K2: objects and the link

[tamnd]

This tracks K2, objects and the link. Every unit of x86-64 defconfig compiles and assembles with rucc, and vmlinux links. The section, symvers and vec-audit differentials are clean. Bring-up flags are allowed for -m16 and -m32 only, since those are the i686 parts that K4 delivers. The full text is in docs/plan/15-milestones.md.

Harness

• rk sections-diff and rk symvers-diff
• rk vec-audit, which finds vector and x87 instructions in objects built without the FPU
• rk modules-audit
• The RK_BRINGUP path in the shim, used only for -m16 and -m32. Done in Retire the as bring-up class and refuse unknown classes #29, where the as class is retired and unknown classes are refused
• rk objtool-report, run early at K2's midpoint to answer open question 2

rucc

• Assembler expressions and relaxation
• The macro processor
• One assembler stream per object
• asm goto with bodies and outputs
• x86 system instructions and operands from the inventory
• Relocations, notes, groups and empty sections
• Kernel code model and non-PIC with R_X86_64_32S, weak symbols without GOT
• -mgeneral-regs-only, -mno-sse and -mno-80387 (-mgeneral-regs-only and -mno-sse as a hard constraint on x86-64 rucc#2277). Done in Keep the vector registers and the x87 stack out under -mno-sse and -mgeneral-regs-only rucc#2318, merged
• 8-byte stack alignment. The boundary is honored in Honor -mpreferred-stack-boundary on x86-64 rucc#2315, merged, and =3 is taken with -mno-sse in Keep the vector registers and the x87 stack out under -mno-sse and -mgeneral-regs-only rucc#2318, merged
• Stack protector guard register and symbol (The -mstack-protector-guard flags, and stack protector on AArch64 rucc#2279). x86-64 is done in Honor the -mstack-protector-guard flags on x86-64 rucc#2319, merged, including gs as the default under -mcmodel=kernel. AArch64 is K6, as The -mstack-protector-guard flags, and stack protector on AArch64 rucc#2279 says, and stays open there
• Retpoline, rethunk, SLS and the IBT jump table rule, -fno-jump-tables (Retpoline, return thunk, SLS and -fno-jump-tables on x86-64 rucc#2280). The extern thunks landed in Honor the x86-64 speculation hardening flags rucc#2331, and the keep attributes are in Honor function_return("keep") and indirect_branch("keep") rucc#2344, merged. thunk-inline for the vDSO, and thunk with it, are done in codegen: honor thunk-inline and thunk for -mindirect-branch and -mfunction-return rucc#2350, merged
• -fzero-call-used-regs=used-gpr and the attribute (Implement -fzero-call-used-regs rucc#2281). Done in Honor -fzero-call-used-regs and the zero_call_used_regs attribute rucc#2345, merged. The vector choices, which the kernel does not use, are Zero the vector registers under -fzero-call-used-regs=used and all rucc#2335
• The __builtin_constant_p contract, global register variables and the missing inline asm modifiers. The contract is done in opt: follow __builtin_constant_p answers to the branches they decide rucc#2347, merged, and current_stack_pointer as a global register variable is done in codegen: the stack pointer as a global register, and three inline asm fixes rucc#2383, merged. The modifiers %V, %a, %z, %n and %H are done in codegen: the x86-64 inline asm modifiers %V, %a, %z, %n and %H rucc#2386, merged.

Corpora

• kernel-asm graded in rucc-compat. The harness is in Add kernel-asm, the assembler differential over an rk build of Linux 7.2.8 rucc-compat#180, merged. The grade itself comes from the nightly against a GCC build
• rucc-corpus facets: constant-p-after-inline, asm-goto, asm-local-labels, gas-macros, mcmodel-kernel, general-regs-only. Done in Add the asm-goto, asm-local-labels and gas-macros facets, the inline assembly the kernel writes rucc-corpus#125 and Add the constant-p-after-inline, mcmodel-kernel and general-regs-only facets, the way the kernel is built rucc-corpus#126, merged

Exit criterion

• rk build 7.2.8 --row X64 --config defconfig with RK_BRINGUP=m16 produces vmlinux and bzImage. All 3042 units compile, and the kernel boots under QEMU with every smoke check passing. The last fixes were Do not count blocks that only jump against the range walk depth rucc#2536, #2537, #2540, #2542 and #2544
• rk config-diff and rk flags-diff are clean, with the K1 expected differences gone. On 7.2.8 the config diff has 6 symbols, all explained, and the flags diff has none
• rk sections-diff and rk symvers-diff show no difference outside the rules of 09.4. On 7.2.8 defconfig sections-diff has 200 differences, all explained in sections-divergences.toml (Explain expected sections-diff differences #44), after asm: write each alias just after what it names, as gcc does rucc#2824 and opt: the output of an asm goto is not known to sccp rucc#2825. symvers-diff finds the same exports and CRCs
• rk vec-audit finds nothing. Run with --reference, the only hits are the six FPU users that gcc builds the same way
• Every .S unit and every GCC -S output assembles byte for byte the same, or has a recorded explanation. kernel-asm is 3042 of 3042 after asm: lay out gas's first round, where a later section's labels are worth their place in their piece rucc#2821 and asm: pad .eh_frame records to four bytes and the last to a pointer, as gas does rucc#2822
• The same on 6.12.y and 6.1.y. On 6.12.111 defconfig all 2934 units build, sections-diff has 399 differences and all are explained, kernel-asm is 2934 of 2934 and objtool reports nothing after Write functions in the order gcc writes them rucc#2859. On 6.1.188 defconfig sections-diff has 382 differences, all explained, kernel-asm is 2763 of 2763 and objtool reports nothing
• RK_TWICE=1 finds every object identical across two compiles. On 7.2.8 defconfig with --twice, all 3046 compiles wrote the same bytes the second time and all 3042 units built

[tamnd]

tamnd/rucc#2315 is up for the stack alignment item. rucc now honors -mpreferred-stack-boundary=N on x86-64 for N from 4 to 12. Frames get rounded to the boundary, and a local that needs more gets realigned with gcc's push %rbp; mov %rsp,%rbp; and sequence. The kernel's =3 is still refused, the same way gcc refuses it while SSE is on, so the real unblock is -mno-sse (tamnd/rucc#2277). I'm taking that next. I split the checklist line in two so each half can be ticked on its own.

[tamnd]

Opened tamnd/rucc#2318 for #2277, stacked on #2315. rucc now takes -mno-sse, -mno-mmx, -mno-sse2, -mno-80387, -mno-fp-ret-in-387 and -msoft-float on x86-64, and -mgeneral-regs-only on both arches. With those flags no function gets a vector or x87 register, the variadic save area is general purpose registers only, and a function that uses a float or double is refused with the type named, like gcc does. -mpreferred-stack-boundary=3 is taken once SSE is off, so the full x86-64 kernel flag set now goes through the driver. The only one still refused is -mcmodel=kernel (#2275). I checked struct copies, memset, memcpy and a printk-style variadic under those flags and saw no xmm and no realigned frame.

[tamnd]

#22 is merged, so rk now has sections-diff, symvers-diff, vec-audit, modules-audit and objtool-report, and the kernel workflow saves sections.json and vec.json. In rucc, -fno-pic and -fno-pie (tamnd/rucc#2316) and -mcmodel=kernel (tamnd/rucc#2317) are merged. Under the kernel model an address is a movq $sym with an R_X86_64_32S relocation, weak variables are read directly, and an object has no GOT relocations left. -mpreferred-stack-boundary for 4 and up is merged too (tamnd/rucc#2315). =3 waits on tamnd/rucc#2318.

[tamnd]

Opened tamnd/rucc#2319 for the stack protector guard on x86-64, stacked on #2318. The canary can now be read from %gs:__ref_stack_chk_guard(%rip) (6.13 and later), from %gs:40 (older kernels) or from the global __stack_chk_guard. I checked each case against gcc 13.3 on server3 and the instructions and relocations match. Two pieces are still open: gs as the default register under -mcmodel=kernel, which is what gcc-x86_64-has-stack-protector.sh checks for and which waits on #2317, and the AArch64 sp_el0 guard, which is K6 work.

[tamnd]

tamnd/rucc#2328 is up for the mitigations. The extern retpoline and return thunks, the cs prefix, -mharden-sls= and -fno-jump-tables now work and match gcc 13 byte for byte on my test files. The keep attributes work too. thunk-inline is still refused. The vDSO uses it, so I opened tamnd/rucc#2326 for it. I also fixed the stack guard line above, since gs under the kernel code model is already in #2319.

[tamnd]

Two assembler items are done. tamnd/rucc#2324 assembles inline asm templates, file scope asm and .pushsection as one stream per object on ELF. tamnd/rucc#2333 lets .skip, .fill and .org take counts from labels further down or in another section, laying the file out again until they settle, and adds every operator gas has plus absolute operands like $sym, sym(,%rax,8) and %gs:sym with R_X86_64_32S. The ALTERNATIVE padding now comes out the same as gas. The x86-64 speculation flags from tamnd/rucc#2280 also landed in tamnd/rucc#2331, which overlaps tamnd/rucc#2328, so one of the two needs to be dropped.

[tamnd]

The macro processor is in, tamnd/rucc#2336. .macro with defaults, :req, :vararg and named arguments, .rept, .irp, .irpc, the whole .if family, .altmacro with LOCAL and %expr, and .error, .warning, .print and .abort. Inline asm templates go through the same layer, so macros from a top level asm() work in later templates. I also fixed two rucc-target tests that tamnd/rucc#2331 broke on main, in tamnd/rucc#2334.

[tamnd]

The x86 system instructions are in, tamnd/rucc#2338. Control, debug and segment registers, descriptor tables, MSRs, swapgs, sysretq, iretq, far jumps and calls, paging and cache control, port I/O in every width, extended state, SMAP and CET, and monitor, mwait, serialize, enqcmd and movdir64b. Segment prefixes may sit on the instruction's line, in gas's prefix order. The byte for byte check against the full kernel-asm corpus is still ahead with kernel-asm in rucc-compat.

[tamnd]

Notes, groups and empty sections are in, tamnd/rucc#2340. @note is SHT_NOTE, declared empty sections are kept, G and comdat make section groups, .linkonce works on ELF, and R sets SHF_GNU_RETAIN. The relocation half of this item came with tamnd/rucc#2333 (R_X86_64_32S absolutes, R_X86_64_PC64, R_X86_64_8 and _16).

[tamnd]

tamnd/rucc#2341 is merged. asm goto with instructions in the template now compiles, %l is written as the block label, and outputs are valid on every edge the way gcc 11 and later promise. It also works in a function with a variable length array and out of a block with a cleanup handler, since each label edge gets its own landing block. The computed goto half of tamnd/rucc#221 is still open, but the kernel does not need it.

[tamnd]

Update: opened tamnd/rucc#2345 for -fzero-call-used-regs, which clears the same registers as gcc 13 on x86-64 and arm64. tamnd/rucc#2328 is closed because tamnd/rucc#2331 already landed the same thunks, and the one piece it had that main did not, the keep attributes, is now tamnd/rucc#2344. tamnd/rucc#2318 and tamnd/rucc#2319 are rebased on main. tamnd/rucc#2342 fixes the i686 cross name test that is failing on main.