jrtc27 wrote:
> …rlyClobber
>
> Clobber-list registers specified via `~{reg}` in inline asm constraints are
> modified _during_ the asm body, not before inputs are read. As such they must
> NOT carry the EarlyClobber flag — inputs are permitted to use them, matching
> GCC's behaviour.
>
> Before this fix, `InlineAsm::Kind::Clobber` fell through into the
> `InlineAsm::Kind::RegDefEarlyClobber` case in
> `InstrEmitter::EmitSpecialNode`, causing every clobbered register to receive
> `RegState::EarlyClobber` in the generated INLINEASM MachineInstr. This
> prevented the register allocator from ever assigning a clobbered register to
> any input operand.
>
> With a heavy clobber list (e.g. rax, rbx, rcx, rdx, r8-r11, r13-r15 — 11
> registers), `-fno-omit-frame-pointer` (rbp reserved), and a gcov counter
> increment between two inlined asm calls (adding register pressure), the
> allocator ran out of registers and emitted: "inline assembly requires more
> registers than available"
>
> The bug was first observed in the Linux kernel's Curve25519 field arithmetic:
> `fsqr()`/`fsqr2()` in `arch/x86/crypto/curve25519-x86_64.c` use this exact
> pattern — 11-register clobber lists with `+&r` early-clobber outputs — and
> fail to compile with clang under CONFIG_GCOV_KERNEL + frame pointer builds.
> The bug is reproducible with clang 17 through 22 but not with clang 10 or GCC.
>
> Fix: give `Kind::Clobber` its own case that emits `RegState::Define` only,
> without `RegState::EarlyClobber`.
Given the style of this commit message is of a certain nature, was an LLM used
in the creation of this PR? Please see https://llvm.org/docs/AIToolPolicy.html
if so.
https://github.com/llvm/llvm-project/pull/206222
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
