[clang] [SPARC] use `divideCeil` to calculate register offset (PR #213739)

[Eli Friedman via cfe-commits]

================
@@ -54,6 +54,29 @@ long double f_longdouble(long a, struct align16_longdouble b) {
 	return b.x;
 }
 
+// CHECK-LABEL: define{{.*}} signext i32 @f_float_aligned(float noundef %a, i64 %0, i64 %b.coerce0, i64 %b.coerce1)
+int f_float_aligned(float a, struct align16_int b) {
+	return b.x;
+}
+
+// CHECK-LABEL: define{{.*}} signext i32 @f_float_pair_aligned(float noundef %a, float noundef %b, i64 %c.coerce0, i64 %c.coerce1)
+int f_float_pair_aligned(float a, float b, struct align16_int c) {
+	return c.x;
+}
+
+// CHECK-LABEL: define{{.*}} i64 @f_char_int128_aligned(i8 noundef signext %x, i64 %0, i128 noundef %v, i64 %q.coerce0, i64 %q.coerce1)
+struct aligned16_struct { long a, b; } __attribute__((aligned(16)));
+long f_char_int128_aligned(char x, __int128 v, struct aligned16_struct q) {
+    return q.a;
+}
+
+// FIXME: alignment on typedefs should be taken into account, but isn't.
----------------
efriedma-quic wrote:

The gcc-style `aligned` attribute on typedefs is a mess; it isn't part of the canonical type, so it doesn't integrate into the type system properly.  This isn't fundamental to the notion of an alignment qualifier; the Microsoft `__unaligned` is properly integrated.  But it's the current state, and I doubt anyone will try to address it anytime soon.

(Note that this is unrelated to the `aligned` attribute on structs or fields, which works differently.)

https://github.com/llvm/llvm-project/pull/213739