Defining FMT_BUILTIN_TYPES=0 causes wrong prints

[grisharav]

The following code breaks when defining DFMT_BUILTIN_TYPES=0:

    int32_t x = 56;
    fmt::println("This is x={}", x);

Works fine without the define, but when defined outputs:

This is x=1073479024

I'm working on an embedded system and would like to reduce the size of the generated binary.

[dinomight]

Likely need more details like what compiler you're using, what architecture, which version of {fmt}. Godbolt seems alright using GCC14.2 on x86 [0]. There's also testing in the {fmt} code base for this case that is also passing against various compilers.

fmt/test/no-builtin-types-test.cc

Lines 11 to 16 in 864bdf9

	#define FMT_BUILTIN_TYPES 0
	#include "fmt/format.h"
	TEST(no_builtin_types_test, format) {
	EXPECT_EQ(fmt::format("{}", 42), "42");
	}

[0] https://godbolt.org/z/Y9s8bd337

[vitaut]

Closing as not reproducible but feel free to reopen with more details (ideally a godbolt repro).

[grisharav]

Please reopen, I don't have permissions. We use xtensa gcc 14.2, https://github.com/fmtlib/fmt#11.1.1 . See https://godbolt.org/z/zvEbjref1 for a repro, can't run in godbolt but you can examine the long assembly...

[grisharav]

Please note we use some additional flags of FMT as well, see https://godbolt.org/z/rebhj5MYb

#define FMT_BUILTIN_TYPES 0
#define FMT_USE_LOCALE 0
#define FMT_OPTIMIZE_SIZE 1
#include <fmt/format.h>

int main()
{
    int32_t x = 53;
    fmt::println("This is x={}", x);
}

[grisharav]

@vitaut verified the same issue exists with fmt with the latest 11.1.4
Also verified that the only thing that causes this is FMT_BUILTIN_TYPES, the other flags don't affect this.

[dinomight]

I went ahead and setup a diff view [0]. I don't know the arch and I also don't have a way to play with this, but I do find it fascinating the difference in what the compiler is generating. This might be a compiler bug since x86-64 is fine. Is {fmt} itself built with the same compiler definitions? It looks like there may be some assumptions just looking at the code-gen difference.

[0] https://godbolt.org/z/1sv1Wa4xz

[grisharav]

I went ahead and setup a diff view [0]. I don't know the arch and I also don't have a way to play with this, but I do find it fascinating the difference in what the compiler is generating. This might be a compiler bug since x86-64 is fine. Is {fmt} itself built with the same compiler definitions? It looks like there may be some assumptions just looking at the code-gen difference.

[0] https://godbolt.org/z/1sv1Wa4xz

Yes, FMT is built as part of the build process. We use https://platformio.org/ and set up:

lib_deps =
    fmtlib=https://github.com/fmtlib/fmt#11.1.1

The architecture is ESP32, https://dl.espressif.com/github_assets/espressif/xtensa-isa-doc/releases/download/latest/Xtensa.pdf it's quite a common microprocessor.

[grisharav]

@dinomight @vitaut - can either one reopen, even if it's just for the record?

[vitaut]

Let's keep it closed unless it's reproducible on some compiler we can actually run. You'll need to either not enable FMT_BUILTIN_TYPES or do debugging on your side to see if it's actually a bug in {fmt} or a compiler.

[vitaut]

Possibly resolved as part of #4394.

[grisharav]

Possibly resolved as part of #4394.

Thanks! Will gladly test it when released

[grisharav]

Tested on 11.2, works as expected, thanks!