Limit inference from apparent types to one level deep (microsoft#27225)

* Limit inference from apparent types to one level deep

* Rename marker & use booleanness not ===

* Undo typo

Author: weswigham

src/compiler/checker.ts

@@ -13439,6 +13439,7 @@ namespace ts {
             let visited: Map<boolean>;
             let contravariant = false;
             let propagationType: Type;
+            let allowComplexConstraintInference = true;
             inferFromTypes(originalSource, originalTarget);
 
             function inferFromTypes(source: Type, target: Type): void {
@@ -13616,7 +13617,15 @@ namespace ts {
                         // getApparentType can return _any_ type, since an indexed access or conditional may simplify to any other type.
                         // If that occurs and it doesn't simplify to an object or intersection, we'll need to restart `inferFromTypes`
                         // with the simplified source.
-                        if (apparentSource !== source && !(apparentSource.flags & (TypeFlags.Object | TypeFlags.Intersection))) {
+                        if (apparentSource !== source && allowComplexConstraintInference && !(apparentSource.flags & (TypeFlags.Object | TypeFlags.Intersection))) {
+                            // TODO: The `allowComplexConstraintInference` flag is a hack! This forbids inference from complex constraints within constraints!
+                            // This isn't required algorithmically, but rather is used to lower the memory burden caused by performing inference
+                            // that is _too good_ in projects with complicated constraints (eg, fp-ts). In such cases, if we did not limit ourselves
+                            // here, we might produce more valid inferences for types, causing us to do more checks and perform more instantiations
+                            // (in addition to the extra stack depth here) which, in turn, can push the already close process over its limit.
+                            // TL;DR: If we ever become generally more memory efficienct (or our resource budget ever increases), we should just
+                            // remove this `allowComplexConstraintInference` flag.
+                            allowComplexConstraintInference = false;
                             return inferFromTypes(apparentSource, target);
                         }
                         source = apparentSource;