Bitfield structs that could fit in a single register are still returned by reference

[pizzarollsroyce]

Bitfield structs that have fewer than 32 bits of data should be returned in r10 as they should, but seem to only do that in some cases.

Test code:

#include <stdint.h>
struct TwoBits {
  uint16_t a: 1;
  uint16_t b: 1;
};
struct ThreeBits {
  uint16_t a: 1;
  uint16_t b: 1;
  uint16_t c: 1;
};

extern struct TwoBits MakeTwoBits() {
  struct TwoBits ret = {};
  ret.a = 1;
  return ret;
}

extern struct ThreeBits MakeThreeBits() {
  struct ThreeBits ret = {};
  ret.a = 1;
  return ret;
}

Compiled with clang -O2 -c struct_return.c -o struct_return.clang.o

Result:

00000000 <MakeTwoBits>:
       0: 41 41        	mov 0x1, r10
       2: 1f 18        	jmp [r31]

00000004 <MakeThreeBits>:
       4: e1 40        	mov 0x1, r7
       6: e6 d4 00 00  	st.h r7, 0x0[r6]
       a: 1f 18        	jmp [r31]

Note that the 3-field struct is now returned via a write to the pointer provided in r6.

Conversely, with GCC v810 (v810-gcc -O2 -c struct_return.c -o struct_return.o)

00000000 <_MakeTwoBits>:
   0:	41 41       	mov	1, r10
   2:	1f 18       	jmp	lp

00000004 <_MakeThreeBits>:
   4:	41 41       	mov	1, r10
   6:	1f 18       	jmp	lp

Changing the uint16_ts to uint8_ts doesn't increase this from 2 (so, it's not like it's erroneously adding up the entire bit widths of the primitive type or something).