[llvm] Fix emulated TLS alignment for large variables (PR #171037)

[Medha Tiwari via llvm-commits]

================
@@ -150,6 +150,17 @@ bool addEmuTlsVar(Module &M, const GlobalVariable *GV) {
   Type *GVType = GV->getValueType();
   Align GVAlignment = DL.getValueOrABITypeAlignment(GV->getAlign(), GVType);
 
+  // For larger TLS variables that might be accessed via memcpy/memmove,
+  // use the preferred alignment to ensure the backend can use optimized
+  // vector instructions (e.g., ARM NEON vld1.64 requires 16-byte alignment).
+  // This fixes the mismatch where emulated TLS allocates with lower alignment
+  // than what the code generator assumes.
+  TypeSize TS = DL.getTypeStoreSize(GVType);
+  if (TS.isScalable() || TS.getFixedValue() >= 32) {
----------------
medhatiwari wrote:

Thanks for the review! I've addressed all three points:

1. **Scalable size check removed** - You're right, global variables can't have scalable types. Removed that check.

2. The original 32 was arbitrary. I've updated it to `> 16` which corresponds to the NEON vector register size (128 bits = 16 bytes). Types larger than a single vector register are candidates for vectorized memcpy/memmove, which may use aligned vector load/store instructions.

3. **Added testcase** - Added `llvm/test/CodeGen/ARM/emutls-alignment.ll` which verifies a 64-byte TLS array gets 16-byte alignment.

https://github.com/llvm/llvm-project/pull/171037