Here's a possible workaround: Replace the "0:" label and the "0b" backreferences with "999:" and "999b", respectively. (Any number greater than 9 should work just as well.)
When I did that, the disassembly suddenly started showing two distinct pairs of ldi statements (with different offsets) where the original (using the "0:" label) hadn't. The idea of giving this change a try was inspired by the following statement from the relevant GNU Assembler docs:
It is also worth noting that the first 10 local labels (0:...9:) are implemented in a slightly more efficient manner than the others.
It seems that "slightly more efficient" also implies "slightly buggy" in our particular case here.
Here's the full repro case:
void setup() {}
bool flag = false;
volatile PGM_P result;
void loop()
{
flag = !flag;
result =
flag
? (__extension__({
PGM_P ptr;
asm volatile
(
".pushsection .progmem.data, \"SM\", @progbits, 1" "\n\t"
"999: .string " "\"false\"" "\n\t"
".popsection" "\n\t"
);
asm volatile
(
"ldi %A0, lo8(999b)" "\n\t"
"ldi %B0, hi8(999b)" "\n\t"
: "=d" (ptr)
);
ptr;
}))
: (__extension__({
PGM_P ptr;
asm volatile
(
".pushsection .progmem.data, \"SM\", @progbits, 1" "\n\t"
"999: .string " "\"true\"" "\n\t"
".popsection" "\n\t"
);
asm volatile
(
"ldi %A0, lo8(999b)" "\n\t"
"ldi %B0, hi8(999b)" "\n\t"
: "=d" (ptr)
);
ptr;
}));
}
Here's the corresponding disassembly of the loop() method -- note the two distinct ldi pairs at offsets ae and b4 that refer to string offsets 104 and 110, respectively:
0000009e <loop>:
9e: 80 91 02 01 lds r24, 0x0102
a2: 91 e0 ldi r25, 0x01 ; 1
a4: 89 27 eor r24, r25
a6: 80 93 02 01 sts 0x0102, r24
aa: 88 23 and r24, r24
ac: 19 f0 breq .+6 ; 0xb4 <loop+0x16>
ae: 88 e6 ldi r24, 0x68 ; 104
b0: 90 e0 ldi r25, 0x00 ; 0
b2: 02 c0 rjmp .+4 ; 0xb8 <loop+0x1a>
b4: 8e e6 ldi r24, 0x6E ; 110
b6: 90 e0 ldi r25, 0x00 ; 0
b8: 90 93 01 01 sts 0x0101, r25
bc: 80 93 00 01 sts 0x0100, r24
c0: 08 95 ret
For comparison, the disassembly when "0:" is used as the local label instead of "999:" -- note that there's only a single ldi pair at offset b0, as you had shown already:
0000009e <loop>:
9e: 80 91 02 01 lds r24, 0x0102
a2: 91 e0 ldi r25, 0x01 ; 1
a4: 89 27 eor r24, r25
a6: 80 93 02 01 sts 0x0102, r24
aa: 88 23 and r24, r24
ac: 09 f0 breq .+2 ; 0xb0 <loop+0x12>
ae: 00 c0 rjmp .+0 ; 0xb0 <loop+0x12>
b0: 8e e6 ldi r24, 0x6E ; 110
b2: 90 e0 ldi r25, 0x00 ; 0
b4: 90 93 01 01 sts 0x0101, r25
b8: 80 93 00 01 sts 0x0100, r24
bc: 08 95 ret