My apologies: this is an opinion, rather than an explicit declaration from the project. Their compiler "gc" starts with a traditional recursive descent parser using bison, and chose not to use llvm. Although one can argue that llvm doesn't provide enough information to implement a great jit, all aot compilers are moving towards using it. By having a common compiler infrastructure, compilers can focus on what is important. They currently have a simple parallel mark-and-sweep garbage collector, and there is some scope for development in the context of concurrent programs (wasn't it their goal to make it easy to write concurrent programs with goroutines?). Nothing spectacular though: the garbage collector in a mature JVM or V8 (which has to deal with a weakly typed language!) is far ahead; that's where the garbage collection research happens.
> what the final programs looks like is secondary concern
The grammar has been specifically engineered so that the language can be analyzed and parsed without a symbol table. "go fmt" is just syntactical sugar: you cannot change the underlying parser. I'm not bikeshedding about which syntax is "better", but pointing out the motivation behind the current syntax. For instance, it's very different from ruby, which was designed ground-up to resemble human language as closely as possible (hence Poignant Ruby etc).
[edit: gc does not use a gcc backend; gccgo does]