Errors from the Testsuite GCC 13.2.0

[JPeterMugaas]

I was running the testsuite on GNU GCC 13.2.0 and I got these errors:

***************
*************
howdy, from the PC
        Testing DirectX Math (aka XNAMath 3) 318
        SSE_INTRINSICS
Warning: Couldn't open math2.dat. f==0

[..]
XMVectorPermute: 3 3 6 6 = 0.000000 0.000000 -0.000000 -0.000000 ... 3.000000 3.000000 6.000000 6.00
0000
XMVectorPermute: 3 2 4 2 = 0.000000 0.000000 0.000000 0.000000 ... 3.000000 2.000000 4.000000 2.0000
00
XMVectorPermute: 1 2 7 1 = -14743.487305 0.000000 0.000000 -14743.487305 ... 1.000000 2.000000 7.000
000 1.000000
XMVectorPermute: 2 4 2 2 = 0.000000 0.000000 0.000000 0.000000 ... 2.000000 4.000000 2.000000 2.0000
00
XMVectorPermute: 4 2 4 5 = 0.000000 0.000000 0.000000 0.000000 ... 4.000000 2.000000 4.000000 5.0000
00
XMVectorPermute: 4 4 7 2 = 0.000000 0.000000 0.000000 0.000000 ... 4.000000 4.000000 7.000000 2.0000
00
XMVectorPermute: 4 5 4 4 = 0.000000 0.000000 0.000000 0.000000 ... 4.000000 5.000000 4.000000 4.0000
00
XMVectorPermute: 7 1 3 7 = 0.000000 -14743.487305 0.000000 0.000000 ... 7.000000 1.000000 3.000000 7
.000000
   test 321 (XMVectorPermute) FAILED

The tests run fine with Clang.

[walbourn]

Is this in DEBUG or RELEASE builds?

The DirectXMath library exercises a lot of intrinsics code-gen, so it has definitely found a lot of compiler bugs over the years.

[JPeterMugaas]

Release

[walbourn]

This is pretty clearly a compiler optimization bug. Perhaps you should file it with the GNUC/MinGW folks and link the report here.

[walbourn]

This bug still exists in MinGW 15.3 so GCC really doesn't handle permute well.

[walbourn]

I think the problem is the "address-of and cast" trick being used in XMVectorPermute and XMVectorSwizzle (not the template version) that GCC considers undefined behavior. I've reworked it for GNUC and it seems to work now.

[walbourn]

Kept the same code pattern as before for MSVC and clang