Can you explain what "It works fine in 64 bits, but not in 32 bits" really means? Error? Incorrect result? Give us the details.

你能解释一下“它在64位下运行得很好，但在32位下就不行了”的真正含义吗？错误？结果不正确？给我们讲讲细节。

ia64 refers to Itanium which I suspect is not what you have. The common 64-bit desktop architecture is called x86-64 or amd64 (or x64 by Microsoft). But since you're compiling as 32-bit code, you're not using that either; this is just x86.

Ia64指的是安腾，我怀疑它不是你所拥有的。常见的64位桌面体系结构称为x86-64或AMD64(或Microsoft的x64)。但是，因为您正在编译为32位代码，所以您也没有使用它；这只是x86。

I think the fabs version might be able to promote the value directly to double, since fabs is defined with a double argument and return value, and then you get the nearest double to -189.55 which is much closer. I'd have to double check C's rules for floating point literals. I suspect if you use fabsf instead you will get the same result as the other versions.

我认为FABS版本可能能够直接将值提升到双倍，因为FABS是用双精度参数和返回值定义的，然后您可以得到最接近的双精度，即-189.55，这更接近。我必须仔细检查C语言中浮点文字的规则。我怀疑，如果您使用frupf，您将得到与其他版本相同的结果。

You didn't say what the problem is.

你没说问题出在哪里。

I'm not sure why you mess with the rounding mode. fabs toggles the sign bit, it should not cause rounding.

我不知道你为什么要弄乱四舍五入模式。FABS切换符号位，它应该不会导致舍入。

Every call to some fabs variant in OP’s code passes -189.55f as argument. This should always be a float of the same value, and the fact one of the calls is to fabs and that fabs has a double parameter and double return type should be irrelevant. -189.55f should produce a float value before it is passed to fabs, and passing it to fabs should not change that value. Peter Cordes already identified the problem, a compiler defect.

每次调用OP代码中的某个FABS变量时，都会传递-189.55f作为参数。它应该始终是一个相同值的浮点数，其中一个调用是对FABS的调用，FABS具有双参数和双返回类型，这一事实应该是无关紧要的。-189.55f应在传递给FABS之前产生浮点值，传递给FABS不应改变该值。Peter Cordes已经发现了这个问题，这是一个编译器缺陷。

@EricPostpischil The precision of float is just a minimum -- the compiler is always permitted to evaluate things at higher precision if it wants to, but it is not required to. But that is irrelevant to the OPs question of why the 3 appears (and how to get rid of it), which is due to using float precision.

@EricPostpischil浮点数的精度只是一个最小值--如果编译器愿意，它总是被允许以更高的精度计算，但这不是必需的。但这与OP为什么出现3（以及如何摆脱它）的问题无关，这是由于使用浮点精度。

The C standard explicitly states that all instances of a floating-point literal of the same form (exactly the same characters in the source text) must convert to the same value. C 2018 6.4.4.2 5: “… All floating constants of the same source form shall convert to the same internal format with the same value.”

C标准明确规定，相同形式(源文本中完全相同的字符)的浮点文本的所有实例必须转换为相同的值。C 2018 6.4.4.2 5：“…同一源格式的所有浮点常量应转换为具有相同值的相同内部格式。“

The compiler is permitted to do operations at higher precision - Yes, but 189.55f means the starting point for any operations must still be a float that can actually have existed, as Eric says. That literal has a value of type float. Gaining precision beyond that requires breaking in to the parsing of the float and changing to parsing it as a double, which as Eric says is not what the standard says should happen. godbolt.org/z/n53nsKhW6 shows that as args to printf, (float)189.55f has the expected low zeros in the mantissa but plain 189.55f doesn't.

编译器被允许以更高的精度进行运算--是的，但189.55f意味着任何运算的起点必须仍然是一个浮点数，而该浮点数实际上可能已经存在，Eric说。该字面值的类型为Float。除此之外，要获得更高的精度，需要进入浮点数的解析，并将其解析为双精度型，正如Eric所说，这不是标准所说的应该发生的事情。Org/z/n53nsKhW6显示，与args to print tf一样，(Float)189.55f在尾数中有预期的低零，而普通的189.55f没有。

@PeterCordes: Ugh. Lousy excuse. I expect the standard’s reason for allowing excess precision is for run-time performance—use one FMA instead of two instructions, use double-precision instructions on a processor without single-precision instructions, etc. Interpreting the standard’s latitude on computation as applying to translating constants is questionable. I would expect most experienced floating-point programmers to think that 189.55f refers to a specific value, so the GCC behavior will be surprising to them.

@PeterCordes：啊。糟糕的借口。我预计该标准允许额外精度的原因是为了运行时性能-使用一个FMA而不是两条指令，在没有单精度指令的处理器上使用双精度指令，等等。将该标准在计算方面的纬度解释为适用于转换常量是值得怀疑的。我预计大多数有经验的浮点程序员都会认为189.55f指的是一个特定值，所以GCC的行为会让他们感到惊讶。