Message-ID: <20241112140454.518823-5-mailhol.vincent@wanadoo.fr>
Date: Tue, 12 Nov 2024 22:59:23 +0900
From: Vincent Mailhol <mailhol.vincent@...adoo.fr>
To: Linus Torvalds <torvalds@...ux-foundation.org>,
	Yury Norov <yury.norov@...il.com>,
	Rasmus Villemoes <linux@...musvillemoes.dk>,
	Luc Van Oostenryck <luc.vanoostenryck@...il.com>
Cc: linux-kernel@...r.kernel.org,
	linux-sparse@...r.kernel.org,
	Rikard Falkeborn <rikard.falkeborn@...il.com>,
	Vincent Mailhol <mailhol.vincent@...adoo.fr>
Subject: [PATCH v3 1/2] compiler.h: add _static_assert()

__builtin_constant_p() is known for not always being able to produce
constant expression [1] which led to the introduction of
__is_constexpr() [2]. Because of its dependency on
__builtin_constant_p(), statically_true() suffers from the same
issues.

For example:

  void foo(int a)
  {
  	 /* fail on GCC */
  	BUILD_BUG_ON_ZERO(statically_true(a));

  	 /* fail on both clang and GCC */
  	static char arr[statically_true(a) ? 1 : 2];
  }

For the same reasons why __is_constexpr() was created to remediate
__builtin_constant_p() edge cases, __is_constexpr() can be used to
resolve statically_true()'s limitations.

Note that, somehow, GCC is not always able to fold this:

  __is_constexpr(x) && (x)

It is OK in BUILD_BUG_ON_ZERO() but not in array declarations or in
static_assert():

  void bar(int a)
  {
  	/* success */
  	BUILD_BUG_ON_ZERO(__is_constexpr(a) && (a));

  	/* fail on GCC */
  	static char arr[__is_constexpr(a) && (a) ? 1 : 2];

  	/* fail on GCC */
  	static_assert(__is_constexpr(a) && (a));
  }

Encapsulating the expression in a __builtin_choose_expr() switch
resolves all these failed examples.

Declare a new _statically_true() macro which, by making use of the
__builtin_choose_expr() and __is_constexpr() combo, always produces a
constant expression.

It should be noted that statically_true() still produces better
folding:

  statically_true(!(var * 8 % 8))

always evaluates to true even if var is unknown, whereas

  _statically_true(!(var * 8 % 8))

fails to fold the expression and returns false.

For this reason, usage of _statically_true() should be the exception.
Reflect in the documentation that _statically_true() is less powerful
and that statically_true() is the overall preferred solution.

[1] __builtin_constant_p cannot resolve to const when optimizing
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=19449

[2] commit 3c8ba0d61d04 ("kernel.h: Retain constant expression output for max()/min()")

Signed-off-by: Vincent Mailhol <mailhol.vincent@...adoo.fr>
---
Bonuses:

  - above examples (and a bit more) in godbolt:

      https://godbolt.org/z/GYeEK5d7s

  - that proof, in godbolt, that statically_true() is bettera at
    constant folding than _statically_true()

      https://godbolt.org/z/vK6KK4hMG
---
 include/linux/compiler.h | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index 4d4e23b6e3e7..c76db8b50202 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -308,6 +308,20 @@ static inline void *offset_to_ptr(const int *off)
  */
 #define statically_true(x) (__builtin_constant_p(x) && (x))
 
+/*
+ * Similar to statically_true() but produces a constant expression
+ *
+ * To be used in conjunction with macros, such as BUILD_BUG_ON_ZERO(),
+ * which require their input to be a constant expression and for which
+ * statically_true() would otherwise fail.
+ *
+ * This is a tradeoff: _statically_true() is less efficient at
+ * constant folding and will fail to optimize any expressions in which
+ * at least one of the subcomponents is not constant. For the general
+ * case, statically_true() is better.
+ */
+#define _statically_true(x) __builtin_choose_expr(__is_constexpr(x), x, false)
+
 /*
  * This is needed in functions which generate the stack canary, see
  * arch/x86/kernel/smpboot.c::start_secondary() for an example.
-- 
2.45.2

Powered by blists - more mailing lists