bugc: fix optimizer for nested call arguments in recursion#208
Merged
gnidan merged 1 commit intocall-returnfrom Apr 2, 2026
Merged
bugc: fix optimizer for nested call arguments in recursion#208gnidan merged 1 commit intocall-returnfrom
gnidan merged 1 commit intocall-returnfrom
Conversation
Move phi resolution from target blocks to predecessor blocks (standard phi deconstruction). The previous approach resolved phis at the target using the layout-order predecessor, which broke for back-edges where the runtime predecessor differs. Also fixes jump optimization eliminating phi-bearing blocks and rewrites TCO to use a pre_entry trampoline with phis on the original entry block.
Contributor
|
gnidan
added a commit
that referenced
this pull request
Apr 16, 2026
* Draft function call context schemas * Organize schema a bit * Disable unevaluated properties * Allow additionalProperties at top-level * Add function identity fields to invoke/return/revert contexts Introduce a shared context/function schema defining identifier, declaration, and type properties (mirroring Variable's pattern) and compose it into invoke, return, and revert via allOf. Switch return and revert from additionalProperties to unevaluatedProperties to support allOf composition. * Fix Playground crash on page refresh in dev server Wrap the Monaco-based Playground component in BrowserOnly with React.lazy so it only loads client-side, preventing SSR-related initialization errors that caused [object Object] crashes on refresh. * Nest invoke/return/revert under function/ category Move function call lifecycle contexts (invoke, return, revert) into a function/ subcategory and extract shared identity fields (identifier, declaration, type) into a base function schema that each context extends via $ref. * Rename invoke discriminant fields: internal->jump, call->message Reduces ambiguity: `call` could mean any function call, and `internal` vs `create` doesn't clearly convey the mechanism. The new names (jump/message/create) map directly to the three EVM invocation mechanisms. Also enforces mutual exclusivity of delegate/static on external calls via `not: required: [delegate, static]`. * Rewrite function context descriptions Frame descriptions from the perspective of contexts marking instructions rather than "representing" things. A context indicates association with a function lifecycle event; it does not represent the event itself. * Rewrite function context examples with realistic scenarios Each example now describes a concrete EVM execution scenario: which instruction the context marks, what the stack and memory layout looks like, and why each pointer points where it does. * Add type specifier schema and use it across the format Introduce schema:ethdebug/format/type/specifier to formalize the "type or type reference" pattern. Update type wrapper, variables, and doc pages to use the new schema instead of inline oneOf/if-then discrimination. Remove the now-unnecessary allow-list entry for the function context schema. * Format * format: add TypeScript types for function call contexts (#186) Add Type.Reference, Type.Specifier, and type guards for the new function call context schemas (invoke, return, revert). - Type.Reference: { id: string | number } for type references - Type.Specifier: Type | Reference union (matches type/specifier schema) - Context.Function.Identity: shared function identity fields - Context.Invoke: internal calls, external calls, contract creation - Context.Return: function return with data and optional success - Context.Revert: function revert with optional reason/panic - Update Variable.type to accept Type.Specifier (was Type only) - Update Type.Wrapper to use Type.Specifier * bugc: emit invoke/return contexts for internal function calls (#185) Add debug context annotations for function call boundaries using the typed Format.Program.Context.Invoke and Context.Return interfaces: - Invoke context on caller's JUMP instruction with target pointer and argument group pointers (stack slots) - Return context on continuation JUMPDEST with data pointer to return value at stack slot 0 - Invoke context on callee entry JUMPDEST with target and argument pointers Includes tests verifying context emission for single-arg, multi-arg, nested, and void function call scenarios. * bugc: add source maps for call setup instructions (#188) The call setup sequence (POP cleanup, PUSH return address, MSTORE, push arguments, PUSH function address) was using remark-only debug contexts, leaving these instructions unmapped in tracing output. Now threads the call terminator's operationDebug (which carries the source code range for the call expression) through all setup instructions, matching how other instruction generators use operationDebug. * Add function call tracing documentation (#187) * Add function call tracing documentation Add invoke/return/revert context documentation across concept, reference, and spec pages: - concepts/programs.mdx: new "Function call contexts" section explaining the three context types with a SchemaExample - tracing.mdx: walkthrough of tracing through an internal function call (Adder contract), plus external call and revert examples - Spec pages: added intro prose to function.mdx, return.mdx, revert.mdx * Make function call tracing example interactive Replace the static BUG code block with an interactive TraceExample component that lets readers compile and step through the Adder contract, seeing invoke/return contexts at function boundaries. Static SchemaExample blocks are kept for the narrative walkthrough and for external call/revert examples (which BUG can't demonstrate). * Fix BUG source indentation in tracing examples Add 2-space indentation inside block bodies (storage, create, code, if) to match the canonical style used in .bug example files. * Revert BUG indentation to match prettier formatting Prettier strips indentation inside template literal strings in MDX JSX props. Revert to the unindented style that prettier enforces. * docs: move 4th BUG example into tracing-examples.ts Move the "Function call and return" (Adder) example into tracing-examples.ts with proper indentation, matching the pattern established for the other three examples. * docs: add invoke.mdx prose intro and cross-link spec pages (#192) Add a descriptive introduction to the invocation contexts spec page, matching the pattern of function.mdx, return.mdx, and revert.mdx. Add cross-links between the function context spec pages and the tracing documentation. * bugc: use format type guards in call context tests (#194) Validate that compiler output satisfies @ethdebug/format type guards (Context.isInvoke, Context.isReturn, Invocation.isInternalCall) instead of casting through Record<string, unknown>. Adds a typed findInstructionsWithContext helper that filters by mnemonic and type guard, providing proper type narrowing. * bugc: add declaration source ranges and param names to invoke/return contexts (#196) The invoke and return contexts emitted by bugc now include declaration source ranges (pointing to the function definition) and named argument pointers (using parameter names from the function signature). This enriches debug info so debuggers can link call stack entries to source declarations and display meaningful parameter names. Changes: - Add loc/sourceId fields to Ir.Function, populated during IR generation from AST function declarations - Thread module.functions map through EVM codegen so call terminators can look up target function metadata - Build declaration source ranges on invoke contexts (caller JUMP, callee entry JUMPDEST) and return contexts (continuation JUMPDEST) - Add parameter names to argument group pointers - Update call-contexts tests for new fields * bugc: add debug contexts to all unmapped bytecodes (#190) * bugc: add debug contexts to all remaining unmapped bytecodes Thread remark/code contexts through all compiler-generated instructions that previously lacked debug info: - Free memory pointer initialization (remark) - Return value spill after call continuation (call expr source range) - STOP guard between main and user functions (remark) - Function prologue MSTORE for param storage (thread existing remark) - Function prologue return PC save sequence (thread existing remark) - Deployment wrapper CODECOPY+RETURN (remark) All 82 instructions across runtime and create programs now carry debug contexts (previously 22 were unmapped). * bugc: add code contexts with source ranges to compiler-generated instructions Add source location info (loc, sourceId) to Ir.Function so EVM codegen can build code contexts for compiler-generated instructions. Instructions that map to a source location now use gather contexts combining both a remark (for debugger tooling) and a code context (for source highlighting): - Free memory pointer init → code block / create block range - Function prologue (param stores, return PC save) → function decl range - STOP guard → code block range Deployment wrapper remains remark-only (no corresponding source). Return value spill already had correct source mapping (call expr). * Add call stack breadcrumb and call info panel (#189) * Add call stack breadcrumb and call info panel components Surface invoke/return/revert context information in the trace viewer: a breadcrumb showing the current call stack, and a panel showing call details with async-resolved pointer ref values. New components: CallStackDisplay, CallInfoPanel New utilities: extractCallInfoFromInstruction, buildCallStack, buildPcToInstructionMap New types: CallInfo, CallFrame, ResolvedCallInfo, ResolvedPointerRef * Integrate CallStackDisplay and CallInfoPanel into TraceViewer The components were exported from programs-react but never rendered in the web package's TraceViewer. Add them to the layout: call stack breadcrumb in the header, call info panel at the top of the right sidebar. * Add call stack breadcrumb and call info banner to TraceDrawer The deploy preview uses TraceDrawer (not TraceViewer) for the interactive trace playground. Add call context display directly to TraceDrawer: a breadcrumb bar showing nested call frames with clickable navigation, and a colored banner showing invoke/return/revert status at the current step. * Show always-visible call stack and fix duplicate frame bug - Call stack bar now always visible with "(top level)" empty state so users know the feature exists - Fix duplicate call stack frames: compiler emits invoke context on both the caller JUMP and callee entry JUMPDEST, so skip push if top frame already matches the same call - Applied fix to both TraceDrawer and programs-react utility * docs: expand return and revert spec pages (#193) Add subsections to return.mdx covering internal returns, external call returns, the success field, and data pointer. Add subsections to revert.mdx covering reason-based reverts, panic codes, and field optionality. Both pages now match the depth of invoke.mdx. * bugc-react: surface invoke/return/revert contexts in BytecodeView (#195) Add context-aware rendering for function call debug contexts in the bytecode disassembly view. Instructions with invoke, return, or revert contexts now show colored badges and inline labels instead of the generic info icon, with colored left borders to make call boundaries visually scannable. - Add classifyContext/summarizeContext utils for extracting human-readable info from debug context objects - Replace generic info icon with arrow/return/x badges for invoke/return/revert instructions - Show inline context labels (e.g. "invoke add", "return add") - Add structured tooltip headers before raw JSON - CSS for context badges, labels, and row highlighting * bugc: map return epilogue instructions to source location (#197) The function-return PUSH/MLOAD/JUMP sequence previously had remark-only debug context. Now uses the IR return terminator's operationDebug, which carries the source location of the return statement. This lets debuggers map the epilogue back to the BUG source. * ui: named arguments, click-to-source, and call stack params (#198) Show named argument parameters throughout the UI: - BytecodeView labels show "invoke add(a, b)" instead of "invoke add" when argument names are available in the invoke context's pointer group entries - Call stack breadcrumbs show "add(a, b)" instead of "add()" in both programs-react CallStackDisplay and web TraceDrawer - CallInfoPanel banner shows "Calling add(a, b)" Add click-to-source for context badges: - BytecodeView accepts onDeclarationClick callback - Clicking an invoke/return/revert badge with a declaration field fires the callback with sourceId, offset, and length - Falls back to pinning the tooltip if no declaration Extract declaration and argument names from contexts: - summarizeContext now returns argumentNames and declaration - New formatCallSignature utility for consistent formatting - New DeclarationRange type for click-to-source data - CallFrame and CallInfo types carry argumentNames * Fix typos in type concepts and add missing CSS classes (#199) Fix "disinction" typo and "languages that whose" grammar error in type concepts spec page. Add missing CSS class definitions for .call-stack-empty-text and .call-info-ref-value in both programs-react and web theme. * docs: fix broken anchor in type concepts page Update internal link to match renamed heading "Type specifiers, wrappers, and references". * bugc: fix multi-block function return values (#200) Functions with if/else branches returning from both sides produced wrong return values. The branch terminator's loadValue(condition) DUPs the condition onto the stack; JUMPI consumes the dup but the original remains. Successor blocks inherit this stale value, so the return epilogue incorrectly treated it as the return value. Add generateReturnEpilogue that loads the return value, cleans stale stack entries via SWAP+POP, then loads the saved return PC and jumps back. Uses the same imperative state pattern as generateCallTerminator. Also changes generateTerminator return type from Transition<S, S> to Transition<S, Stack> to accommodate the stack-cleaning epilogue. * bugc: add runtime call stack for recursion support (#201) Implement FMP-based call frames so functions can call themselves recursively. Each call allocates a frame (saved FP, saved return PC, locals) from the free memory pointer; returns deallocate by restoring FP and FMP. Also fixes pre-existing SUB/DIV/MOD operand ordering bug where non-commutative operations computed right op left instead of left op right. * docs: add recursive function call tracing example (#202) Add a recursive count/succ example to the tracing docs, showing how nested invoke/return contexts appear when functions call themselves repeatedly. * Fix call stack display for recursive function calls (#203) The call stack dedup logic compared only function name and call type, causing recursive calls (e.g. count -> count) to be collapsed into a single frame. Now also checks whether the previous frame was pushed on the immediately preceding step, which distinguishes the compiler's duplicate invoke contexts (caller JUMP + callee JUMPDEST on consecutive steps) from genuine recursive calls (same name but steps far apart). Fixed in both programs-react buildCallStack and web TraceDrawer's duplicated call stack logic. * Add source-level stepping to trace viewer controls (#204) Use outline triangles (◁/▷) for single-step and filled triangles (◀/▶) for source-level stepping, alongside skip-to-start (⏮) and skip-to-end (⏭) controls. * docs: replace recursive example with isEven/isOdd mutual recursion (#205) Replace the succ/count recursive example with an isEven/isOdd mutual recursion example, which better demonstrates alternating invoke/return contexts between two functions. * bugc: add skipped test for optimizer recursion bug Nested call arguments (e.g. count(succ(n), target)) fail at optimizer level 2+. Add a skipped test to track the issue. * Resolve argument values in call stack display (#206) Store argument pointers on call stack frames and resolve them against historical machine state at each frame's invoke step. Display as "add(a: 3, b: 4)" in both CallStackDisplay and TraceDrawer breadcrumbs. When the compiler emits duplicate invoke contexts (caller JUMP + callee JUMPDEST), use the callee entry step for resolution since argument pointers reference stack slots valid at the JUMPDEST, not the JUMP. Fix stack peek indexing in both traceStepToMachineState and traceStepToState: depth 0 should read from the top of the stack (last array element), not the bottom. The evm package stores stacks bottom-first. Values are cached by step index to avoid re-resolving unchanged frames. Small values (<=9999) display as decimal, larger values as hex. * bugc: fix optimizer for nested call arguments in recursion (#208) Move phi resolution from target blocks to predecessor blocks (standard phi deconstruction). The previous approach resolved phis at the target using the layout-order predecessor, which broke for back-edges where the runtime predecessor differs. Also fixes jump optimization eliminating phi-bearing blocks and rewrites TCO to use a pre_entry trampoline with phis on the original entry block. * format: fix invoke schema pointer semantics and examples (#207) * format: fix invoke schema pointer semantics and examples Align invoke context examples with the "following execution" semantics defined by instruction.schema.yaml. The internal call example previously referenced pre-JUMP stack state (destination at slot 0, arguments at slots 2-3), but JUMP consumes its destination operand. Fix the example to mark the callee's entry JUMPDEST with the correct post-JUMP stack layout: return label at slot 0, arguments at slots 1-2, and target as a code pointer to the function's entry offset. Also fix the CREATE2 example: salt was incorrectly shown at stack slot 1 instead of slot 3 (the EVM pops value, offset, length, salt in that order). Add clarifying text to the schema description and spec page explaining when internal call contexts use callee JUMPDEST placement vs. external call/create contexts on the instruction itself, and how stack-based pointers reference pre-execution state for instructions that consume their operands. * format: clarify context semantics — facts vs pointer state Resolve the tension between "following execution" semantics and pointer evaluation. The instruction schema's "following execution" describes when a context's semantic facts hold (e.g., "a function was invoked"), not the machine state that pointers reference. Pointers reference the state at the instruction's trace step — the state a debugger observes when it encounters the instruction. Update instruction.schema.yaml to make this two-level model explicit. Simplify the invoke schema description and spec page to present this cleanly rather than restating the contradiction. * chore: fix formatting * bugc: update invoke contexts per spec fix (#207) (#209) * bugc: update invoke contexts per spec fix (#207) - Callee JUMPDEST: full invoke with arg pointers and code target (resolved at module-level patching) - Caller JUMP: identity + code target only, no arg pointers - Target pointers changed from stack to code location * fix: use Pointer.Region.isCode type guard in tests * programs-react: propagate arg names from callee JUMPDEST The invoke spec change moved argument pointers from the caller JUMP to the callee entry JUMPDEST. The call stack builder was updating argumentPointers from the duplicate callee step but not argumentNames, causing names to show as _0 instead of the actual parameter name. * web: propagate arg names from callee JUMPDEST in TraceDrawer Same fix as the programs-react mockTrace.ts change — the duplicated call stack builder in TraceDrawer.tsx also needs to update argumentNames from the callee entry step, not just argumentPointers.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes the level 2+ optimizer failure with nested call arguments like
count(succ(n), target).