Assembler error on larger than 32-bit bitfields

[GabrielRavier]

struct S { unsigned long long int b:32; } s;

void f()
{
    s.b = 0;
}

Trying to compile this results in this error:

/tmp/chibicc-0G8AuI: Assembler messages:
/tmp/chibicc-0G8AuI:60: Error: operand type mismatch for `and'

[tavianator]

codegen.c is attempting to and with an immediate that is too large when writing out the bitfield. This patch fixes it for me:

diff --git a/codegen.c b/codegen.c
index da11fd7..c3164aa 100644
--- a/codegen.c
+++ b/codegen.c
@@ -785,7 +785,12 @@ static void gen_expr(Node *node) {
       // from memory and merge it with a new value.
       Member *mem = node->lhs->member;
       println("  mov %%rax, %%rdi");
-      println("  and $%ld, %%rdi", (1L << mem->bit_width) - 1);
+      if (mem->bit_width >= 32) {
+        println("  mov $%ld, %%rax", (1L << mem->bit_width) - 1);
+        println("  and %%rax, %%rdi");
+      } else {
+        println("  and $%ld, %%rdi", (1L << mem->bit_width) - 1);
+      }
       println("  shl $%d, %%rdi", mem->bit_offset);

       println("  mov (%%rsp), %%rax");

[rui314]

@tavianator Yes, it looks like a bug in codegen. Your code should work, but there's another way. You can shift to the left to clear higher bits and then shift to the right to align bits at a proper bit position. That's I believe what gcc and clang do too: https://godbolt.org/z/G3678rrKz

[tavianator]

@rui314 This is for assignments though, not reads. I think you already do the left+right shift for reads. You could do it to mask the old value but gcc and clang seem to just mask it explicitly: https://godbolt.org/z/6EseE3vo6

[rui314]

@tavianator Ah, interesting! So gcc and clang use an extra register instead of shifting to the left and to the right. I'm not very familiar with instruction selection, but I guess there must be a reason to choose that instruction sequence. Thank you for letting me know!

[tavianator]

@rui314 You're welcome! Actually though I misread that asm, the mask is applied to the value to be stored, not to the old value from the read/modify/write. Other cases don't use masking at all: https://godbolt.org/z/PxMvMqa4h

[rui314]

@tavianator It looks like it is an optimization for a field that is at the leftmost bit position in a word. For other cases, and is still being used: https://godbolt.org/z/hThfor7o3