Fix alignment in s390x and cross test

Requested 32-alignment for s390x but thread local storage does not
supply it. Lower requested align to 16 in general to avoid having this
problem pop up on other platforms too.

.github/workflows/ci.yml

@@ -10,6 +10,7 @@ env:
   CARGO_TERM_COLOR: always
   CARGO_INCREMENTAL: 0
   MATMUL_NUM_THREADS: 4
+  RUST_BACKTRACE: full
 
 jobs:
   tests:
@@ -119,6 +120,9 @@ jobs:
     strategy:
       matrix:
         include:
+          - rust: stable
+            target: s390x-unknown-linux-gnu
+            features: constconf cgemm threading
           - rust: stable
             target: aarch64-unknown-linux-gnu
             features: constconf cgemm threading
@@ -146,7 +150,11 @@ jobs:
         run: cross test --target "${{ matrix.target }}" --features "${{ matrix.features }}"
         env:
           MMTEST_FAST_TEST: 1
-          RUSTFLAGS: -Copt-level=2
+      - name: Tests (Release)
+        run: cross test --release --target "${{ matrix.target }}" --features "${{ matrix.features }}"
+        env:
+          MMTEST_FAST_TEST: 1
+
 
   cargo-careful:
     runs-on: ubuntu-latest

src/gemm.rs

@@ -344,7 +344,8 @@ const MASK_BUF_SIZE: usize = KERNEL_MAX_SIZE + KERNEL_MAX_ALIGN - 1;
 // bugs we have seen on certain platforms (macos) that look like
 // we don't get aligned allocations out of TLS - 16- and 8-byte
 // allocations have been seen, make the minimal align request we can.
-#[cfg_attr(not(target_os = "macos"), repr(align(32)))]
+// Align(32) would not work with TLS for s390x.
+#[cfg_attr(not(target_os = "macos"), repr(align(16)))]
 struct MaskBuffer {
     buffer: [u8; MASK_BUF_SIZE],
 }