Re: [PATCH] x86/irqflags: Force a register output in native_save_fl()

["H. Peter Anvin" <hpa@zytor.com>]

On December 18, 2025 1:54:33 AM PST, Eric Dumazet <edumazet@google.com> wrote:
>clang is generating very inefficient code for native_save_fl() which
>is used for local_irq_save() in few critical spots.
>
>Allowing the "pop %0" to use memory :
>
>1) forces the compiler to add annoying stack canaries when
>   CONFIG_STACKPROTECTOR_STRONG=y in many places.
>
>2) Almost always is followed by an immediate "move memory,register"
>
>One good example is _raw_spin_lock_irqsave, with 8 extra instructions
>
>ffffffff82067a30 <_raw_spin_lock_irqsave>:
>ffffffff82067a30:       ...
>ffffffff82067a39:       53                      push   %rbx
>
>// Three instructions to ajust the stack, read the per-cpu canary
>// and copy it to 8(%rsp)
>ffffffff82067a3a:       48 83 ec 10             sub    $0x10,%rsp
>ffffffff82067a3e:       65 48 8b 05 da 15 45 02 mov    %gs:0x24515da(%rip),%rax        # <__stack_chk_guard>
>ffffffff82067a46:       48 89 44 24 08          mov    %rax,0x8(%rsp)
>
>ffffffff82067a4b:       9c                      pushf
>
>// instead of pop %rbx, compiler uses 2 instructions.
>ffffffff82067a4c:       8f 04 24                pop    (%rsp)
>ffffffff82067a4f:       48 8b 1c 24             mov    (%rsp),%rbx
>
>ffffffff82067a53:       fa                      cli
>ffffffff82067a54:       b9 01 00 00 00          mov    $0x1,%ecx
>ffffffff82067a59:       31 c0                   xor    %eax,%eax
>ffffffff82067a5b:       f0 0f b1 0f             lock cmpxchg %ecx,(%rdi)
>ffffffff82067a5f:       75 1d                   jne    ffffffff82067a7e <_raw_spin_lock_irqsave+0x4e>
>
>// three instructions to check the stack canary
>ffffffff82067a61:       65 48 8b 05 b7 15 45 02 mov    %gs:0x24515b7(%rip),%rax        # <__stack_chk_guard>
>ffffffff82067a69:       48 3b 44 24 08          cmp    0x8(%rsp),%rax
>ffffffff82067a6e:       75 17                   jne    ffffffff82067a87
>
>...
>
>// One extra instruction to adjust the stack.
>ffffffff82067a73:       48 83 c4 10             add    $0x10,%rsp
>...
>
>// One more instruction in case the stack was mangled.
>ffffffff82067a87:       e8 a4 35 ff ff          call   ffffffff8205b030 <__stack_chk_fail>
>
>This patch changes almost nothing for gcc, but saves 23153 bytes
>of text with clang, while allowing more functions to be inlined.
>
>$ size vmlinux.gcc.before vmlinux.gcc.after vmlinux.clang.before vmlinux.clang.after
>    text     data     bss      dec     hex	filename
>45564911 25005415 4708896 75279221 47cab75	vmlinux.gcc.before
>45564901 25005414 4708896 75279211 47cab6b	vmlinux.gcc.after
>45115647 24638569 5537072 75291288 47cda98	vmlinux.clang.before
>45092494 24638585 5545000 75276079 47c9f2f	vmlinux.clang.after
>
>$ scripts/bloat-o-meter -t vmlinux.clang.before vmlinux.clang.after
>add/remove: 0/3 grow/shrink: 22/533 up/down: 5162/-25088 (-19926)
>Function                                     old     new   delta
>__noinstr_text_start                         640    3584   +2944
>wakeup_cpu_via_vmgexit                      1002    1447    +445
>rcu_tasks_trace_pregp_step                  1052    1454    +402
>snp_kexec_finish                            1290    1527    +237
>check_all_holdout_tasks_trace                909    1106    +197
>x2apic_send_IPI_mask_allbutself               38     198    +160
>hpet_set_rtc_irq_bit                         118     265    +147
>x2apic_send_IPI_mask                          38     184    +146
>ring_buffer_poll_wait                        261     405    +144
>rb_watermark_hit                             253     386    +133
>...
>tcp_wfree                                    402     332     -70
>stacktrace_trigger                           133      62     -71
>w1_touch_bit                                 418     343     -75
>w1_triplet                                   446     370     -76
>link_create                                  980     902     -78
>drain_dead_softirq_workfn                    425     347     -78
>kcryptd_queue_crypt                          253     174     -79
>perf_event_aux_pause                         448     368     -80
>idle_worker_timeout                          320     240     -80
>srcu_funnel_exp_start                        418     333     -85
>call_rcu                                     751     666     -85
>enable_IR_x2apic                             279     191     -88
>bpf_link_free                                432     342     -90
>synchronize_rcu                              497     403     -94
>identify_cpu                                2665    2569     -96
>ftrace_modify_all_code                       355     258     -97
>load_gs_index                                212     104    -108
>verity_end_io                                369     257    -112
>bpf_prog_detach                              672     555    -117
>__x2apic_send_IPI_mask                       552     275    -277
>snp_cleanup_vmsa                             284       -    -284
>__end_rodata                              606208  602112   -4096
>Total: Before=28577927, After=28558001, chg -0.07%
>
>Signed-off-by: Eric Dumazet <edumazet@google.com>
>---
> arch/x86/include/asm/irqflags.h | 7 +------
> 1 file changed, 1 insertion(+), 6 deletions(-)
>
>diff --git a/arch/x86/include/asm/irqflags.h b/arch/x86/include/asm/irqflags.h
>index b30e5474c18e1be63b7c69354c26ae6a6cb02731..a06bdc51b6e3d04c06352c28389b10ec66551a0e 100644
>--- a/arch/x86/include/asm/irqflags.h
>+++ b/arch/x86/include/asm/irqflags.h
>@@ -18,14 +18,9 @@ extern __always_inline unsigned long native_save_fl(void)
> {
> 	unsigned long flags;
> 
>-	/*
>-	 * "=rm" is safe here, because "pop" adjusts the stack before
>-	 * it evaluates its effective address -- this is part of the
>-	 * documented behavior of the "pop" instruction.
>-	 */
> 	asm volatile("# __raw_save_flags\n\t"
> 		     "pushf ; pop %0"
>-		     : "=rm" (flags)
>+		     : "=r" (flags)
> 		     : /* no input */
> 		     : "memory");
> 

Maybe report a bug to the clang team?