Commit `4e357151a5`

William Sengir <william@sengir.com>

2022-03-20 08:00:13

stage2: align store for vector-to-array bitcast in LLVM backend

This was causing a very rare segfault when LLVM would emit `vmovdqa` using an unaligned memory operand on the stack.

master

1 parent 961248c

Changed files (1)

src

codegen

llvm.zig

@@ -5563,7 +5563,8 @@ pub const FuncGen = struct {
             if (bitcast_ok) {
                 const llvm_vector_ty = try self.dg.llvmType(operand_ty);
                 const casted_ptr = self.builder.buildBitCast(array_ptr, llvm_vector_ty.pointerType(0), "");
-                _ = self.builder.buildStore(operand, casted_ptr);
+                const llvm_store = self.builder.buildStore(operand, casted_ptr);
+                llvm_store.setAlignment(inst_ty.abiAlignment(target));
             } else {
                 // If the ABI size of the element type is not evenly divisible by size in bits;
                 // a simple bitcast will not work, and we fall back to extractelement.

Commit 4e357151a5

Commit `4e357151a5`