Question SSE floating kernel

NTDEV
1

I created a new thread using PsCreateSystemThread to avoid the complications of performing floating-point operations in the main kernel context.
Managed the floating-point environment manually, saving and restoring state using MASM instructions, like this:

stmxcsr dword ptr [rsp - 54h]  
ldmxcsr dword ptr gs:[180h]   

movaps [rsp - 10h], xmm0
movaps [rsp + 0h], xmm1
movaps [rsp + 10h], xmm2

The thread executes floating-point operations.

The Problem:
Despite the setup, the results are inconsistent or erroneous. It seems like the floating-point state isn’t being managed properly, or there’s interference with other parts of the kernel.

Is there a specific way to correctly set up the floating-point environment in a kernel thread?
Are there pitfalls I need to watch out for when using PsCreateSystemThread for this purpose?
Should I be using MMX/SSE or other registers differently to ensure correctness?

2

You should be using SSE. The issue is not interfering with other routines, the issue is that some of the floating point state is not saved/restored at interrupt time, so other threads (user or kernel) can interfere with YOUR results.

1 Like
5

i fix it and works now :slight_smile: for hours using LLVM its dirty fix but does the job

__m128 xmm0_backup, xmm1_backup, xmm2_backup;

 __declspec(align(16)) __m128 xmm_backup_stack[3];


    _asm {
        push r8
        push r9
        push r10
    }

    _asm {
        movaps xmm0_backup, xmm0
        movaps xmm1_backup, xmm1
        movaps xmm2_backup, xmm2
    }