fix(asm): 内联汇编输出约束修复(去重+多输出 struct)

- _asm_add_operand 去重: 模板内嵌 AsmOut/AsmInp 与 out=/inp= 声明同一变量不再重复收集
- ret_ty 多输出: output_count>1 返回匿名 struct, 元素数=输出数
- 输出处理: 多输出 extractvalue 逐元素 store(rdrand*_step 的 val+ok)
- extractvalue SSA 名改字母前缀 %asm.N.e{i} 避免数字 ID 后缀非法

App/lib/core/Gen/CallGenMixin.vp

@@ -1219,9 +1219,20 @@ class CallGenMixin:
                          val: llvmlite.Value | t.CPtr,
                          constraint: str,
                          is_output: int) -> int:
-        """添加操作数到数组，返回新计数（失败返回原计数）"""
+        """添加操作数到数组，返回新计数（失败返回原计数）
+
+        去重：f-string 模板内嵌的 AsmOut/AsmInp 与 out=/inp= 声明指向
+        同一变量时会重复收集（如 vrandom.vp 的 c.Asm(f"...{AsmOut(v)}",
+        out=[AsmOut(v)])），同一 (Value, IsOutput) 只保留第一个，
+        避免约束数多于 asm 返回元素数（llc: output constraints mismatch）。
+        """
         if count >= ASM_MAX_OPERANDS:
             return count
+        if val is not None:
+            for ei in range(count):
+                eop: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, ei)
+                if eop is not None and eop.Value == val and eop.IsOutput == is_output:
+                    return count
         op: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, count)
         op.Value = val
         op.Constraint = constraint
@@ -1481,8 +1492,35 @@ class CallGenMixin:
                 input_count += 1
 
         # 确定返回类型
+        # 多输出（如 vrandom rdrand*_step 的 val+ok 两个 AsmOut）返回
+        # 匿名 struct {ty1, ty2, ...}，元素数 = 输出操作数（LLVM asm
+        # 要求输出约束数 = 返回 struct 元素数）。
         ret_ty: llvmlite.LLVMType | t.CPtr = llvmlite.Void(pool)
         if output_count > 0 and first_output_idx >= 0:
+            if output_count > 1:
+                out_pn_head: llvmlite.ParamNode | t.CPtr = None
+                out_pn_last: llvmlite.ParamNode | t.CPtr = None
+                out_pn_n: int = 0
+                for oi5 in range(operand_count):
+                    op5: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, oi5)
+                    if op5 is None or op5.IsOutput == 0:
+                        continue
+                    if op5.Value is not None and op5.Value.Ty is not None \
+                            and op5.Value.Ty.Pointee is not None:
+                        pn5: llvmlite.ParamNode | t.CPtr = llvmlite.new_param_node(
+                            pool, op5.Value.Ty.Pointee)
+                        if pn5 is not None:
+                            if out_pn_head is None:
+                                out_pn_head = pn5
+                            else:
+                                out_pn_last.Next = pn5
+                            out_pn_last = pn5
+                            out_pn_n += 1
+                if out_pn_head is not None and out_pn_n > 0:
+                    ret_ty = llvmlite.Struct(pool, out_pn_head, out_pn_n, None)
+                else:
+                    ret_ty = llvmlite.Int32(pool)
+            else:
                 out_op: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, first_output_idx)
                 if out_op.Value is not None and out_op.Value.Ty is not None:
                     # alloca 的类型是指针，Pointee 是目标类型
@@ -1676,9 +1714,34 @@ class CallGenMixin:
         # ============================================================
         if is_void == 0 and first_output_idx >= 0:
             result_val: llvmlite.Value | t.CPtr = llvmlite.SSAValue(pool, ret_ty, ssa_name)
-            out_op2: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, first_output_idx)
-            if out_op2.Value is not None:
-                llvmlite.build_store(builder, result_val, out_op2.Value)
+            out_emit_idx: int = 0
+            for oi_out in range(operand_count):
+                orig_o: int = new_order[oi_out]
+                oo: HandlesExprCall.AsmOperand | t.CPtr = self._asm_get_operand(operands, orig_o)
+                if oo is None or oo.IsOutput == 0:
+                    continue
+                if output_count > 1:
+                    # 多输出：extractvalue 取第 out_emit_idx 个元素后 store
+                    if oo.Value is not None and oo.Value.Ty is not None \
+                            and oo.Value.Ty.Pointee is not None:
+                        ev_name: t.CChar | t.CPtr = pool.alloc(32)
+                        ev_line: t.CChar | t.CPtr = pool.alloc(ASM_LINE_BUF_SIZE)
+                        if ev_name is not None and ev_line is not None:
+                            # ssa_name 形如 "%N"（纯数字 ID），构造唯一名
+                            # "%asm.N.e{i}"（字母开头，避免 %N.e{i} 被解析为
+                            # 数字 ID + 非法后缀导致 "expected '=' after id"）
+                            viperlib.snprintf(ev_name, 32, "%%asm.%s.e%d", ssa_name + 1, out_emit_idx)
+                            viperlib.snprintf(ev_line, ASM_LINE_BUF_SIZE,
+                                "%s = extractvalue %s %s, %d",
+                                ev_name, ret_ty_buf, ssa_name, out_emit_idx)
+                            llvmlite.block_append_text(pool, builder.CurBlock, ev_line)
+                            ev_val: llvmlite.Value | t.CPtr = llvmlite.SSAValue(
+                                pool, oo.Value.Ty.Pointee, ev_name)
+                            llvmlite.build_store(builder, ev_val, oo.Value)
+                else:
+                    if oo.Value is not None:
+                        llvmlite.build_store(builder, result_val, oo.Value)
+                out_emit_idx += 1
             return result_val
 
         return None