Why does floating-point arithmetic produce different results between x86 and x64 architectures?

Discrepancies in Floating-Point Arithmetic: x86 vs. x64

In a code snippet involving floating-point arithmetic, inconsistencies arise between MS VS 2010 builds targeting x86 and x64 architectures. The code is as follows:

float a = 50.0f;
float b = 65.0f;
float c =  1.3f;
float d = a*c;
bool bLarger1 = d <p><strong>Discrepancies:</strong></p>

• x86 Build: Variable bLarger1 is false (both d and b are set to 65.0), while bLarger2 is true.
• x64 Build: Both bLarger1 and bLarger2 are false.

Underlying Issue:

The discrepancy stems from the expression bool bLarger2 = (a*c)

Difference in Floating-Point Units:

The key difference lies in the floating-point units employed by the two architectures. The x86 architecture uses the x87 floating-point unit, which performs calculations at a higher precision than single-precision (typically double-precision). In contrast, the x64 architecture uses the SSE floating-point unit, which performs pure single-precision calculations.

Impact on the Multiplication:

In the bLarger1 expression, the multiplication of a and c is performed by the hardware multiply instruction. This instruction uses double-precision precision, resulting in d being set to 65.0.

However, in the bLarger2 expression, the multiplication is explicitly performed in single-precision due to the type conversion (a*c). This results in (a*c) being set to 64.999992.

x87 Precision Control:

By default, the x87 unit operates at double-precision. However, it is possible to persuade the unit to perform single-precision calculations using the _controlfp function.

_controlfp(_PC_24, _MCW_PC);

By adding this line to the 32-bit code, both bLarger1 and bLarger2 will be set to false.

Compiler Options:

In more recent versions of Visual Studio, the compiler may emit SSE instructions even for 32-bit targets. This ensures consistency in floating-point arithmetic across different architectures.

The above is the detailed content of Why does floating-point arithmetic produce different results between x86 and x64 architectures?. For more information, please follow other related articles on the PHP Chinese website!