LLVM學習筆記（45）
3.7. 描述呼叫慣例的資料結構
選項“-gen-callingconv”用於生成處理函式呼叫慣例的程式碼。呼叫慣例是函式呼叫者與被呼叫者之間關於引數及返回值傳遞方式的一個共識。存在多個呼叫慣例，以適合各種機器架構。LLVM目前已經基本能完全通過TableGen生成處理呼叫慣例的程式碼。
3.7.1. TD的基本型別與描述
在檔案TargetCallingConv.td裡，首先出現的是CCAction。這是一個空類，作為表示呼叫慣例操作的基類。從它出發，有這些派生類：
19 class CCCustom <string fn> : CCAction {
20 string FuncName = fn;
21 }
CCCustom記錄處理引數的定製方法。
25 class CCPredicateAction <CCAction A> : CCAction {
26 CCAction SubAction = A;
27 }
CCPredicateAction作為基類，用在檢測特定謂詞，在謂詞成立時執行CCAction A的定義中。
31 class CCIfType <list<ValueType> vts, CCAction A> : CCPredicateAction<A> {
32 list<ValueType> VTs = vts;
33 }
CCPredicateAction派生類CCIfType用作判定指定型別，在指定型別之一出現時執行CCAction A。
36 class CCIf <string predicate, CCAction A> : CCPredicateAction<A> {
37 string Predicate = predicate;
38 }
而派生類CCIf作為基類，用在執行指定謂詞程式碼片段，在成立時執行CCAction A的定義中。
42 class CCIfByVal <CCAction A> : CCIf<"ArgFlags.isByVal()", A> {
43 }
CCIf派生類CCIfByVal則固化了謂詞程式碼片段——引數必須按值傳遞。
47 class CCIfConsecutiveRegs <CCAction A> : CCIf<"ArgFlags.isInConsecutiveRegs()", A> {
48 }
派生類CCIfConsecutiveRegs固化了另一種謂詞程式碼片段——引數必須在連續的暫存器裡。
51 class CCIfCC <string CC, CCAction A>
52 : CCIf<!strconcat("State.getCallingConv() == ", CC), A> {}
56 class CCIfInReg <CCAction A> : CCIf<"ArgFlags.isInReg()", A> {}
60 class CCIfNest <CCAction A> : CCIf<"ArgFlags.isNest()", A> {}
64 class CCIfSplit <CCAction A> : CCIf<"ArgFlags.isSplit()", A> {}
68 class CCIfSRet <CCAction A> : CCIf<"ArgFlags.isSRet()", A> {}
71 class CCIfVarArg <CCAction A> : CCIf<"State.isVarArg()", A> {}
74 class CCIfNotVarArg <CCAction A> : CCIf<"!State.isVarArg()", A> {}
上面出現的ArgFlags是原始檔CallingConvLower.h及CallingConvLower.cpp中型別為ISD::ArgFlagsTy 的函式引數，TableGen使用這個型別來描述被編譯函式的引數。
79 class CCAssignToReg <list<Register> regList> : CCAction {
80 list<Register> RegList = regList;
81 }
CCAssignToReg的語義由其名字隱含表述（即TableGen會解釋如下）：如果regList中仍有一個暫存器可用，就將值（定義中沒有出現）賦給該暫存器並返回成功。
85 class CCAssignToRegWithShadow <list<Register> regList,
86 list<Register> shadowList> : CCAction {
87 list<Register> RegList = regList;
88 list<Register> ShadowRegList = shadowList;
89 }
CCAssignToRegWithShadow類似於CCAssignToReg，不過還包括了一組成功時將被遮蔽的暫存器。
95 class CCAssignToStack <int size, int align> : CCAction {
96 int Size = size;
97 int Align = align;
98 }
CCAssignToStack則總是使得TableGen產生將值賦給指定大小與對齊邊界的棧區間。如果大小或對齊是0，使用ABI的定義值。
103 class CCAssignToStackWithShadow <int size,
104 int align,
105 list<Register> shadowList> : CCAction {
106 int Size = size;
107 int Align = align;
108 list<Register> ShadowRegList = shadowList;
109 }
同樣CCAssignToStackWithShadow類似CCAssignToStack，但還包括一組成功時將被遮蔽的暫存器。
114 class CCPassByVal <int size, int align> : CCAction {
115 int Size = size;
116 int Align = align;
117 }
CCPassByVal則會使TableGen產生將在棧上按值傳遞聚集引數的程式碼。
121 class CCPromoteToType <ValueType destTy> : CCAction {
122 ValueType DestTy = destTy;
123 }
CCPromoteToType告訴TableGen產生將當前值（定義中沒有出現）提升到destTy型別。
127 class CCPromoteToUpperBitsInType <ValueType destTy> : CCAction {
128 ValueType DestTy = destTy;
129 }
CCPromoteToUpperBitsInType類似CCPromoteToType，但還要將值移到高位。
133 class CCBitConvertToType <ValueType destTy> : CCAction {
134 ValueType DestTy = destTy;
135 }
CCBitConvertToType則是將當前值按位轉換（bitconvert）到destTy型別。
139 class CCPassIndirect <ValueType destTy> : CCAction {
140 ValueType DestTy = destTy;
141 }
CCPassIndirect告訴TableGen產生將值（定義中沒有出現）存入棧，將對應指標作為值傳遞的程式碼。
145 class CCDelegateTo <CallingConv cc> : CCAction {
146 CallingConv CC = cc;
147 }
CCDelegateTo指定所要委派執行的呼叫慣例。
接下來是描述呼叫慣例的基本型別。首先是CallingConv，它通過一組CCAction來刻畫呼叫慣例的具體行為。
151 class CallingConv <list<CCAction> actions> {
152 list<CCAction> Actions = actions;
153 bit Custom = 0;
154 }
然後是CustomCallingConv。實際使用上從CustomCallingConv派生的定義，表示使用LLVM中同名的函式來處理該呼叫慣例（這是少數不能由機器描述生成處理呼叫慣例程式碼的例子）。
158 class CustomCallingConv : CallingConv<[]> {
159 let Custom = 1;
160 }
最後就是CalleeSavedRegs的定義，描述被呼叫者儲存的暫存器，我們在一節中已經看過它的定義，以及X86對應的派生定義。
3.7.2. X86呼叫慣例的TD描述
X86呼叫慣例的描述在檔案X86CallingConv.td中。首先是CCIfSubtarget定義，它用於判定目標機器是否具有指定的特徵（feature）。
16 class CCIfSubtarget <string F, CCAction A>
17 : CCIf<!strconcat("static_cast<const X86Subtarget&>"
18 "(State.getMachineFunction().getSubtarget()).", F),
這裡State是CCState型別的物件，我們後面會詳細瞭解這個類。
3.7.2.1. 返回值慣例
接著是定義如何傳遞返回值的返回值慣例。首先是RetCC_X86Common。
26 def RetCC_X86Common : CallingConv<[
27 // Scalar values are returned in AX first, then DX. For i8, the ABI
28 // requires the values to be in AL and AH, however this code uses AL and DL
29 // instead. This is because using AH for the second register conflicts with
30 // the way LLVM does multiple return values -- a return of {i16,i8} would end
31 // up in AX and AH, which overlap. Front-ends wishing to conform to the ABI
32 // for functions that return two i8 values are currently expected to pack the
33 // values into an i16 (which uses AX, and thus AL:AH).
34 //
35 // For code that doesn't care about the ABI, we allow returning more than two
36 // integer values in registers.
37 <[i1], <i8>>,
38 CCIfType<[i8] ,<[AL, DL, CL]>>,
39 CCIfType<[i16], CCAssignToReg<[AX, DX, CX]>>,
40 CCIfType<[i32], CCAssignToReg<[EAX, EDX, ECX]>>,
41 CCIfType<[i64], CCAssignToReg<[RAX, RDX, RCX]>>,
42
43 // Boolean vectors of AVX-512 are returned in SIMD registers.
44 // The call from AVX to AVX-512 function should work,
45 // since the boolean types in AVX/AVX2 are promoted by default.
46 CCIfType<[v2i1], CCPromoteToType<v2i64>>,
47 CCIfType<[v4i1], CCPromoteToType<v4i32>>,
48 CCIfType<[v8i1], CCPromoteToType<v8i16>>,
49 CCIfType<[v16i1], CCPromoteToType<v16i8>>,
50 CCIfType<[v32i1], CCPromoteToType<v32i8>>,
51 CCIfType<[v64i1], CCPromoteToType<v64i8>>,
52
53 // Vector types are returned in XMM0 and XMM1, when they fit. XMM2 and XMM3
54 // can only be used by ABI non-compliant code. If the target doesn't have XMM
55 // registers, it won't have vector types.
56 CCIfType<[v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
57 CCAssignToReg<[XMM0,XMM1,XMM2,XMM3]>>,
58
59 // 256-bit vectors are returned in YMM0 and XMM1, when they fit. YMM2 and YMM3
60 // can only be used by ABI non-compliant code. This vector type is only
61 // supported while using the AVX target feature.
62 CCIfType<[v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
63 CCAssignToReg<[YMM0,YMM1,YMM2,YMM3]>>,
64
65 // 512-bit vectors are returned in ZMM0 and ZMM1, when they fit. ZMM2 and ZMM3
66 // can only be used by ABI non-compliant code. This vector type is only
67 // supported while using the AVX-512 target feature.
68 CCIfType<[v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
69 CCAssignToReg<[ZMM0,ZMM1,ZMM2,ZMM3]>>,
70
71 // MMX vector types are always returned in MM0. If the target doesn't have
72 // MM0, it doesn't support these vector types.
73 CCIfType<[x86mmx], CCAssignToReg<[MM0]>>,
74
75 // Long double types are always returned in FP0 (even with SSE).
76 CCIfType<[f80], CCAssignToReg<[FP0, FP1]>>
77 ]>;
這個呼叫慣例由一系列CCIfType定義組成，這使得其定義十分清晰。比如37行的CCIfType定義說明，如果當前引數型別是i1，把它提升至i8（即位元組型別）。這些CCIfType在TableGen生成程式碼時將被展開為一系列的if語句塊。
接著是32位的C返回值慣例RetCC_X86_32_C。除了型別為f32或f64的引數，其他處理它交由RetCC_X86Common來代勞。
80 def RetCC_X86_32_C : CallingConv<[
81 // The X86-32 calling convention returns FP values in FP0, unless marked
82 // with "inreg" (used here to distinguish one kind of reg from another,
83 // weirdly; this is really the sse-regparm calling convention) in which
84 // case they use XMM0, otherwise it is the same as the common X86 calling
85 // conv.
86 <<"hasSSE2()",
87 CCIfType<[f32, f64], CCAssignToReg<[XMM0,XMM1,XMM2]>>>>,
88 CCIfType<[f32,f64], CCAssignToReg<[FP0, FP1]>>,
89 <RetCC_X86Common>
90 ]>;
然後是32位的FastCC返回值慣例RetCC_X86_32_Fast。
93 def RetCC_X86_32_Fast : CallingConv<[
94 // The X86-32 fastcc returns 1, 2, or 3 FP values in XMM0-2 if the target has
95 // SSE2.
96 // This can happen when a float, 2 x float, or 3 x float vector is split by
97 // target lowering, and is returned in 1-3 sse regs.
98 CCIfType<[f32], CCIfSubtarget<"hasSSE2()", CCAssignToReg<[XMM0,XMM1,XMM2]>>>,
99 CCIfType<[f64], CCIfSubtarget<"hasSSE2()", CCAssignToReg<[XMM0,XMM1,XMM2]>>>,
100
101 // For integers, ECX can be used as an extra return register
102 CCIfType<[i8], CCAssignToReg<[AL, DL, CL]>>,
103 CCIfType<[i16], CCAssignToReg<[AX, DX, CX]>>,
104 CCIfType<[i32], CCAssignToReg<[EAX, EDX, ECX]>>,
105
106 // Otherwise, it is the same as the common X86 calling convention.
107 CCDelegateTo<RetCC_X86Common>
108 ]>;
以及Intel_OCL_BI的返回值慣例RetCC_Intel_OCL_BI（Intel OpenCL內建函式的呼叫慣例）。
111 def RetCC_Intel_OCL_BI : CallingConv<[
112 // Vector types are returned in XMM0,XMM1,XMMM2 and XMM3.
113 CCIfType<[f32, f64, v4i32, v2i64, v4f32, v2f64],
114 CCAssignToReg<[XMM0,XMM1,XMM2,XMM3]>>,
115
116 // 256-bit FP vectors
117 // No more than 4 registers
118 CCIfType<[v8f32, v4f64, v8i32, v4i64],
119 CCAssignToReg<[YMM0,YMM1,YMM2,YMM3]>>,
120
121 // 512-bit FP vectors
122 CCIfType<[v16f32, v8f64, v16i32, v8i64],
123 CCAssignToReg<[ZMM0,ZMM1,ZMM2,ZMM3]>>,
124
125 // i32, i64 in the standard way
126 CCDelegateTo<RetCC_X86Common>
127 ]>;
32位HiPE返回值慣例RetCC_X86_32_HiPE（高效能Erlang編譯器的呼叫慣例）。
130 def RetCC_X86_32_HiPE : CallingConv<[
131 // Promote all types to i32
132 CCIfType<[i8, i16], CCPromoteToType<i32>>,
133
134 // Return: HP, P, VAL1, VAL2
135 CCIfType<[i32], CCAssignToReg<[ESI, EBP, EAX, EDX]>>
136 ]>;
32位MSVC返回值慣例RetCC_X86_32_VectorCall，通過SSE暫存器傳遞向量。
139 def RetCC_X86_32_VectorCall : CallingConv<[
140 // Vector types are returned in XMM0,XMM1,XMMM2 and XMM3.
141 CCIfType<[f32, f64, v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
142 CCAssignToReg<[XMM0,XMM1,XMM2,XMM3]>>,
143
144 // 256-bit FP vectors
145 CCIfType<[v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
146 CCAssignToReg<[YMM0,YMM1,YMM2,YMM3]>>,
147
148 // 512-bit FP vectors
149 CCIfType<[v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
150 CCAssignToReg<[ZMM0,ZMM1,ZMM2,ZMM3]>>,
151
152 // Return integers in the standard way.
153 CCDelegateTo<RetCC_X86Common>
154 ]>;
64位的C返回值慣例RetCC_X86_64_C。
157 def RetCC_X86_64_C : CallingConv<[
158 // The X86-64 calling convention always returns FP values in XMM0.
159 CCIfType<[f32], CCAssignToReg<[XMM0, XMM1]>>,
160 CCIfType<[f64], CCAssignToReg<[XMM0, XMM1]>>,
161
162 // MMX vector types are always returned in XMM0.
163 CCIfType<[x86mmx], CCAssignToReg<[XMM0, XMM1]>>,
164 CCDelegateTo<RetCC_X86Common>
165 ]>;
64位Windows返回值慣例RetCC_X86_Win64_C。
168 def RetCC_X86_Win64_C : CallingConv<[
169 // The X86-Win64 calling convention always returns __m64 values in RAX.
170 CCIfType<[x86mmx], CCBitConvertToType<i64>>,
171
172 // Otherwise, everything is the same as 'normal' X86-64 C CC.
173 CCDelegateTo<RetCC_X86_64_C>
174 ]>;
64位HiPE返回值慣例RetCC_X86_64_HiPE。
177 def RetCC_X86_64_HiPE : CallingConv<[
178 // Promote all types to i64
179 CCIfType<[i8, i16, i32], CCPromoteToType<i64>>,
180
181 // Return: HP, P, VAL1, VAL2
182 CCIfType<[i64], CCAssignToReg<[R15, RBP, RAX, RDX]>>
183 ]>;
64位JScript返回值慣例RetCC_X86_64_WebKit_JS。
186 def RetCC_X86_64_WebKit_JS : CallingConv<[
187 // Promote all types to i64
188 CCIfType<[i8, i16, i32], CCPromoteToType<i64>>,
189
190 // Return: RAX
191 CCIfType<[i64], CCAssignToReg<[RAX]>>
192 ]>;
AnyReg返回值慣例RetCC_X86_64_AnyReg。它允許暫存器分配器選擇任何空閒的暫存器。中Debug build時，RetCC_X86_64_AnyReg將產生一個assert，在Release build時，會落入C慣例。
201 def RetCC_X86_64_AnyReg : CallingConv<[
202 CCCustom<"CC_X86_AnyReg_Error">
203 ]>;
最後三個定義將生成最終的返回值處理分派函式。
206 def RetCC_X86_32 : CallingConv<[
207 // If FastCC, use RetCC_X86_32_Fast.
208 CCIfCC<"CallingConv::Fast", CCDelegateTo<RetCC_X86_32_Fast>>,
209 // If HiPE, use RetCC_X86_32_HiPE.
210 CCIfCC<"CallingConv::HiPE", CCDelegateTo<RetCC_X86_32_HiPE>>,
211 CCIfCC<"CallingConv::X86_VectorCall", CCDelegateTo<RetCC_X86_32_VectorCall>>,
212
213 // Otherwise, use RetCC_X86_32_C.
214 CCDelegateTo<RetCC_X86_32_C>
215 ]>;
218 def RetCC_X86_64 : CallingConv<[
219 // HiPE uses RetCC_X86_64_HiPE
220 CCIfCC<"CallingConv::HiPE", CCDelegateTo<RetCC_X86_64_HiPE>>,
221
222 // Handle JavaScript calls.
223 CCIfCC<"CallingConv::WebKit_JS", CCDelegateTo<RetCC_X86_64_WebKit_JS>>,
224 CCIfCC<"CallingConv::AnyReg", CCDelegateTo<RetCC_X86_64_AnyReg>>,
225
226 // Handle explicit CC selection
227 CCIfCC<"CallingConv::X86_64_Win64", CCDelegateTo<RetCC_X86_Win64_C>>,
228 CCIfCC<"CallingConv::X86_64_SysV", CCDelegateTo<RetCC_X86_64_C>>,
229
230 // Mingw64 and native Win64 use Win64 CC
231 CCIfSubtarget<"isTargetWin64()", CCDelegateTo<RetCC_X86_Win64_C>>,
232
233 // Otherwise, drop to normal X86-64 CC
234 CCDelegateTo<RetCC_X86_64_C>
235 ]>;
238 def RetCC_X86 : CallingConv<[
239
240 // Check if this is the Intel OpenCL built-ins calling convention
241 CCIfCC<"CallingConv::Intel_OCL_BI", CCDelegateTo<RetCC_Intel_OCL_BI>>,
242
243 CCIfSubtarget<"is64Bit()", CCDelegateTo<RetCC_X86_64>>,
244 CCDelegateTo<RetCC_X86_32>
245 ]>;