When the Compiler Bites

Update: There are discussions on Reddit and on Hacker News.

So far this year I’ve been bitten three times by compiler edge cases in GCC and Clang, each time catching me totally by surprise. Two were caused by historical artifacts, where an ambiguous specification lead to diverging implementations. The third was a compiler optimization being far more clever than I expected, behaving almost like an artificial intelligence.

In all examples I’ll be using GCC 7.3.0 and Clang 6.0.0 on Linux.

x86-64 ABI ambiguity

The first time I was bit — or, well, narrowly avoided being bit — was when I examined a missed floating point optimization in both Clang and GCC. Consider this function:

double
zero_multiply(double x)
{
    return x * 0.0;
}

The function multiplies its argument by zero and returns the result. Any number multiplied by zero is zero, so this should always return zero, right? Unfortunately, no. IEEE 754 floating point arithmetic supports NaN, infinities, and signed zeros. This function can return NaN, positive zero, or negative zero. (In some cases, the operation could also potentially produce a hardware exception.)

As a result, both GCC and Clang perform the multiply:

zero_multiply:
    xorpd  xmm1, xmm1
    mulsd  xmm0, xmm1
    ret

The -ffast-math option relaxes the C standard floating point rules, permitting an optimization at the cost of conformance and consistency:

zero_multiply:
    xorps  xmm0, xmm0
    ret

Side note: -ffast-math doesn’t necessarily mean “less precise.” Sometimes it will actually improve precision.

Here’s a modified version of the function that’s a little more interesting. I’ve changed the argument to a short:

double
zero_multiply_short(short x)
{
    return x * 0.0;
}

It’s no longer possible for the argument to be one of those special values. The short will be promoted to one of 65,535 possible double values, each of which results in 0.0 when multiplied by 0.0. GCC misses this optimization (-Os):

zero_multiply_short:
    movsx     edi, di       ; sign-extend 16-bit argument
    xorps     xmm1, xmm1    ; xmm1 = 0.0
    cvtsi2sd  xmm0, edi     ; convert int to double
    mulsd     xmm0, xmm1
    ret

Clang also misses this optimization:

zero_multiply_short:
    cvtsi2sd xmm1, edi
    xorpd    xmm0, xmm0
    mulsd    xmm0, xmm1
    ret

But hang on a minute. This is shorter by one instruction. What happened to the sign-extension (movsx)? Clang is treating that short argument as if it were a 32-bit value. Why do GCC and Clang differ? Is GCC doing something unnecessary?

It turns out that the x86-64 ABI didn’t specify what happens with the upper bits in argument registers. Are they garbage? Are they zeroed? GCC takes the conservative position of assuming the upper bits are arbitrary garbage. Clang takes the boldest position of assuming arguments smaller than 32 bits have been promoted to 32 bits by the caller. This is what the ABI specification should have said, but currently it does not.

Fortunately GCC also conservative when passing arguments. It promotes arguments to 32 bits as necessary, so there are no conflicts when linking against Clang-compiled code. However, this is not true for Intel’s ICC compiler: Clang and ICC are not ABI-compatible on x86-64.

I don’t use ICC, so that particular issue wouldn’t bite me, but if I was ever writing assembly routines that called Clang-compiled code, I’d eventually get bit by this.

Floating point precision

Without looking it up or trying it, what does this function return? Think carefully.

int
float_compare(void)
{
    float x = 1.3f;
    return x == 1.3f;
}

Confident in your answer? This is a trick question, because it can return either 0 or 1 depending on the compiler. Boy was I confused when this comparison returned 0 in my real world code.

$ gcc   -std=c99 -m32 cmp.c  # float_compare() == 0
$ clang -std=c99 -m32 cmp.c  # float_compare() == 1

So what’s going on here? The original ANSI C specification wasn’t clear about how intermediate floating point values get rounded, and implementations all did it differently. The C99 specification cleaned this all up and introduced FLT_EVAL_METHOD. Implementations can still differ, but at least you can now determine at compile-time what the compiler would do by inspecting that macro.

Back in the late 1980’s or early 1990’s when the GCC developers were deciding how GCC should implement floating point arithmetic, the trend at the time was to use as much precision as possible. On the x86 this meant using its support for 80-bit extended precision floating point arithmetic. Floating point operations are performed in long double precision and truncated afterward (FLT_EVAL_METHOD == 2).

In float_compare() the left-hand side is truncated to a float by the assignment, but the right-hand side, despite being a float literal, is actually “1.3” at 80 bits of precision as far as GCC is concerned. That’s pretty unintuitive!

The remnants of this high precision trend are still in JavaScript, where all arithmetic is double precision (even if simulated using integers), and great pains have been made to work around the performance consequences of this. Until recently, Mono had similar issues.

The trend reversed once SIMD hardware became widely available and there were huge performance gains to be had. Multiple values could be computed at once, side by side, at lower precision. So on x86-64, this became the default (FLT_EVAL_METHOD == 0). The young Clang compiler wasn’t around until well after this trend reversed, so it behaves differently than the backwards compatible GCC on the old x86.

I’m a little ashamed that I’m only finding out about this now. However, by the time I was competent enough to notice and understand this issue, I was already doing nearly all my programming on the x86-64.

Built-in Function Elimination

I’ve saved this one for last since it’s my favorite. Suppose we have this little function, new_image(), that allocates a greyscale image for, say, some multimedia library.

static unsigned char *
new_image(size_t w, size_t h, int shade)
{
    unsigned char *p = 0;
    if (w == 0 || h <= SIZE_MAX / w) { // overflow?
        p = malloc(w * h);
        if (p) {
            memset(p, shade, w * h);
        }
    }
    return p;
}

It’s a static function because this would be part of some slick header library (and, secretly, because it’s necessary for illustrating the issue). Being a responsible citizen, the function even checks for integer overflow before allocating anything.

I write a unit test to make sure it detects overflow. This function should return 0.

/* expected return == 0 */
int
test_new_image_overflow(void)
{
    void *p = new_image(2, SIZE_MAX, 0);
    return !!p;
}

So far my test passes. Good.

I’d also like to make sure it correctly returns NULL — or, more specifically, that it doesn’t crash — if the allocation fails. But how can I make malloc() fail? As a hack I can pass image dimensions that I know cannot ever practically be allocated. Essentially I want to force a malloc(SIZE_MAX), e.g. allocate every available byte in my virtual address space. For a conventional 64-bit machine, that’s 16 exibytes of memory, and it leaves space for nothing else, including the program itself.

/* expected return == 0 */
int
test_new_image_oom(void)
{
    void *p = new_image(1, SIZE_MAX, 0xff);
    return !!p;
}

I compile with GCC, test passes. I compile with Clang and the test fails. That is, the test somehow managed to allocate 16 exibytes of memory, and initialize it. Wat?

Disassembling the test reveals what’s going on:

test_new_image_overflow:
    xor  eax, eax
    ret

test_new_image_oom:
    mov  eax, 1
    ret

The first test is actually being evaluated at compile time by the compiler. The function being tested was inlined into the unit test itself. This permits the compiler to collapse the whole thing down to a single instruction. The path with malloc() became dead code and was trivially eliminated.

In the second test, Clang correctly determined that the image buffer is not actually being used, despite the memset(), so it eliminated the allocation altogether and then simulated a successful allocation despite it being absurdly large. Allocating memory is not an observable side effect as far as the language specification is concerned, so it’s allowed to do this. My thinking was wrong, and the compiler outsmarted me.

I soon realized I can take this further and trick Clang into performing an invalid optimization, revealing a bug. Consider this slightly-optimized version that uses calloc() when the shade is zero (black). The calloc() function does its own overflow check, so new_image() doesn’t need to do it.

static void *
new_image(size_t w, size_t h, int shade)
{
    unsigned char *p = 0;
    if (shade == 0) { // shortcut
        p = calloc(w, h);
    } else if (w == 0 || h <= SIZE_MAX / w) { // overflow?
        p = malloc(w * h);
        if (p) {
            memset(p, color, w * h);
        }
    }
    return p;
}

With this change, my overflow unit test is now also failing. The situation is even worse than before. The calloc() is being eliminated despite the overflow, and replaced with a simulated success. This time it’s actually a bug in Clang. While failing a unit test is mostly harmless, this could introduce a vulnerability in a real program. The OpenBSD folks are so worried about this sort of thing that they’ve disabled this optimization.

Here’s a slightly-contrived example of this. Imagine a program that maintains a table of unsigned integers, and we want to keep track of how many times the program has accessed each table entry. The “access counter” table is initialized to zero, but the table of values need not be initialized, since they’ll be written before first access (or something like that).

struct table {
    unsigned *counter;
    unsigned *values;
};

static int
table_init(struct table *t, size_t n)
{
    t->counter = calloc(n, sizeof(*t->counter));
    if (t->counter) {
        /* Overflow already tested above */
        t->values = malloc(n * sizeof(*t->values));
        if (!t->values) {
            free(t->counter);
            return 0; // fail
        }
        return 1; // success
    }
    return 0; // fail
}

This function relies on the overflow test in calloc() for the second malloc() allocation. However, this is a static function that’s likely to get inlined, as we saw before. If the program doesn’t actually make use of the counter table, and Clang is able to statically determine this fact, it may eliminate the calloc(). This would also eliminate the overflow test, introducing a vulnerability. If an attacker can control n, then they can overwrite arbitrary memory through that values pointer.

The takeaway

Besides this surprising little bug, the main lesson for me is that I should probably isolate unit tests from the code being tested. The easiest solution is to put them in separate translation units and don’t use link-time optimization (LTO). Allowing tested functions to be inlined into the unit tests is probably a bad idea.

The unit test issues in my real program, which was a bit more sophisticated than what was presented here, gave me artificial intelligence vibes. It’s that situation where a computer algorithm did something really clever and I felt it outsmarted me. It’s creepy to consider how far that can go. I’ve gotten that even from observing AI I’ve written myself, and I know for sure no human taught it some particularly clever trick.

My favorite AI story along these lines is about an AI that learned how to play games on the Nintendo Entertainment System. It didn’t understand the games it was playing. It’s optimization task was simply to choose controller inputs that maximized memory values, because that’s generally associated with doing well — higher scores, more progress, etc. The most unexpected part came when playing Tetris. Eventually the screen would fill up with blocks, and the AI would face the inevitable situation of losing the game, with all that memory being reinitialized to low values. So what did it do?

Just before the end it would pause the game and wait… forever.

Load Comments

null program

Chris Wellons