Re: [PATCH] x86/boot: Use __ASM_SIZE() to reduce ifdeffery in cpuflags.c

[Uros Bizjak]

On Thu, Jul 18, 2024 at 10:56 AM H. Peter Anvin <hpa@xxxxxxxxx> wrote:
>
> On July 18, 2024 1:52:17 AM PDT, Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> >On Thu, Jul 18, 2024 at 8:36 AM H. Peter Anvin <hpa@xxxxxxxxx> wrote:
> >>
> >> On July 17, 2024 11:32:18 PM PDT, Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> >> >Use __ASM_SIZE() macro to add correct insn suffix to pushf/popf.
> >> >
> >> >Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> >> >Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> >> >Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> >> >Cc: Borislav Petkov <bp@xxxxxxxxx>
> >> >Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> >> >Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> >> >---
> >> > arch/x86/boot/cpuflags.c | 10 +++-------
> >> > 1 file changed, 3 insertions(+), 7 deletions(-)
> >> >
> >> >diff --git a/arch/x86/boot/cpuflags.c b/arch/x86/boot/cpuflags.c
> >> >index d75237ba7ce9..aacabe431fd5 100644
> >> >--- a/arch/x86/boot/cpuflags.c
> >> >+++ b/arch/x86/boot/cpuflags.c
> >> >@@ -2,6 +2,7 @@
> >> > #include <linux/types.h>
> >> > #include "bitops.h"
> >> >
> >> >+#include <asm/asm.h>
> >> > #include <asm/processor-flags.h>
> >> > #include <asm/required-features.h>
> >> > #include <asm/msr-index.h>
> >> >@@ -36,13 +37,8 @@ static int has_fpu(void)
> >> >  * compressed/ directory where it may be 64-bit code, and thus needs
> >> >  * to be 'pushfq' or 'popfq' in that case.
> >> >  */
> >> >-#ifdef __x86_64__
> >> >-#define PUSHF "pushfq"
> >> >-#define POPF "popfq"
> >> >-#else
> >> >-#define PUSHF "pushfl"
> >> >-#define POPF "popfl"
> >> >-#endif
> >> >+#define PUSHF __ASM_SIZE(pushf)
> >> >+#define POPF __ASM_SIZE(popf)
> >> >
> >> > int has_eflag(unsigned long mask)
> >> > {
> >>
> >> Just use pushf/popf. gas hasn't needed that suffix for a long time as far as I know.
> >
> >Unfortunately, clang does not do the right thing when pushf/popf
> >without suffix are used.
> >
> >arch/x86/boot/cpuflags.c compiles to:
> >
> >00000000 <has_eflag>:
> >   0:    9c                       pushf
> >   1:    9c                       pushf
> >   2:    66 5a                    pop    %edx
> >   4:    66 89 d1                 mov    %edx,%ecx
> >   7:    66 31 c1                 xor    %eax,%ecx
> >   a:    66 51                    push   %ecx
> >   c:    9d                       popf
> >   d:    9c                       pushf
> >   e:    66 59                    pop    %ecx
> >  10:    9d                       popf
> >  11:    66 31 ca                 xor    %ecx,%edx
> >  14:    66 31 c9                 xor    %ecx,%ecx
> >  17:    66 85 c2                 test   %eax,%edx
> >  1a:    0f 95 c1                 setne  %cl
> >  1d:    66 89 c8                 mov    %ecx,%eax
> >  20:    66 c3                    retl
> >
> >instead of:
> >
> >00000000 <has_eflag>:
> >   0:    66 9c                    pushfl
> >   2:    66 9c                    pushfl
> >   4:    66 5a                    pop    %edx
> >   6:    66 89 d1                 mov    %edx,%ecx
> >   9:    66 31 c1                 xor    %eax,%ecx
> >   c:    66 51                    push   %ecx
> >   e:    66 9d                    popfl
> >  10:    66 9c                    pushfl
> >  12:    66 59                    pop    %ecx
> >  14:    66 9d                    popfl
> >  16:    66 31 ca                 xor    %ecx,%edx
> >  19:    66 31 c9                 xor    %ecx,%ecx
> >  1c:    66 85 c2                 test   %eax,%edx
> >  1f:    0f 95 c1                 setne  %cl
> >  22:    66 89 c8                 mov    %ecx,%eax
> >  25:    66 c3                    retl
> >
> >Please note missing 0x66 operand size override prefixes with pushfl
> >and popfl. This is 16bit code, operand prefixes are mandatory to push
> >32-bit EFLAGS register (ID flag lives in bit 21).
> >
> >So, the original patch is the way to go.
> >
> >Uros.
> >
>
> You do know that has_eflag can be completely elided on x86-64, or you can use %z with one of the register operands.

It is 32-bit PUSHFL insn that requires the "L" suffix in 16-bit code.
This is x86_32 issue and clang was just lucky that the instruction was
always defined with explicit L suffix. Please note that PUSHF has no
register operand, so %z can't be used.

> One more reason why clang really needs to shape up.

Indeed.

Thanks,
Uros.