Issue 243457
Summary [AsmPrinter] `trunc` to a non-byte-sized type in a global initializer is emitted untruncated
Labels llvm:codegen, miscompilation
Assignees
Reporter chfast
    
```llvm
@a = global [2 x i8] zeroinitializer
@b = global [2 x i8] zeroinitializer
@g = global i1 trunc (i64 sub (i64 ptrtoint (ptr @b to i64), i64 ptrtoint (ptr @a to i64)) to i1)

define i32 @load_g() {
  %v = load i1, ptr @g
  %z = zext i1 %v to i32
  ret i32 %z
}
```

`llc -mtriple=x86_64-linux-gnu -O2`:

```asm
load_g:
        movq    g@GOTPCREL(%rip), %rax
 movzbl  (%rax), %eax
        retq
...
g:
        .byte   b-a
```

`b - a` is 2. Truncated to i1 it is 0, but the byte holds 2. Codegen reads an `i1` as a zero-extended byte, so `load_g()` returns 2, which `zext i1` cannot produce. With this driver it prints `g byte = 2, load_g() = 2`:

```c
#include <stdio.h>
extern unsigned char g;
int load_g(void);
int main(void) {
  printf("g byte = %d, load_g() = %d\n", g, load_g());
  return 0;
}
```

`AsmPrinter::lowerConstant` handles `Instruction::Trunc` by emitting the operand _expression_ unchanged. The comment there says it relies on "the assembler to truncate the generated _expression_ properly". The assembler only truncates to the emitted size, one byte here, not to the bit width of the type.

So for a destination type that is not byte-sized, the truncation is lost. One option is to emit the _expression_ masked to the type width, `(b-a)&1`, which the assembler still folds for label differences (llvm-mc gives 0 for `.byte (b-a)&1` here). The other is to reject it.

LangRef has no sentence on the padding bits of global initializers. For stores and loads it says:
- **store:** a non-byte-sized store "behaves like a `zext` followed by a byte-sized store";
- **load:** a load treats a non-zero padding bit as poison.

So the initializer is expected to have zero padding.

This is with trunk e0b437836a21.

_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to