LoongArch build failure after LLVM commit 1a2e77cf9e11dbf56b5720c607313a566eebb16e

[nathanchance]

After llvm/llvm-project@1a2e77c, the LoongArch allyesconfig build fails when building kernel/sched/build_policy.c:

$ echo "CONFIG_CRASH_DUMP=n
CONFIG_KCOV=n
CONFIG_MODULES=n
CONFIG_RELOCATABLE=n
CONFIG_WERROR=n" >allyes.config

$ make -skj"$(nproc)" ARCH=loongarch KCONFIG_ALLCONFIG=1 LLVM=1 mrproper allyesconfig kernel/sched/build_policy.o
In file included from kernel/sched/build_policy.c:17:
In file included from include/linux/sched/cputime.h:5:
In file included from include/linux/sched/signal.h:5:
In file included from include/linux/rculist.h:11:
In file included from include/linux/rcupdate.h:26:
In file included from include/linux/irqflags.h:18:
arch/loongarch/include/asm/percpu.h:97:3: error: call to '__compiletime_assert_51' declared with 'error' attribute: BUILD_BUG failed
   97 |                 BUILD_BUG();
      |                 ^
include/linux/build_bug.h:59:21: note: expanded from macro 'BUILD_BUG'
   59 | #define BUILD_BUG() BUILD_BUG_ON_MSG(1, "BUILD_BUG failed")
      |                     ^
include/linux/build_bug.h:39:37: note: expanded from macro 'BUILD_BUG_ON_MSG'
   39 | #define BUILD_BUG_ON_MSG(cond, msg) compiletime_assert(!(cond), msg)
      |                                     ^
include/linux/compiler_types.h:425:2: note: expanded from macro 'compiletime_assert'
  425 |         _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
      |         ^
include/linux/compiler_types.h:413:2: note: expanded from macro '_compiletime_assert'
  413 |         __compiletime_assert(condition, msg, prefix, suffix)
      |         ^
include/linux/compiler_types.h:406:4: note: expanded from macro '__compiletime_assert'
  406 |                         prefix ## suffix();                             \
      |                         ^
<scratch space>:86:1: note: expanded from here
   86 | __compiletime_assert_51
      | ^
1 error generated.
...

This does not appear to be LoongArch specific though, as the reduced reproducer from cvise triggers for arm64 too.

int lockdep_recursion, local_cpu_mask;
static void __percpu_read(int size) {
  switch (size) {
  case 4:
  case 8:
    break;
  default: {
    void __compiletime_assert_51()
        __attribute__((__error__("BUILD_BUG failed")));
    __compiletime_assert_51();
  }
  }
}
__attribute__((__always_inline__)) void ct_cpuidle_enter() {
  switch (lockdep_recursion)
  case 4:
    __percpu_read(lockdep_recursion);
}
void __attribute__((__no_profile_instrument_function__)) cpu_idle_poll() {
  ct_cpuidle_enter();
}
void find_lock_lowest_rq() { __percpu_read(sizeof &local_cpu_mask); }

$ clang --version | head -2
ClangBuiltLinux clang version 18.0.0 (https://github.com/llvm/llvm-project ebc2c4bde3ec0d37d1bcb224e7ffcc18ac4e2620)
Target: aarch64-redhat-linux-gnu

$ clang -O2 -c -o /dev/null build_policy.i

$ clang --target=loongarch64-linux-gnusf -O2 -c -o /dev/null build_policy.i

$ clang --version | head -2
ClangBuiltLinux clang version 18.0.0 (https://github.com/llvm/llvm-project 1a2e77cf9e11dbf56b5720c607313a566eebb16e)
Target: aarch64-redhat-linux-gnu

$ clang -O2 -c -o /dev/null build_policy.i
build_policy.i:10:5: error: call to '__compiletime_assert_51' declared with 'error' attribute: BUILD_BUG failed
   10 |     __compiletime_assert_51();
      |     ^
1 error generated.

$ clang --target=loongarch64-linux-gnusf -O2 -c -o /dev/null build_policy.i
build_policy.i:10:5: error: call to '__compiletime_assert_51' declared with 'error' attribute: BUILD_BUG failed
   10 |     __compiletime_assert_51();
      |     ^
1 error generated.

I have an LLVM IR reproducer that was generated by testing for the presence of the error with opt | llc at the bad revision but not the good revision using llvm-reduce.

target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-redhat-linux-gnu"

; Function Attrs: alwaysinline
define void @ct_cpuidle_enter() #0 {
entry:
  call void @__percpu_read(i32 4)
  ret void
}

define internal void @__percpu_read(i32 %size) #1 {
entry:
  switch i32 %size, label %sw.default [
    i32 4, label %common.ret
    i32 0, label %common.ret
  ]

common.ret:                                       ; preds = %sw.default, %entry, %entry
  ret void

sw.default:                                       ; preds = %entry
  call void @__compiletime_assert_51()
  br label %common.ret
}

define void @cpu_idle_poll() {
entry:
  call void @ct_cpuidle_enter()
  ret void
}

define void @find_lock_lowest_rq() #1 {
entry:
  call void @__percpu_read(i32 0)
  ret void
}

declare void @__compiletime_assert_51(...) #2

attributes #0 = { alwaysinline "target-features"="+fp-armv8,+neon,+outline-atomics,+v8a,-fmv" }
attributes #1 = { "target-features"="+fp-armv8,+neon,+outline-atomics,+v8a,-fmv" }
attributes #2 = { "dontcall-error"="BUILD_BUG failed" }

# Good
$ opt -o - -O2 reduced.ll | llc -o /dev/null -

# Bad
$ opt -o - -O2 reduced.ll | llc -o /dev/null -
error: call to __compiletime_assert_51 marked "dontcall-error": BUILD_BUG failed

[nickdesaulniers]

I'm guessing they're missing an always_inline on their definition of __percpu_read. The semantics of __attribute__((error())) is that if a call to it is emitted, it's an error.

If the caller needs to be inlined to elide calls to that error fn, then the caller probably should have been marked always_inline.

At least the reduced test cases make me think this is a kernel source problem, not an llvm problem. Sources should not assume inline substitution is performed or not, unless they explicitly use noinline/always_inline attributes.

[nathanchance]

Heh, duh... This fixes the issue:

diff --git a/arch/loongarch/include/asm/percpu.h b/arch/loongarch/include/asm/percpu.h
index b9f567e66016..8fb857ae962b 100644
--- a/arch/loongarch/include/asm/percpu.h
+++ b/arch/loongarch/include/asm/percpu.h
@@ -32,7 +32,7 @@ static inline void set_my_cpu_offset(unsigned long off)
 #define __my_cpu_offset __my_cpu_offset

 #define PERCPU_OP(op, asm_op, c_op)                                    \
-static inline unsigned long __percpu_##op(void *ptr,                   \
+static __always_inline unsigned long __percpu_##op(void *ptr,                  \
                        unsigned long val, int size)                    \
 {                                                                      \
        unsigned long ret;                                              \
@@ -63,7 +63,7 @@ PERCPU_OP(and, and, &)
 PERCPU_OP(or, or, |)
 #undef PERCPU_OP

-static inline unsigned long __percpu_read(void *ptr, int size)
+static __always_inline unsigned long __percpu_read(void *ptr, int size)
 {
        unsigned long ret;

I will send this with a commit message shortly.

[nathanchance]

Patch submitted: https://lore.kernel.org/20231101-loongarch-always-inline-percpu-ops-v1-1-b8f2e9a71729@kernel.org/

[nathanchance]

v2: https://lore.kernel.org/20231102-loongarch-always-inline-percpu-ops-v2-1-31c51959a5c0@kernel.org/

[nathanchance]

Patch accepted: https://git.kernel.org/chenhuacai/linux-loongson/c/71945968d8b128c955204baa33ec03bdd91bdc26

[nathanchance]

https://git.kernel.org/linus/71945968d8b128c955204baa33ec03bdd91bdc26