wasm: support ThinLTO

Using ThinLTO manages to optimize binaries quite significantly. The
exact amount varies a lot by program but it's about 10-15% usually.

Don't remove non-ThinLTO support yet. It would not surprise me if this
triggered some unintended side effect. Eventually, non-ThinLTO support
should be removed though.

compiler/compiler.go

@@ -1099,6 +1099,11 @@ func (b *builder) createFunctionStart(intrinsic bool) {
 		// otherwise the function is not exported.
 		functionAttr := b.ctx.CreateStringAttribute("wasm-export-name", b.info.linkName)
 		b.llvmFn.AddFunctionAttr(functionAttr)
+		// Unlike most targets, exported functions are actually visible in
+		// WebAssembly (even if it's not called from within the WebAssembly
+		// module). But LTO generally optimizes such functions away. Therefore,
+		// exported functions must be explicitly marked as used.
+		llvmutil.AppendToGlobal(b.mod, "llvm.used", b.llvmFn)
 	}
 
 	// Some functions have a pragma controlling the inlining level.

compiler/testdata/pragma.ll

@@ -6,6 +6,7 @@ target triple = "wasm32-unknown-wasi"
 @extern_global = external global [0 x i8], align 1
 @main.alignedGlobal = hidden global [4 x i32] zeroinitializer, align 32
 @main.alignedGlobal16 = hidden global [4 x i32] zeroinitializer, align 16
+@llvm.used = appending global [2 x ptr] [ptr @extern_func, ptr @exportedFunctionInSection]
 @main.globalInSection = hidden global i32 0, section ".special_global_section", align 4
 @undefinedGlobalNotInSection = external global i32, align 4
 @main.multipleGlobalPragmas = hidden global i32 0, section ".global_section", align 1024