Exclude alignment for MaskBuffer for i686-win7-windows-msvc

Windows 7 does not respect 16 byte alignment for thread locals on i686
builds. Rust 1.79 changed i686 windows builds to use native thread local
support. As a result, using `matrixmultiply` on i686 win7 builds leads
to a UB check panic nounwind when run on Windows 7. See
<rust-lang/rust#138903> for more info.

This change adds `i686-win7-windows-msvc` as an excluded target for the
alignment attribute on `MaskBuffer`.

Author: drewkett

Diff for: src/gemm.rs

@@ -345,7 +345,19 @@ const MASK_BUF_SIZE: usize = KERNEL_MAX_SIZE + KERNEL_MAX_ALIGN - 1;
 // we don't get aligned allocations out of TLS - 16- and 8-byte
 // allocations have been seen, make the minimal align request we can.
 // Align(32) would not work with TLS for s390x.
-#[cfg_attr(not(target_os = "macos"), repr(align(16)))]
+#[cfg_attr(
+    not(any(
+        target_os = "macos",
+        // Target i686-win7-windows-msvc <https://github.com/rust-lang/rust/issues/138903>
+        all(
+            target_arch = "x86",
+            target_vendor = "win7",
+            target_os = "windows",
+            target_env = "msvc"
+        )
+    )),
+    repr(align(16))
+)]
 struct MaskBuffer {
     buffer: [u8; MASK_BUF_SIZE],
 }