diff --git a/Test/TestFBProject/App/main.vp b/Test/TestFBProject/App/main.vp
index e107a17..0ca823f 100644
--- a/Test/TestFBProject/App/main.vp
+++ b/Test/TestFBProject/App/main.vp
@@ -138,7 +138,8 @@ def main() -> t.CInt | t.CExport:
     testcheck.section("Test LLVMIR: inline LLVM IR add")
     x: t.CInt = 10
     y: t.CInt = 20
-    r: t.CInt = c.LLVMIR(f"add i32 {c.LInp(x)}, {c.LInp(y)}", t.CInt)
+    r: t.CInt
+    c.LLVMIR(f"{c.LOut(r)} = add i32 {c.LInp(x)}, {c.LInp(y)}")
     print("LLVMIR add result:", r)
     testcheck.check(r == 30, "LLVMIR add (10+20=30)", "LLVMIR add result mismatch")
 
@@ -152,7 +153,7 @@ def main() -> t.CInt | t.CExport:
 
     testcheck.section("Test LLVMIR with LOut")
     out_val: t.CInt = 0
-    c.LLVMIR(f"{c.LOut(out_val)} = add i32 {c.LInp(x)}, {c.LInp(y)}", t.CInt)
+    c.LLVMIR(f"{c.LOut(out_val)} = add i32 {c.LInp(x)}, {c.LInp(y)}")
     print("LLVMIR LOut result:", out_val)
     testcheck.check(out_val == 30, "LLVMIR LOut (10+20=30)", "LLVMIR LOut result mismatch")
 
diff --git a/includes/vipersimd.vp b/includes/vipersimd.vp
index 53b0f22..e567495 100644
--- a/includes/vipersimd.vp
+++ b/includes/vipersimd.vp
@@ -378,42 +378,58 @@ vzeroupper""",
 
 def simd_sqrt(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: sqrt(x)"""
-    return c.LLVMIR(f"call double @llvm.sqrt.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.sqrt.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_fabs(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: |x|"""
-    return c.LLVMIR(f"call double @llvm.fabs.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.fabs.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_floor(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: floor(x)"""
-    return c.LLVMIR(f"call double @llvm.floor.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.floor.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_ceil(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: ceil(x)"""
-    return c.LLVMIR(f"call double @llvm.ceil.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.ceil.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_round(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: round(x)"""
-    return c.LLVMIR(f"call double @llvm.round.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.round.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_trunc(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: trunc(x)"""
-    return c.LLVMIR(f"call double @llvm.trunc.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.trunc.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_fma(a: t.CDouble, b: t.CDouble, c_val: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: a*b + c (single rounding)"""
-    return c.LLVMIR(f"call double @llvm.fma.f64(double {c.LInp(a)}, double {c.LInp(b)}, double {c.LInp(c_val)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.fma.f64(double {c.LInp(a)}, double {c.LInp(b)}, double {c.LInp(c_val)})")
+    return out
 
 
 def simd_copysign(mag: t.CDouble, sign: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: copysign(mag, sign)"""
-    return c.LLVMIR(f"call double @llvm.copysign.f64(double {c.LInp(mag)}, double {c.LInp(sign)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.copysign.f64(double {c.LInp(mag)}, double {c.LInp(sign)})")
+    return out
 
 
 def simd_neg(x: t.CDouble) -> t.CDouble:
@@ -422,39 +438,55 @@ def simd_neg(x: t.CDouble) -> t.CDouble:
     copysign(x, -1.0) is wrong for x < 0 (returns -|x| instead of +|x|).
     copysign(x, -x) = |x| * sign(-x) = -x is correct for all IEEE 754 values.
     """
-    return c.LLVMIR(f"call double @llvm.copysign.f64(double {c.LInp(x)}, double {c.LInp(t.CDouble(0.0) - x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.copysign.f64(double {c.LInp(x)}, double {c.LInp(t.CDouble(0.0) - x)})")
+    return out
 
 
 def simd_minnum(a: t.CDouble, b: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: min(a, b) (NaN-aware)"""
-    return c.LLVMIR(f"call double @llvm.minnum.f64(double {c.LInp(a)}, double {c.LInp(b)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.minnum.f64(double {c.LInp(a)}, double {c.LInp(b)})")
+    return out
 
 
 def simd_maxnum(a: t.CDouble, b: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: max(a, b) (NaN-aware)"""
-    return c.LLVMIR(f"call double @llvm.maxnum.f64(double {c.LInp(a)}, double {c.LInp(b)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.maxnum.f64(double {c.LInp(a)}, double {c.LInp(b)})")
+    return out
 
 
 def simd_exp(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: e^x"""
-    return c.LLVMIR(f"call double @llvm.exp.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.exp.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_log(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: ln(x)"""
-    return c.LLVMIR(f"call double @llvm.log.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.log.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_sin(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: sin(x)"""
-    return c.LLVMIR(f"call double @llvm.sin.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.sin.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_cos(x: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: cos(x)"""
-    return c.LLVMIR(f"call double @llvm.cos.f64(double {c.LInp(x)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.cos.f64(double {c.LInp(x)})")
+    return out
 
 
 def simd_pow(x: t.CDouble, y: t.CDouble) -> t.CDouble:
     """LLVM intrinsic: x^y"""
-    return c.LLVMIR(f"call double @llvm.pow.f64(double {c.LInp(x)}, double {c.LInp(y)})", t.CDouble)
+    out: t.CDouble
+    c.LLVMIR(f"{c.LOut(out)} = call double @llvm.pow.f64(double {c.LInp(x)}, double {c.LInp(y)})")
+    return out
diff --git a/lib/includes/c.py b/lib/includes/c.py
index 5881be5..c24a3db 100644
--- a/lib/includes/c.py
+++ b/lib/includes/c.py
@@ -615,13 +615,14 @@ class CUndef:
 
 
 class LLVMIR:
-    """内联 LLVM IR
-    
-    用法: c.LLVMIR(f"add i32 {c.LInp(a)}, {c.LInp(b)}", t.CInt)
-    或:   c.LLVMIR(f"%{c.LOut(result)} = add i32 {c.LInp(a)}, {c.LInp(b)}", t.CInt)
-    
-    c.LInp(expr) - 输入操作数，翻译后替换为对应的 LLVM 临时变量名
-    c.LOut(expr) - 输出操作数，翻译后替换为对应的 LLVM 临时变量名
+    """内联 LLVM IR（无返回值，直接嵌入原始 IR 文本）
+
+    用法: c.LLVMIR(f"add i32 {c.LInp(a)}, {c.LInp(b)}")
+    或:   c.LLVMIR(f"{c.LOut(result)} = add i32 {c.LInp(a)}, {c.LInp(b)}")
+
+    c.LInp(expr) - 输入操作数，替换为 expr 的 LLVM 值引用文本（%x 或常量）
+    c.LOut(expr) - 输出操作数：为 expr 变量分配唯一 SSA 寄存器 %llvmirN，
+                   模板嵌入后自动追加 store 写回变量 alloca
     """
     def __init__(self, ir_template: Any, ret_type: Any = None) -> None:
         self.ir_template: Any = ir_template
diff --git a/wiki/08-c-operations.md b/wiki/08-c-operations.md
index 253a74c..2c51ac0 100644
--- a/wiki/08-c-operations.md
+++ b/wiki/08-c-operations.md
@@ -284,10 +284,10 @@ errno: t.CExtern | t.CInt        # extern int errno;
 
 ### c.LLVMIR
 
-直接嵌入 LLVM IR 指令来实现高效且跨平台的汇编操作：
+直接嵌入原始 LLVM IR 文本来实现高效且跨平台的汇编操作（无返回值、不做指令解析）：
 
 ```python
-c.LLVMIR(f"add i32 {c.LInp(a)}, {c.LInp(b)}", t.CInt)
+c.LLVMIR(f"{c.LOut(r)} = add i32 {c.LInp(a)}, {c.LInp(b)}")
 ```
 
 ### c.LInp / c.LOut
@@ -295,8 +295,9 @@ c.LLVMIR(f"add i32 {c.LInp(a)}, {c.LInp(b)}", t.CInt)
 标记 LLVM IR 的输入/输出操作数：
 
 ```python
-c.LInp(expr)    # 输入操作数
-c.LOut(expr)    # 输出操作数
+c.LInp(expr)    # 输入操作数：替换为 expr 的 LLVM 值引用文本（%x 或常量文本）
+c.LOut(expr)    # 输出操作数：为 expr 变量分配唯一 SSA 寄存器 %llvmirN，
+                # 模板嵌入后自动追加 store 写回变量 alloca
 ```
 
 ## 运算符重载