From 37b9da32d3b911091deb254f6cba2a137c471287 Mon Sep 17 00:00:00 2001 From: Marc Vertes Date: Thu, 12 Oct 2023 10:51:58 +0200 Subject: move to a direct byte code compiler (#8) * chore: refactor to keep only the new parser and bytecode vm * scanner: remove Token.value field * scanner: remove scanner.kind field * chore: move language specification in lang package This avoid a cyclic dependency in scanner_test which can now use the golang/GoSpec language specification for Go. * clean code * scanner: export scanner fields Also parser now generate function calls, including externals. * chore: fix lint issues * parser: handle strings * wip * parser: implement support for 'if, else, else if' statements Resolving labels in the compiler still in progress. * parser: support if statements, improve compiler * improve handling of functions * improve support of local variables * scanner: trim leading and trailing spaces * fixes to make fibonacci work * parser: improve README, fix function parameters parsing --- vm/README.md | 38 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 38 insertions(+) create mode 100644 vm/README.md (limited to 'vm/README.md') diff --git a/vm/README.md b/vm/README.md new file mode 100644 index 0000000..92b1ac8 --- /dev/null +++ b/vm/README.md @@ -0,0 +1,38 @@ +# vm + +`vm` is a bytecode based stack machine. + +The purpose of `vm` is to provide a simple, fast, embeddable and +portable Go execution environment. + +```mermaid +graph LR +s[ ] --> |source| a(scanner) +--> |tokens| b(parser) +--> |AST| c(codegen) +--> |bytecode| d[vm] +subgraph vm + d +end +style s height:0px; +``` + +The bytecode consists of a dozen of instructions, each taking 0 or 1 +immediate argument (non-immediate arguments are only passed through the +stack). Only a few operators for a few types are implemented. I expect +to have 1 instruction per operator per numerical type, all with the same +pattern, which would be generated from a template. Estimate is around 20 +operators and 10 numerical types, so around 200 instructions in final. + +Structurally, the vm implements logical and arithmetic operators, +condional jumps for `if`, `for` and `switch` control flow, and function +call, return and frame management. + +the memory state of the vm is a slice of Go interfaces (`[]any`). + +The whole vm is coded in a single function of 80 lines with no +dependencies. The size will grow as we add missing instructions, but the +code complexity will remain the same. + +the vm1 package is totally standalone and could be used for any purpose +outside of parscan and/or gno. -- cgit v1.2.3