cmd/compile: run write barriers before LLVM emission

[zhouguangyuan0718]

Summary

• run the Go writebarrier SSA pass in LLVM mode before expand calls
• emit LLVM from that form so pointer stores retain Go write-barrier semantics
• lower scalar barriers to llvm.go.gc.write.barrier while preserving the current branchy Go control flow
• preserve the native raw AuxCall as physical ABI authority for runtime.wbZero/runtime.wbMove, while restoring their semantic LLVM parameters to ptr
• unconditionally attach go-async-unsafe to every LLVM-emitted function; LLVM PR cmd/objview: add metadata-aware text disassembly #19 therefore emits whole-function PCDATA_UnsafePoint=-2
• add codegen, IR, GoObj, concurrent-GC, aggregate move/zero, and stack-growth coverage

Unsafe-point model

This version intentionally uses the complete conservative model requested for the first functional write-barrier implementation: every LLVM-emitted function is asynchronous-unsafe for its entire body.

There is no interval scan, unsafe-point start/end intrinsic, SelectionDAG marker, or partial range protocol in this PR. Removing the old WriteMem marker also restores normal optimization; the memory-order test accepts inlining and load forwarding while still checking the write barrier and resulting store.

Exact interval marking is deferred. The previous range-emission prototype is preserved at codex/post-expand-llvm-emission-ranges-backup.

Rebase and prerequisites

Rebased onto go1.27.master@912359eeecb1b00c280bd857e220857de32dbf83, including the merged post-construction nilcheck-marker expansion (#38). This branch preserves that implementation unchanged and does not restore the old BlockEnds CFG bypass.

Pointer allocas are handled exclusively by the #33 / LLVM #18 deopt ptrmap model. This PR contains no capture analysis, fixed-stack preload, alloca relocation, or post-call write-back implementation.

Merged prerequisites:

• AArch64 write-barrier opcode lowering: AArch64: lower Go write barrier intrinsics llvm-project#17
• pointer-allocation deopt ptrmaps: GoObj: describe fixed alloca roots in statepoint deopts llvm-project#18 and cmd/llvmplugin: describe pointer allocas as memory roots #33
• whole-function unsafe GoObj PCDATA: GoObj: mark asynchronous-unsafe functions llvm-project#19
• runtime.newproc pointer typing: cmd/compile: preserve newproc pointer argument in LLVM IR #34
• supported statepoint pointer parameter attributes: cmd/llvmplugin: accept non-ABI pointer parameter attributes #35
• post-construction explicit nilcheck expansion: cmd/compile: expand LLVM nil checks after IR construction #38

Open prerequisite:

• preserve fixed incoming argument homes across consecutive statepoints: SelectionDAG: reuse fixed argument homes across statepoints llvm-project#22

Statepoint and SSA integration

• keep original calls alive until every precomputed statepoint live set has been consumed; LLVM block layout is not CFG dominance order
• validate the actual post-nilcheck call queries for mixedABI: four ordinary safepoints select bitmap {2}, while morestack selects entry bitmap {2,4,18}
• recognize only the exact dead OpNilCheck tombstone that writebarrier leaves before early deadcode
• remove the unreachable BlockFirst lowering and all four BlockInvalid skip paths; unexpected block kinds now use the existing fail-closed path
• handle runtime.newproc, wbZero, and wbMove in one llvmStaticCallSignature path: validate each raw ABIInternal uintptr signature, then construct the semantic LLVM pointer signature without ptrtoint adapter values
• keep wbZero/wbMove as direct gc-leaf calls; focused checks reject any statepoint that targets either helper
• update writebarrier helper checks to distinguish the ordinary call from two expanded runtime.panicmem statepoints

Validation

• clean make.bash toolchain rebuild with the latest vendored go-llvm and exact LLVM payload
• installed and freshly built statepoint plugins have identical SHA-256
• plugin CTest: 48/48
• go test -count=1 -tags="llvm23 dynamicllvm" cmd/compile/internal/ssa cmd/llvmtoolexec
• direct IR verification, opt -passes=default<O2>, RS4GC rewritten verification, llc -verify-machineinstrs, MIR and object generation through focused tests
• cmd/internal/testdir Test/LLVM: all 51 codegen and all 47 runtime whitelist cases
• ABI differential, nilcheck, alloca-statepoint, semantic pointer wbZero/wbMove, writebarrier IR, concurrent GC, and stack-growth suites
• LLVM AArch64/X86 fixed argument home lit tests: 2/2
• gofmt, diff-only clang-format, and git diff --check

Scope

This first functional version keeps the branches generated by the Go writebarrier pass. A later optimization can use a more constrained late-expanded intrinsic and exact unsafe intervals. StackObjects precision and the postponed linker indirect-call nosplit graph issue are not included.