Put back getfield :boundscheck hack (#48677)

We used to have this hack before #48246, but I removed it because I had
hoped we don't need. Unfortunately, we weren't able to infer consistency of
```
@inbounds (1,2)[1]
```
With this hack, constprop is able to independently prove inbounded-ness,
overriding the usual consistency taintaing that happens for inbounds.

Author: Keno

base/compiler/abstractinterpretation.jl

@@ -1974,6 +1974,14 @@ function abstract_call_known(interp::AbstractInterpreter, @nospecialize(f),
         end
         rt = abstract_call_builtin(interp, f, arginfo, sv, max_methods)
         effects = builtin_effects(𝕃ᵢ, f, arginfo, rt)
+        if f === getfield && (fargs !== nothing && isexpr(fargs[end], :boundscheck)) && !is_nothrow(effects) && isa(sv, InferenceState)
+            # As a special case, we delayed tainting `noinbounds` for getfield calls in case we can prove
+            # in-boundedness indepedently. Here we need to put that back in other cases.
+            # N.B.: This isn't about the effects of the call itself, but a delayed contribution of the :boundscheck
+            # statement, so we need to merge this directly into sv, rather than modifying thte effects.
+            merge_effects!(interp, sv, Effects(EFFECTS_TOTAL; noinbounds=false,
+                consistent = (get_curr_ssaflag(sv) & IR_FLAG_INBOUNDS) != 0 ? ALWAYS_FALSE : ALWAYS_TRUE))
+        end
         return CallMeta(rt, effects, NoCallInfo())
     elseif isa(f, Core.OpaqueClosure)
         # calling an OpaqueClosure about which we have no information returns no information
@@ -2195,6 +2203,15 @@ function abstract_eval_value_expr(interp::AbstractInterpreter, e::Expr, vtypes::
         return rt
     elseif head === :boundscheck
         if isa(sv, InferenceState)
+            stmt = sv.src.code[sv.currpc]
+            if isexpr(stmt, :call)
+                f = abstract_eval_value(interp, stmt.args[1], vtypes, sv)
+                if f isa Const && f.val === getfield
+                    # boundscheck of `getfield` call is analyzed by tfunc potentially without
+                    # tainting :inbounds or :consistent when it's known to be nothrow
+                    @goto delay_effects_analysis
+                end
+            end
             # If there is no particular `@inbounds` for this function, then we only taint `:noinbounds`,
             # which will subsequently taint `:consistent`-cy if this function is called from another
             # function that uses `@inbounds`. However, if this `:boundscheck` is itself within an
@@ -2203,6 +2220,7 @@ function abstract_eval_value_expr(interp::AbstractInterpreter, e::Expr, vtypes::
             merge_effects!(interp, sv, Effects(EFFECTS_TOTAL; noinbounds=false,
                 consistent = (get_curr_ssaflag(sv) & IR_FLAG_INBOUNDS) != 0 ? ALWAYS_FALSE : ALWAYS_TRUE))
         end
+        @label delay_effects_analysis
         rt = Bool
     elseif head === :inbounds
         @assert false && "Expected this to have been moved into flags"

test/compiler/effects.jl

@@ -739,15 +739,19 @@ end |> Core.Compiler.is_foldable_nothrow
 
 # Test that dead `@inbounds` does not taint consistency
 # https://github.com/JuliaLang/julia/issues/48243
-@test Base.infer_effects() do
-    false && @inbounds (1,2,3)[1]
+@test Base.infer_effects(Tuple{Int64}) do i
+    false && @inbounds (1,2,3)[i]
     return 1
 end |> Core.Compiler.is_foldable_nothrow
 
 @test Base.infer_effects(Tuple{Int64}) do i
     @inbounds (1,2,3)[i]
 end |> !Core.Compiler.is_consistent
 
+@test Base.infer_effects(Tuple{Tuple{Int64}}) do x
+    @inbounds x[1]
+end |> Core.Compiler.is_foldable_nothrow
+
 # Test that :new of non-concrete, but otherwise known type
 # does not taint consistency.
 @eval struct ImmutRef{T}