MinNegZeroConst(+0.0)));
        Assert.Equal(BitConverter.DoubleToInt64Bits(-0.0), BitConverter.DoubleToInt64Bits(MinNumberZeroConst(-0.0)));

        Assert.Equal(BitConverter.SingleToInt32Bits(-0.0f), BitConverter.SingleToInt32Bits(MinNegZeroConst(+0.0f)));
        Assert.Equal(BitConverter.SingleToInt32Bits(-0.0f), BitConverter.SingleToInt32Bits(MinNumberZeroConst(-0.0f)));
    }

• Files reviewed: 4/4 changed files
• Comments generated: 0

[tannergooding]

The AllSubsets_Mono_LLVMAot_RuntimeTests legs failed on Runtime_98068.TestMinNumber() (double.MinNumber(NaN, +1.0) returned garbage). The llvm.minimumnum/maximumnum intrinsics from the first commit have the right IEEE 754-2019 semantics, but are miscompiled by the x86 backend in the LLVM version Mono builds against. Confirmed this was the regression (baseline main passes the same leg).

Fixed by composing MinNumber/MaxNumber from llvm.minimum/maximum (sign-of-zero aware, already used for Math.Min/Max) plus an explicit NaN fixup, rather than lowering to the broken intrinsic. This keeps the sign-of-zero fix while avoiding the miscompile. Verified the composed IR bitwise across the NaN / signed-zero edge cases; end-to-end validation is on the re-run of the Mono LLVM AOT leg here.

Note

This comment was drafted by Copilot.

[Copilot]

Copilot's findings

Comments suppressed due to low confidence (2)

src/tests/JIT/Regression/JitBlue/Runtime_130831/Runtime_130831.cs:21

• This test now runs on Mono LLVM AOT, but it only validates the Min/MinNumber signed-zero behavior. The PR also changes the MaxNumber lowering (OP_*MAXNUM), so adding symmetric assertions for Max/MaxNumber would help prevent regressions in that path as well.

        Assert.Equal(BitConverter.DoubleToInt64Bits(-0.0), BitConverter.DoubleToInt64Bits(MinNegZeroConst(+0.0)));
        Assert.Equal(BitConverter.DoubleToInt64Bits(-0.0), BitConverter.DoubleToInt64Bits(MinNumberZeroConst(-0.0)));

        Assert.Equal(BitConverter.SingleToInt32Bits(-0.0f), BitConverter.SingleToInt32Bits(MinNegZeroConst(+0.0f)));
        Assert.Equal(BitConverter.SingleToInt32Bits(-0.0f), BitConverter.SingleToInt32Bits(MinNumberZeroConst(-0.0f)));

src/mono/mono/mini/mini-llvm.c:7639

• The PR description says OP_*MINNUM/OP_*MAXNUM are lowered to llvm.minimumnum/maximumnum, but this implementation instead composes llvm.minimum/maximum with explicit NaN fixups and even states minimumnum/maximumnum are miscompiled on x86. Please reconcile (either switch to the new intrinsics as described, or update the PR description / add an LLVM issue reference + clarify the expected codegen impact, especially on AArch64).

			 * We compose this from llvm.minimum/maximum (sign-of-zero aware but
			 * NaN-propagating) plus an explicit NaN fixup, rather than lowering
			 * directly to an intrinsic: llvm.minnum/maxnum leave the sign of zero
			 * unspecified (minnum(+0, -0) may return +0 on x86, see dotnet/runtime
			 * #131130) and llvm.minimumnum/maximumnum are miscompiled by the x86

• Files reviewed: 4/4 changed files
• Comments generated: 1

[akoeplinger]

The llvm.minimumnum/maximumnum intrinsics from the first commit have the right IEEE 754-2019 semantics, but are miscompiled by the x86 backend in the LLVM version Mono builds against.

should we file an llvm bug for this?

[tannergooding]

should we file an llvm bug for this?

Was already working on that. I've logged a few issues against Clang/LLVM in the past few days 😅

[tannergooding]

@akoeplinger, looks like this is already fixed upstream

Copilot reports back, and then I verified after:

LLVM minimumnum miscompile — pin & fix investigation

What Mono is pinned to

• Package: 23.1.0-alpha.1.26357.1 from dotnet/llvm-project (eng/Version.Details.xml).
• Snapshot commit: 04bf14b2 (2026-07-07) — an arcade dep-update only; it did not advance LLVM sources.
• Effective LLVM sources: upstream llvm/llvm-project main @ a72232481fe7 (2026-06-13) — the fork's last merge-upstream-main (c9811b58) — plus a handful of unrelated Mono-local patches.

The bug is upstream, not a Mono patch — proven

The dnceng LLVM tree = upstream base + Mono patches. I compared git blob SHAs of all 9 files implementing float min/max lowering across the two forks; every one is byte-identical to upstream a72232481fe7:

X86ISelLowering.cpp, SelectionDAG/{LegalizeDAG, LegalizeFloatTypes, LegalizeVectorTypes, LegalizeVectorOps, SelectionDAGBuilder, DAGCombiner}.cpp, TargetLoweringBase.cpp, Analysis/ConstantFolding.cpp — all SAME.

The Mono-local patches (mono.this metadata, MonoException.cpp, libatomic, LLVMConstInt truncation) don't touch min/max lowering.

Regression window

Revision	minimumnum(NaN, 1.0)
LLVM 22.1.0 release	✅ correct
upstream main @ 2026-06-13 (a72232481fe7, the pin)	❌ -1.0 (per Mono CI)
current trunk	✅ correct

Trunk-only regression: introduced after the 22.x branch, already fixed on trunk. The dnceng snapshot landed inside the window. Bug is asymmetric — minimumnum broken, maximumnum fine.

Supposed fix (not yet in the fork)

dotnet/main last merged upstream at c9811b58 (2026-06-13), so the fork lacks these later X86 changes:

• 961ab7a3 (2026-07-08) — [SDAG][X86] Uplift pseudo fmin/fmax from X86 (#188489) — leading suspect; removes X86's own 177-line pseudo fmin/fmax lowering (the exact path the pin still uses) and re-homes it in SDAG.
• 8a3b7152 (2026-07-01) — [X86] Fold FP UNORD/ORD compare against known-non-NaN to self-compare (#206943) — touches the cmpordsd NaN logic; lower confidence.

Caveat: couldn't build llc at a72232481fe7 to execute + git-bisect the exact fix, so #188489 is the leading suspect by code path, not proven. The "not a Mono patch" conclusion is proven via blob comparison.

Fix path

Fresh merge-upstream-main into dotnet/llvm-project past 2026-07-08 (pulls in #188489), then bump the snapshot in eng/Version.Details.xml.

Item	Value
Mono LLVM pin	23.1.0-alpha.1.26357.1 → dotnet/llvm-project 04bf14b2
Effective upstream base	a72232481fe7 (main @ 2026-06-13)
Suspected upstream fix	#188489 961ab7a3 (2026-07-08)

[akoeplinger]

ok. I see commit 961ab7a3 made it into the 23.x release branch which we're planning to sync to in a bit.

[tannergooding]

made it into the 23.x release branch which we're planning to sync to in a bit.

@akoeplinger, if you can ping me when this happens I can either sign-off on or remove the extra workaround here