[v3] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64

Message ID CAMe9rOrXqn2=00FT4Rw2Eq7q9y_uCXYZ=nxd3LMOtSv9RJNg+Q@mail.gmail.com
State Superseded
Delegated to: Carlos O'Donell
Headers
Series [v3] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64 |

Commit Message

H.J. Lu March 9, 2021, 12:09 a.m. UTC
  On Mon, Mar 8, 2021 at 2:28 PM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 3/2/21 9:21 AM, H.J. Lu wrote:
> > On Mon, Mar 1, 2021 at 5:30 AM Carlos O'Donell <carlos@redhat.com> wrote:
> >>
> >> On 2/2/21 2:12 PM, H.J. Lu via Libc-alpha wrote:
> >>> Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
> >>> "movq reg64, mem64" to store 64-bit constant in TCB.
> >>> ---
> >>>  sysdeps/x86_64/nptl/tls.h | 11 +++++++++++
> >>>  1 file changed, 11 insertions(+)
> >>>
> >>> diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
> >>> index 20f0958780..9ec8703e45 100644
> >>> --- a/sysdeps/x86_64/nptl/tls.h
> >>> +++ b/sysdeps/x86_64/nptl/tls.h
> >>> @@ -269,6 +269,11 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
> >>>         asm volatile ("movl %0,%%fs:%P1" :                                  \
> >>>                    : IMM_MODE (value),                                      \
> >>>                      "i" (offsetof (struct pthread, member)));              \
> >>> +     else if (__builtin_constant_p (value)                                 \
> >>> +           && (int64_t) (int32_t) (uintptr_t) value != (uintptr_t) value)  \
> >>> +       asm volatile ("movq %0,%%fs:%P1" :                                  \
> >>> +                  : "r" (value),                                           \
> >>> +                    "i" (offsetof (struct pthread, member)));              \
> >>
> >> (1) Move conditional into 'else /* 8 */' section.
> >>
> >> The blocks of conditionals are predicated on the size of the member we are
> >> about to write to.
> >>
> >> In the block below...
> >>
> >>>       else /* 8 */                                                          \
> >>
> >> ... here, we are about to write value into a member that is size 8.
> >>
> >> Your code changes the logical construction of the code, but in way that makes
> >> it more difficult to understand.
> >>
> >> We previously had:
> >>
> >> if (sizeof() == 1)
> >>
> >> else if (sizeof() == 4)
> >>
> >> else /* Assume 8 */
> >>
> >> In your case we must already be in the 'else /* Assume 8 */' because otherwise
> >> we've be writing a 64-bit constant into a < 64-bit structure member.
> >>
> >> I think we should put your code into the else clause.
> >>
> >> 272      else /* 8 */                                                             \
> >> 273        {                                                                      \
> >>
> >>              if (__builtin_constant_p (value)
> >>                  && ([the value is a >32-bit constant])
> >>                asm volatile ([use movq reg64, mem64]);
> >>              else
> >>
> >> 274          asm volatile ("movq %q0,%%fs:%P1" :                                  \
> >> 275                        : IMM_MODE ((uint64_t) cast_to_integer (value)),       \
> >> 276                          "i" (offsetof (struct pthread, member)));            \
> >> 277        }})
> >>
> >> (2) What if gcc can't prove it's constant?
> >>
> >> If the constant is >32-bit, but gcc can't prove it's constant, then don't we
> >> try to put a >32-bit constant into a 32-bit immediate?
> >>
> >> Shouldn't the code be structured the other way around?
> >>
> >> e.g.
> >>
> >> else /* 8 */
> >> {
> >>   if (__builtin_constant_p (value)
> >>       && ([the value is a <= 32-bit constant])
> >>     asm volatile ([use movq imm32, mem64]);
> >>   else
> >>     asm volatile ([use movq reg64, mem64]);
> >> }
> >>
> >> This way the code is always correct?
> >
> > I changed it to
> >
> >          if (!__builtin_constant_p (value)                                    \
> >              || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
> >            asm volatile ("movq %q0,%%fs:%P1" :                                \
> >                          : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
> >                            "i" (offsetof (struct pthread, member)));          \
> >          else                                                                 \
> >            asm volatile ("movq %0,%%fs:%P1" :                                 \
> >                          : "r" (value),                                       \
> >                            "i" (offsetof (struct pthread, member)));          \
> >
> >>>         {                                                                   \
> >>>        asm volatile ("movq %q0,%%fs:%P1" :                                  \
> >>> @@ -294,6 +299,12 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
> >>>                    : IMM_MODE (value),                                      \
> >>>                      "i" (offsetof (struct pthread, member[0])),            \
> >>>                      "r" (idx));                                            \
> >>> +     else if (__builtin_constant_p (value)                                 \
> >>> +           && (int64_t) (int32_t) (uintptr_t) value != (uintptr_t) value)  \
> >>> +       asm volatile ("movq %0,%%fs:%P1(,%q2,8)" :                          \
> >>> +                  : "r" (value),                                           \
> >>> +                    "i" (offsetof (struct pthread, member[0])),            \
> >>> +                    "r" (idx));                                            \
> >>>       else /* 8 */                                                          \
> >>>         {                                                                   \
> >>>        asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :                          \
> >>>
> >>
> >
> > Here is the v2 patch.  OK for master?
> >
> > Thanks.
> >
>
> > From 4fbc9ab67933d662506a01658d4c81a922e50fdf Mon Sep 17 00:00:00 2001
> > From: "H.J. Lu" <hjl.tools@gmail.com>
> > Date: Fri, 8 Jan 2021 15:38:14 -0800
> > Subject: [PATCH v2] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64
> >
> > Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
> > "movq reg64, mem64" to store 64-bit constant.
> > ---
> >  sysdeps/x86_64/nptl/tls.h | 27 ++++++++++++++++++++-------
> >  1 file changed, 20 insertions(+), 7 deletions(-)
> >
> > diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
> > index 20f0958780..0dbd66209c 100644
> > --- a/sysdeps/x86_64/nptl/tls.h
> > +++ b/sysdeps/x86_64/nptl/tls.h
> > @@ -271,9 +271,15 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
> >                      "i" (offsetof (struct pthread, member)));              \
> >       else /* 8 */                                                          \
> >         {                                                                   \
> > -      asm volatile ("movq %q0,%%fs:%P1" :                                  \
> > -                    : IMM_MODE ((uint64_t) cast_to_integer (value)),       \
> > -                      "i" (offsetof (struct pthread, member)));            \
> > +      if (!__builtin_constant_p (value)                                    \
>
> Whis is this a "!__builtin_constant_p?"
>
> I would have expected:
>
> if ([the value is a constant a therefore known quantity]
>     || [it is a 32-bit value])

IMM_MODE constraint is valid only if VALUE isn't a constant or
is a signed 32-bit constant.

> Perhaps a quick single line comment here explaining the conditionals would help?

Done.

Here is the v3 patch with comments.   OK for master?

Thanks.
  

Comments

Carlos O'Donell March 15, 2021, 12:49 p.m. UTC | #1
On 3/8/21 7:09 PM, H.J. Lu wrote:
> On Mon, Mar 8, 2021 at 2:28 PM Carlos O'Donell <carlos@redhat.com> wrote:
>>
>> On 3/2/21 9:21 AM, H.J. Lu wrote:
>>> On Mon, Mar 1, 2021 at 5:30 AM Carlos O'Donell <carlos@redhat.com> wrote:
>>>>
>>>> On 2/2/21 2:12 PM, H.J. Lu via Libc-alpha wrote:
>>>>> Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
>>>>> "movq reg64, mem64" to store 64-bit constant in TCB.
>>>>> ---
>>>>>  sysdeps/x86_64/nptl/tls.h | 11 +++++++++++
>>>>>  1 file changed, 11 insertions(+)
>>>>>
>>>>> diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
>>>>> index 20f0958780..9ec8703e45 100644
>>>>> --- a/sysdeps/x86_64/nptl/tls.h
>>>>> +++ b/sysdeps/x86_64/nptl/tls.h
>>>>> @@ -269,6 +269,11 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
>>>>>         asm volatile ("movl %0,%%fs:%P1" :                                  \
>>>>>                    : IMM_MODE (value),                                      \
>>>>>                      "i" (offsetof (struct pthread, member)));              \
>>>>> +     else if (__builtin_constant_p (value)                                 \
>>>>> +           && (int64_t) (int32_t) (uintptr_t) value != (uintptr_t) value)  \
>>>>> +       asm volatile ("movq %0,%%fs:%P1" :                                  \
>>>>> +                  : "r" (value),                                           \
>>>>> +                    "i" (offsetof (struct pthread, member)));              \
>>>>
>>>> (1) Move conditional into 'else /* 8 */' section.
>>>>
>>>> The blocks of conditionals are predicated on the size of the member we are
>>>> about to write to.
>>>>
>>>> In the block below...
>>>>
>>>>>       else /* 8 */                                                          \
>>>>
>>>> ... here, we are about to write value into a member that is size 8.
>>>>
>>>> Your code changes the logical construction of the code, but in way that makes
>>>> it more difficult to understand.
>>>>
>>>> We previously had:
>>>>
>>>> if (sizeof() == 1)
>>>>
>>>> else if (sizeof() == 4)
>>>>
>>>> else /* Assume 8 */
>>>>
>>>> In your case we must already be in the 'else /* Assume 8 */' because otherwise
>>>> we've be writing a 64-bit constant into a < 64-bit structure member.
>>>>
>>>> I think we should put your code into the else clause.
>>>>
>>>> 272      else /* 8 */                                                             \
>>>> 273        {                                                                      \
>>>>
>>>>              if (__builtin_constant_p (value)
>>>>                  && ([the value is a >32-bit constant])
>>>>                asm volatile ([use movq reg64, mem64]);
>>>>              else
>>>>
>>>> 274          asm volatile ("movq %q0,%%fs:%P1" :                                  \
>>>> 275                        : IMM_MODE ((uint64_t) cast_to_integer (value)),       \
>>>> 276                          "i" (offsetof (struct pthread, member)));            \
>>>> 277        }})
>>>>
>>>> (2) What if gcc can't prove it's constant?
>>>>
>>>> If the constant is >32-bit, but gcc can't prove it's constant, then don't we
>>>> try to put a >32-bit constant into a 32-bit immediate?
>>>>
>>>> Shouldn't the code be structured the other way around?
>>>>
>>>> e.g.
>>>>
>>>> else /* 8 */
>>>> {
>>>>   if (__builtin_constant_p (value)
>>>>       && ([the value is a <= 32-bit constant])
>>>>     asm volatile ([use movq imm32, mem64]);
>>>>   else
>>>>     asm volatile ([use movq reg64, mem64]);
>>>> }
>>>>
>>>> This way the code is always correct?
>>>
>>> I changed it to
>>>
>>>          if (!__builtin_constant_p (value)                                    \
>>>              || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
>>>            asm volatile ("movq %q0,%%fs:%P1" :                                \
>>>                          : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
>>>                            "i" (offsetof (struct pthread, member)));          \
>>>          else                                                                 \
>>>            asm volatile ("movq %0,%%fs:%P1" :                                 \
>>>                          : "r" (value),                                       \
>>>                            "i" (offsetof (struct pthread, member)));          \
>>>
>>>>>         {                                                                   \
>>>>>        asm volatile ("movq %q0,%%fs:%P1" :                                  \
>>>>> @@ -294,6 +299,12 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
>>>>>                    : IMM_MODE (value),                                      \
>>>>>                      "i" (offsetof (struct pthread, member[0])),            \
>>>>>                      "r" (idx));                                            \
>>>>> +     else if (__builtin_constant_p (value)                                 \
>>>>> +           && (int64_t) (int32_t) (uintptr_t) value != (uintptr_t) value)  \
>>>>> +       asm volatile ("movq %0,%%fs:%P1(,%q2,8)" :                          \
>>>>> +                  : "r" (value),                                           \
>>>>> +                    "i" (offsetof (struct pthread, member[0])),            \
>>>>> +                    "r" (idx));                                            \
>>>>>       else /* 8 */                                                          \
>>>>>         {                                                                   \
>>>>>        asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :                          \
>>>>>
>>>>
>>>
>>> Here is the v2 patch.  OK for master?
>>>
>>> Thanks.
>>>
>>
>>> From 4fbc9ab67933d662506a01658d4c81a922e50fdf Mon Sep 17 00:00:00 2001
>>> From: "H.J. Lu" <hjl.tools@gmail.com>
>>> Date: Fri, 8 Jan 2021 15:38:14 -0800
>>> Subject: [PATCH v2] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64
>>>
>>> Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
>>> "movq reg64, mem64" to store 64-bit constant.
>>> ---
>>>  sysdeps/x86_64/nptl/tls.h | 27 ++++++++++++++++++++-------
>>>  1 file changed, 20 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
>>> index 20f0958780..0dbd66209c 100644
>>> --- a/sysdeps/x86_64/nptl/tls.h
>>> +++ b/sysdeps/x86_64/nptl/tls.h
>>> @@ -271,9 +271,15 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
>>>                      "i" (offsetof (struct pthread, member)));              \
>>>       else /* 8 */                                                          \
>>>         {                                                                   \
>>> -      asm volatile ("movq %q0,%%fs:%P1" :                                  \
>>> -                    : IMM_MODE ((uint64_t) cast_to_integer (value)),       \
>>> -                      "i" (offsetof (struct pthread, member)));            \
>>> +      if (!__builtin_constant_p (value)                                    \
>>
>> Whis is this a "!__builtin_constant_p?"
>>
>> I would have expected:
>>
>> if ([the value is a constant a therefore known quantity]
>>     || [it is a 32-bit value])
> 
> IMM_MODE constraint is valid only if VALUE isn't a constant or
> is a signed 32-bit constant.
> 
>> Perhaps a quick single line comment here explaining the conditionals would help?
> 
> Done.
> 
> Here is the v3 patch with comments.   OK for master?

No, this still doesn't make logical sense to me.

See my comments below.

> From 0a5d266206be231ab11f3cfa0e213db1e04cd7bd Mon Sep 17 00:00:00 2001
> From: "H.J. Lu" <hjl.tools@gmail.com>
> Date: Fri, 8 Jan 2021 15:38:14 -0800
> Subject: [PATCH v3] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64
> 
> Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
> "movq reg64, mem64" to store 64-bit constant.
> ---
>  sysdeps/x86_64/nptl/tls.h | 31 ++++++++++++++++++++++++-------
>  1 file changed, 24 insertions(+), 7 deletions(-)
> 
> diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
> index 20f0958780..acea951d49 100644
> --- a/sysdeps/x86_64/nptl/tls.h
> +++ b/sysdeps/x86_64/nptl/tls.h
> @@ -271,9 +271,17 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
>  		       "i" (offsetof (struct pthread, member)));	      \
>       else /* 8 */							      \
>         {								      \
> -	 asm volatile ("movq %q0,%%fs:%P1" :				      \
> -		       : IMM_MODE ((uint64_t) cast_to_integer (value)),	      \
> -			 "i" (offsetof (struct pthread, member)));	      \
> +	 /* NB: IMM_MODE constraint is valid only if VALUE isn't a constant   \
> +	    or is a signed 32-bit constant.  */ 			      \

We are trying to gate this asm volatile so it is only used by the compiler
when the immediate is known to be 32-bit wide.

The old code was using IMM_MODE, which resolves to "nr" which should as a
constraint attempt to pass in a compile-time constant 32-bit value. My
understanding is that this constraint could fail to be met and the we get
a compile-time failure?

The value of !__builtin_constant_p only tells you that the value is not a
constant. What prevents that value from being a 64-bit register with a 64-bit
value which can't be encoded via 'nr' with movq?

Why isn't the right solution?

if (__builtin_constant_p (value)
    && (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value))
  /* The value is constant and known to be <= 32-bit immediate value.
     In this case we allow the compiler to optimize and use movq with
     an immediate value.  */
  ...
else
  /* The value is not known and so we must assume it is at worst a 64-bit value.
     We can't optimize this and so we must use a register for the move. */
  ...

> +	 if (!__builtin_constant_p (value)				      \
> +	     || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
> +	   asm volatile ("movq %q0,%%fs:%P1" :				      \
> +			 : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
> +			   "i" (offsetof (struct pthread, member)));	      \
> +	 else								      \
> +	   asm volatile ("movq %0,%%fs:%P1" :				      \
> +			 : "r" (value),					      \
> +			   "i" (offsetof (struct pthread, member)));	      \
>         }})
>  
>  
> @@ -296,10 +304,19 @@ _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
>  		       "r" (idx));					      \
>       else /* 8 */							      \
>         {								      \
> -	 asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :			      \
> -		       : IMM_MODE ((uint64_t) cast_to_integer (value)),	      \
> -			 "i" (offsetof (struct pthread, member[0])),	      \
> -			 "r" (idx));					      \
> +	 /* NB: IMM_MODE constraint is valid only if VALUE isn't a constant   \
> +	    or is a signed 32-bit constant.  */ 			      \
> +	 if (!__builtin_constant_p (value)				      \
> +	     || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
> +	   asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :			      \
> +			 : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
> +			   "i" (offsetof (struct pthread, member[0])),	      \
> +			   "r" (idx));					      \
> +	 else								      \
> +	   asm volatile ("movq %0,%%fs:%P1(,%q2,8)" :			      \
> +			 : "r" (value),					      \
> +			   "i" (offsetof (struct pthread, member[0])),	      \
> +			   "r" (idx));					      \
>         }})
>  
>  
> -- 
> 2.29.2
>
  

Patch

From 0a5d266206be231ab11f3cfa0e213db1e04cd7bd Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Fri, 8 Jan 2021 15:38:14 -0800
Subject: [PATCH v3] x86_64: Update THREAD_SETMEM/THREAD_SETMEM_NC for IMM64

Since there is only "movq imm32s, mem64" and no "movq imm64, mem64", use
"movq reg64, mem64" to store 64-bit constant.
---
 sysdeps/x86_64/nptl/tls.h | 31 ++++++++++++++++++++++++-------
 1 file changed, 24 insertions(+), 7 deletions(-)

diff --git a/sysdeps/x86_64/nptl/tls.h b/sysdeps/x86_64/nptl/tls.h
index 20f0958780..acea951d49 100644
--- a/sysdeps/x86_64/nptl/tls.h
+++ b/sysdeps/x86_64/nptl/tls.h
@@ -271,9 +271,17 @@  _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
 		       "i" (offsetof (struct pthread, member)));	      \
      else /* 8 */							      \
        {								      \
-	 asm volatile ("movq %q0,%%fs:%P1" :				      \
-		       : IMM_MODE ((uint64_t) cast_to_integer (value)),	      \
-			 "i" (offsetof (struct pthread, member)));	      \
+	 /* NB: IMM_MODE constraint is valid only if VALUE isn't a constant   \
+	    or is a signed 32-bit constant.  */ 			      \
+	 if (!__builtin_constant_p (value)				      \
+	     || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
+	   asm volatile ("movq %q0,%%fs:%P1" :				      \
+			 : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
+			   "i" (offsetof (struct pthread, member)));	      \
+	 else								      \
+	   asm volatile ("movq %0,%%fs:%P1" :				      \
+			 : "r" (value),					      \
+			   "i" (offsetof (struct pthread, member)));	      \
        }})
 
 
@@ -296,10 +304,19 @@  _Static_assert (offsetof (tcbhead_t, __glibc_unused2) == 0x80,
 		       "r" (idx));					      \
      else /* 8 */							      \
        {								      \
-	 asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :			      \
-		       : IMM_MODE ((uint64_t) cast_to_integer (value)),	      \
-			 "i" (offsetof (struct pthread, member[0])),	      \
-			 "r" (idx));					      \
+	 /* NB: IMM_MODE constraint is valid only if VALUE isn't a constant   \
+	    or is a signed 32-bit constant.  */ 			      \
+	 if (!__builtin_constant_p (value)				      \
+	     || (int64_t) (int32_t) (uintptr_t) value == (uintptr_t) value)   \
+	   asm volatile ("movq %q0,%%fs:%P1(,%q2,8)" :			      \
+			 : IMM_MODE ((uint64_t) cast_to_integer (value)),     \
+			   "i" (offsetof (struct pthread, member[0])),	      \
+			   "r" (idx));					      \
+	 else								      \
+	   asm volatile ("movq %0,%%fs:%P1(,%q2,8)" :			      \
+			 : "r" (value),					      \
+			   "i" (offsetof (struct pthread, member[0])),	      \
+			   "r" (idx));					      \
        }})
 
 
-- 
2.29.2