allow external lattice to provide its own field-merge implementation

When merging `PartialStruct`, we are currently merging their fields a
bit aggressively (#47307) in order to accelerate convergence.
However, when `PartialStruct` wraps external lattice elements, this can
be too aggressive since it does not use `tmerge(𝕃, fields...)` recursively
and thus the external lattice elements are not merged as expected.

This commit adds an additional lattice hook, `tmerge_field`, inside
`tmerge(::PartialsLattice)` so that external lattice implementation
can provide its own field-merge strategies.

Author: aviatesk

base/compiler/abstractlattice.jl

@@ -131,6 +131,23 @@ remaining mostly associative and commutative.
 """
 function tmerge end
 
+"""
+    tmerge_field(𝕃::AbstractLattice, a, b) -> nothing or lattice element
+
+Compute a lattice join of elements `a` and `b` over the lattice `𝕃`,
+where `a` and `b` are fields of `PartialStruct` or `Const`.
+This is an opt-in interface to allow external lattice implementation to provide its own
+field-merge strategy. If it returns `nothing`, `tmerge(::PartialsLattice, ...)`
+will use the default aggressive type merge implementation that does not use `tmerge`
+recursively to reach convergence.
+"""
+function tmerge_field end
+
+function tmerge_field(𝕃::AbstractLattice, @nospecialize(a), @nospecialize(b))
+    return tmerge_field(widenlattice(𝕃), a, b)
+end
+tmerge_field(::JLTypeLattice, @nospecialize(a), @nospecialize(b)) = nothing
+
 """
     ⊑(𝕃::AbstractLattice, a, b)
 

base/compiler/typelimits.jl

@@ -515,8 +515,12 @@ function tmerge(lattice::PartialsLattice, @nospecialize(typea), @nospecialize(ty
                     tyi = ai
                 elseif is_lattice_equal(lattice, bi, ft)
                     tyi = bi
+                elseif (tyi′ = tmerge_field(lattice, ai, bi); tyi′ !== nothing)
+                    # allow external lattice implementation to provide a custom field-merge strategy
+                    tyi = tyi′
                 else
-                    # Otherwise choose between using the fieldtype or some other simple merged type.
+                    # Otherwise use the default aggressive field-merge implementation, and
+                    # choose between using the fieldtype or some other simple merged type.
                     # The wrapper type never has restrictions on complexity,
                     # so try to use that to refine the estimated type too.
                     tni = _typename(widenconst(ai))