[jogibear9988]

Calling a mutating struct method on a readonly instance field currently changes the original field. Load its value into the existing temporary-address path before calling the member, matching System.Linq.Expressions. Writable fields and explicit assignment/ref paths retain their address handling.

The regression fixture checks repeated calls, writable controls, static readonly fields, nested readonly receivers, and receiver evaluation exactly once, using strict CompileFast and System comparison.

Allocation tradeoff: compiling the measured readonly struct receiver costs 8 additional bytes per compilation (944 to 952 B/op) in both Standard and LightExpression, on .NET 8 and 10. The defensive copy requires a local. This is an accepted correctness cost; it is not allocation-neutral. Runtime allocation remains zero. Other measured working paths retain their allocation counts and emitted IL.

Validation

Baseline: fetched dadhi/master at e2c78de038244228f978f93ed89f40f30daec680. The regression fixture fails on master and passes with this change, for both input APIs.

Release solution build: zero warnings/errors. Full runner results:

Runtime	TestX passed	Legacy passed
.NET 10	308	1756
.NET 8	308	1755
.NET Framework 4.7.2	289	1639

Performance measurements

Windows ARM64, .NET 8.0.28 and 10.0.9, Release, tiering disabled. Prebuilt trees; Light conversion occurs outside timing. Separate assembly contexts, one CPU, nine alternating samples per phase, two fresh process rounds with reversed load order, roughly 100 ms pair calibration, and GC allocation counters. Tables show the median of the 18 samples for compilation and delegate invocation; ns/op and allocated B/op are master → patch. All measured expressions already produce correct results on master.

Timing limitation: this machine shows substantial variation even comparing master with itself (individual process medians ranged from -43.5% to +16.8% for compilation and -46.9% to +55.5% for invocation). These results do not establish a speedup or prove CPU-time neutrality. Slower observations are included below. Allocation and IL comparisons are more stable.

All measured working cases, including controls

Runtime / API / workload	Compile ns	Compile B	Invoke ns	Invoke B	Same IL
net10.0 / Standard / add_control	5078.7 → 4953.5	872 → 872	2.9 → 2.9	0 → 0	yes
net10.0 / Standard / nested_field_read	3488.9 → 3570.9	952 → 952	2.1 → 1.9	0 → 0	yes
net10.0 / Standard / readonly_read	5324.5 → 5744.2	944 → 952	2.0 → 1.9	0 → 0	no
net10.0 / Standard / writable_read	5125.8 → 4139.7	944 → 944	1.9 → 2.0	0 → 0	yes
net10.0 / Light / add_control	2505.5 → 2508.9	872 → 872	2.1 → 2.0	0 → 0	yes
net10.0 / Light / nested_field_read	3295.9 → 3753.0	952 → 952	2.1 → 1.9	0 → 0	yes
net10.0 / Light / readonly_read	2626.4 → 2768.1	944 → 952	2.0 → 2.2	0 → 0	no
net10.0 / Light / writable_read	3507.3 → 3554.5	944 → 944	2.6 → 2.1	0 → 0	yes
net8.0 / Standard / add_control	1145.8 → 1120.4	872 → 872	1.9 → 1.9	0 → 0	yes
net8.0 / Standard / nested_field_read	1096.7 → 1108.3	952 → 952	1.8 → 1.9	0 → 0	yes
net8.0 / Standard / readonly_read	1234.0 → 1257.5	944 → 952	1.8 → 1.8	0 → 0	no
net8.0 / Standard / writable_read	1278.9 → 1227.8	944 → 944	1.9 → 1.9	0 → 0	yes
net8.0 / Light / add_control	966.1 → 1056.9	872 → 872	2.0 → 1.9	0 → 0	yes
net8.0 / Light / nested_field_read	1227.9 → 1184.6	952 → 952	2.1 → 2.1	0 → 0	yes
net8.0 / Light / readonly_read	1192.2 → 1236.8	944 → 952	2.0 → 2.0	0 → 0	no
net8.0 / Light / writable_read	1388.5 → 1411.1	944 → 944	2.3 → 2.2	0 → 0	yes

[dadhi]

@jogibear9988 thank you