Fix `auto_cast` matrix bug

1 files changed, 12 insertions, 1 deletions

diff --git a/src/llvm_backend_expr.cpp b/src/llvm_backend_expr.cpp
index b5f6437a4..8ad44035d 100644
--- a/src/llvm_backend_expr.cpp
+++ b/src/llvm_backend_expr.cpp
@@ -2363,12 +2363,23 @@ gb_internal lbValue lb_emit_conv(lbProcedure *p, lbValue value, Type *t) {
 		GB_ASSERT(src->kind == Type_Matrix);
 		lbAddr v = lb_add_local_generated(p, t, true);
 
-		if (is_matrix_square(dst) && is_matrix_square(dst)) {
+		if (dst->Matrix.row_count    == src->Matrix.row_count &&
+		    dst->Matrix.column_count == src->Matrix.column_count) {
+			for (i64 j = 0; j < dst->Matrix.column_count; j++) {
+				for (i64 i = 0; i < dst->Matrix.row_count; i++) {
+					lbValue d = lb_emit_matrix_epi(p, v.addr, i, j);
+					lbValue s = lb_emit_matrix_ev(p, value, i, j);
+					s = lb_emit_conv(p, s, dst->Matrix.elem);
+					lb_emit_store(p, d, s);
+				}
+			}
+		} else if (is_matrix_square(dst) && is_matrix_square(dst)) {
 			for (i64 j = 0; j < dst->Matrix.column_count; j++) {
 				for (i64 i = 0; i < dst->Matrix.row_count; i++) {
 					if (i < src->Matrix.row_count && j < src->Matrix.column_count) {
 						lbValue d = lb_emit_matrix_epi(p, v.addr, i, j);
 						lbValue s = lb_emit_matrix_ev(p, value, i, j);
+						s = lb_emit_conv(p, s, dst->Matrix.elem);
 						lb_emit_store(p, d, s);
 					} else if (i == j) {
 						lbValue d = lb_emit_matrix_epi(p, v.addr, i, j);