assertion failure: inlinable function call in a function with debug info must have a !dbg location

[yuanfang-chen]

This code when compiled with clang -g -O1 -fsanitize=kernel-address -c x.c

void m();
void *memcpy(void *a, void *b, unsigned long c) {return 0;}
typedef struct {
  int a;
  int c;
} d;
typedef struct {
  d f;
  int g;
} h;
typedef struct {
  d f;
  h g;
} i;
void j() { memcpy(0, 0, 0); }
struct i *e;
int k(i l) {
  i b;
  b.g = l.g;
  e = &b;
  m();
  return 0;
}
i o;
int n() {
  k(o);
  return 0;
}

produces

inlinable function call in a function with debug info must have a !dbg location
  %18 = call ptr @memcpy(ptr %4, ptr @o, i64 20)
inlinable function call in a function with debug info must have a !dbg location
  %18 = call ptr @memcpy(ptr %4, ptr @o, i64 20)
fatal error: error in backend: Broken module found, compilation aborted!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace, preprocessed source, and associated run script.

I traced it down to HandleByValArgumentInit where newly-inserted memcpy has no debug loc. I think we may attribute the memcpy to some debug loc. But the general problem is that the compiler could synthesize intrinsic calls without debug loc. Maybe we should suppress the assertion for intrinsic calls?

@adrian-prantl

[nikic]

This assertion is already suppressed for intrinsics. However, you're not using an intrinsic here (memcpy vs llvm.memcpy).

[yuanfang-chen]

The llvm.memcpy is inserted by HandleByValArgumentInit, not from the user code.

[yuanfang-chen]

Hi @adrian-prantl, what's your opinion on this?

[adrian-prantl]

The verifier error is there because LLVM will crash in the backend if that function ever becomes inlined. If there is no location from the sourrounding instructions we could use then the correct solution is to use "line 0" (which is reserved for compiler-generated code).