[IN] Global register allocator

[matgla]

and a lot of other things + benchmarks and tests

[Copilot]

Pull request overview

This PR significantly expands ARM Thumb (ARMv8-M) backend infrastructure and documentation while removing unused ARM64 stubs, aiming to support a global register allocator and related optimizations, benchmarks, and test workflows.

Changes:

• Added ARM Thumb backend components (scratch register management, call-site tracking, ABI classification, arch/FPU configs).
• Added extensive design/validation documentation and benchmark analysis notes for IR/codegen optimizations.
• Added container + CI scaffolding and repo tooling metadata (Dockerfile, workflows, submodules, editor configs); removed ARM64 dummy assembler/linker code.

Reviewed changes

Copilot reviewed 48 out of 536 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
asm_token_fix_plan.md	Documents migration plan for new ASM token layout in Thumb assembler
arm64-link.c	Removes AArch64 linker/relocation implementation from this branch
arm64-asm.c	Removes ARM64 dummy assembler stub
arm-tok.h	Removes legacy ARM token definitions header
arm-thumb-scratch.c	Adds scratch register allocation/save/restore helpers for Thumb codegen
arm-thumb-defs.h	Introduces Thumb target defs, ABI enums, generator state structs, suffix parsing types
arm-thumb-callsite.c	Adds call-site table + IR scan utilities for ABI call layout reconstruction
arch/fpu/arm/fpv5-sp-d16.h	Declares FPv5-SP-D16 floating-point config
arch/fpu/arm/fpv5-sp-d16.c	Defines FPv5-SP-D16 floating-point capabilities/config
arch/armv8m.c	Adds ARMv8-M ArchitectureConfig defaults
arch/arm_aapcs.c	Adds AAPCS-like argument classification + call layout capacity helpers
TCC_GCC_CODEGEN_COMPARISON.md	Documents codegen gaps vs GCC and targeted optimizations
TAL_TOKSTR_OPTIMIZATION.md	Documents plan to reduce peak memory by changing tok_str allocation strategy
REFACTORING_SUMMARY.md	Summarizes “architecture independence” refactor and new machine/opt modules
POSTINC_EMBEDDED_DEREF_PLAN.md	Documents plan for post-inc fusion when deref is embedded in expressions
OPTIMIZATION_VALIDATION_REPORT.md	Records which optimization phases are implemented/validated
OPTIMIZATION_QUICK_WIN_1_FP_CACHE.md	Documents FP-offset cache implementation and tests
OPTIMIZATION_PLAN_V2.md	Updated optimization plan and analysis (LR usage, indexed loads, post-inc, etc.)
OPTIMIZATION_FP_CACHE_IMPROVEMENT_PLAN.md	Documents FP cache “register-agnostic” approach and limitations
MLA_WITH_DEREFERENCES_PLAN.md	Documents plan to enable MLA fusion when MUL operands are dereferences
LICM_IMPLEMENTATION_STATUS.md	Status/report on LICM implementation and known bugs
LICM_IMPLEMENTATION_PLAN.md	Detailed LICM design plan
LAZY_SECTION_LOADING_ALL_PLAN.md	Notes reverted attempt to defer-load all sections and failure reasons
LAZY_SECTION_LOADING.md	Documents lazy section loading design/implementation
FP_CACHE_IR_LEVEL_PLAN.md	Proposes IR-level FP offset caching as a cleaner approach
Dockerfile	Adds Ubuntu-based build/test container with ARM toolchain + QEMU
BUBBLE_SORT_COMPARISON.md	Detailed bubble-sort codegen diff analysis and optimization priorities
AGENTS.md	Adds project agent/guide documentation, structure, build/test notes
.vscode/tasks.json	Adds QEMU run/stop tasks
.vscode/settings.json	Adds terminal auto-approve configuration for tooling
.vscode/launch.json	Adds debug configs (tcc under gdb, QEMU gdb attach)
.vscode/c_cpp_properties.json	Adds IntelliSense settings/defines for ARM Thumb target
.gitmodules	Adds new submodules (newlib, pico-sdk, mibench)
.github/workflows/docker-build.yml	Adds workflow to build/push container image to GHCR
.github/workflows/ci.yml	Adds CI workflow running tests inside GHCR container
.github/prompts/plan-aliasAnalysisStoreLoadOptimizations.prompt.md	Adds prompt doc for alias-analysis/store-load optimization plan
.dockerignore	Adds docker ignore rules for repo build artifacts and tooling dirs
.clang-format	Updates formatting rules (including brace style + column limit)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Defining USING_GLOBALS in a header is very likely to cause multiple-definition / ODR-style linker errors because every translation unit including this header will pull in global definitions from tcc.h. Move #define USING_GLOBALS into exactly one .c file (typically the implementation file) or remove it entirely from the header and include only what’s needed (e.g., forward-declare FloatingPointConfig or include a smaller header that doesn’t depend on USING_GLOBALS).

Suggested change

#define USING_GLOBALS

[Copilot]

layout->capacity is increased without reallocating the backing arrays (the tcc_realloc/memset calls are commented out). This can desynchronize capacity from actual allocated sizes and can also cause tcc_abi_call_layout_ensure_capacity() to incorrectly skip reallocation later, leading to out-of-bounds writes. Replace this block with a call to tcc_abi_call_layout_ensure_capacity(layout, needed) (or fully re-enable correct realloc/zero-init here) and keep capacity consistent with the real allocations.

Suggested change

	if (layout->capacity < needed)
	{
	int new_cap = layout->capacity ? layout->capacity : 8;
	while (new_cap < needed)
	new_cap *= 2;
	// layout->locs = (TCCAbiArgLoc *)tcc_realloc(layout->locs, sizeof(TCCAbiArgLoc) * new_cap);
	/* Zero new tail for determinism. */
	// memset(&layout->locs[layout->capacity], 0, sizeof(TCCAbiArgLoc) * (new_cap - layout->capacity));
	layout->capacity = new_cap;
	}
	tcc_abi_call_layout_ensure_capacity(layout, needed);

[Copilot]

When you emit a single th_push(regs_to_save) with a register mask, the architectural stack layout is in register-number order, but scratch_push_stack records registers in allocation order (result.regs[]). If allocation order differs from ascending register number order, later POP validation/deferral and the “auto restore” path can pop registers in an order that doesn’t match what’s actually on the stack, restoring wrong values. A robust fix is to track pushes as masks on a push-mask stack (LIFO of uint32_t masks) instead of per-register, or record the saved registers in the exact order they are stored on the stack (ascending register number) for each mask push.

[Copilot]

These fprintf(stderr, ...) calls are unconditional (not behind #ifdef ARM_THUMB_DEBUG_SCRATCH) and will spam stderr in normal builds/tests. Wrap them in the same debug macro used elsewhere in this file (or route through a project logging macro that can be compiled out).

[Copilot]

These fprintf(stderr, ...) calls are unconditional (not behind #ifdef ARM_THUMB_DEBUG_SCRATCH) and will spam stderr in normal builds/tests. Wrap them in the same debug macro used elsewhere in this file (or route through a project logging macro that can be compiled out).

[Copilot]

Committing editor settings that auto-approve terminal commands can create a footgun for contributors (commands may run without explicit confirmation). Consider removing this from the repo, scoping it to a local/user settings example, or setting these to false by default with documentation on how to opt-in locally.

Suggested change

	"make": true,
	"heaptrack_print": true
	"make": false,
	"heaptrack_print": false

[Copilot]

BreakBeforeBraces: Allman conflicts with the brace style documented in AGENTS.md (K&R style). This will cause churn and inconsistent formatting expectations. Align the documented style and .clang-format settings (either update the docs to Allman or adjust .clang-format to match the documented K&R approach).

Suggested change

	BreakBeforeBraces: Allman
	BreakBeforeBraces: Attach