1 | //===-- SIProgramInfo.cpp ----------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file |
10 | /// |
11 | /// The SIProgramInfo tracks resource usage and hardware flags for kernels and |
12 | /// entry functions. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | // |
16 | |
17 | #include "SIProgramInfo.h" |
18 | #include "GCNSubtarget.h" |
19 | #include "SIDefines.h" |
20 | #include "Utils/AMDGPUBaseInfo.h" |
21 | #include "llvm/MC/MCExpr.h" |
22 | |
23 | using namespace llvm; |
24 | |
25 | void SIProgramInfo::reset(const MachineFunction &MF) { |
26 | MCContext &Ctx = MF.getContext(); |
27 | |
28 | const MCExpr *ZeroExpr = MCConstantExpr::create(Value: 0, Ctx); |
29 | |
30 | CodeSizeInBytes.reset(); |
31 | |
32 | VGPRBlocks = ZeroExpr; |
33 | SGPRBlocks = ZeroExpr; |
34 | Priority = 0; |
35 | FloatMode = 0; |
36 | Priv = 0; |
37 | DX10Clamp = 0; |
38 | DebugMode = 0; |
39 | IEEEMode = 0; |
40 | WgpMode = 0; |
41 | MemOrdered = 0; |
42 | FwdProgress = 0; |
43 | RrWgMode = 0; |
44 | ScratchSize = ZeroExpr; |
45 | |
46 | LDSBlocks = 0; |
47 | ScratchBlocks = ZeroExpr; |
48 | |
49 | ScratchEnable = ZeroExpr; |
50 | UserSGPR = 0; |
51 | TrapHandlerEnable = 0; |
52 | TGIdXEnable = 0; |
53 | TGIdYEnable = 0; |
54 | TGIdZEnable = 0; |
55 | TGSizeEnable = 0; |
56 | TIdIGCompCount = 0; |
57 | EXCPEnMSB = 0; |
58 | LdsSize = 0; |
59 | EXCPEnable = 0; |
60 | |
61 | ComputePGMRSrc3 = ZeroExpr; |
62 | |
63 | NumVGPR = ZeroExpr; |
64 | NumArchVGPR = ZeroExpr; |
65 | NumAccVGPR = ZeroExpr; |
66 | AccumOffset = ZeroExpr; |
67 | TgSplit = 0; |
68 | NumSGPR = ZeroExpr; |
69 | SGPRSpill = 0; |
70 | VGPRSpill = 0; |
71 | LDSSize = 0; |
72 | FlatUsed = ZeroExpr; |
73 | |
74 | NumSGPRsForWavesPerEU = ZeroExpr; |
75 | NumVGPRsForWavesPerEU = ZeroExpr; |
76 | Occupancy = ZeroExpr; |
77 | DynamicCallStack = ZeroExpr; |
78 | VCCUsed = ZeroExpr; |
79 | } |
80 | |
81 | static uint64_t getComputePGMRSrc1Reg(const SIProgramInfo &ProgInfo, |
82 | const GCNSubtarget &ST) { |
83 | uint64_t Reg = S_00B848_PRIORITY(ProgInfo.Priority) | |
84 | S_00B848_FLOAT_MODE(ProgInfo.FloatMode) | |
85 | S_00B848_PRIV(ProgInfo.Priv) | |
86 | S_00B848_DEBUG_MODE(ProgInfo.DebugMode) | |
87 | S_00B848_WGP_MODE(ProgInfo.WgpMode) | |
88 | S_00B848_MEM_ORDERED(ProgInfo.MemOrdered); |
89 | |
90 | if (ST.hasDX10ClampMode()) |
91 | Reg |= S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp); |
92 | |
93 | if (ST.hasIEEEMode()) |
94 | Reg |= S_00B848_IEEE_MODE(ProgInfo.IEEEMode); |
95 | |
96 | // TODO: in the long run we will want to enable this unconditionally. |
97 | if (ST.getTargetTriple().getOS() == Triple::OSType::AMDHSA) |
98 | Reg |= S_00B848_FWD_PROGRESS(ProgInfo.FwdProgress); |
99 | |
100 | if (ST.hasRrWGMode()) |
101 | Reg |= S_00B848_RR_WG_MODE(ProgInfo.RrWgMode); |
102 | |
103 | return Reg; |
104 | } |
105 | |
106 | static uint64_t getPGMRSrc1Reg(const SIProgramInfo &ProgInfo, |
107 | CallingConv::ID CC, const GCNSubtarget &ST) { |
108 | uint64_t Reg = S_00B848_PRIORITY(ProgInfo.Priority) | |
109 | S_00B848_FLOAT_MODE(ProgInfo.FloatMode) | |
110 | S_00B848_PRIV(ProgInfo.Priv) | |
111 | S_00B848_DEBUG_MODE(ProgInfo.DebugMode); |
112 | |
113 | if (ST.hasDX10ClampMode()) |
114 | Reg |= S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp); |
115 | |
116 | if (ST.hasIEEEMode()) |
117 | Reg |= S_00B848_IEEE_MODE(ProgInfo.IEEEMode); |
118 | |
119 | if (ST.hasRrWGMode()) |
120 | Reg |= S_00B848_RR_WG_MODE(ProgInfo.RrWgMode); |
121 | |
122 | switch (CC) { |
123 | case CallingConv::AMDGPU_PS: |
124 | Reg |= S_00B028_MEM_ORDERED(ProgInfo.MemOrdered); |
125 | break; |
126 | case CallingConv::AMDGPU_VS: |
127 | Reg |= S_00B128_MEM_ORDERED(ProgInfo.MemOrdered); |
128 | break; |
129 | case CallingConv::AMDGPU_GS: |
130 | Reg |= S_00B228_WGP_MODE(ProgInfo.WgpMode) | |
131 | S_00B228_MEM_ORDERED(ProgInfo.MemOrdered); |
132 | break; |
133 | case CallingConv::AMDGPU_HS: |
134 | Reg |= S_00B428_WGP_MODE(ProgInfo.WgpMode) | |
135 | S_00B428_MEM_ORDERED(ProgInfo.MemOrdered); |
136 | break; |
137 | default: |
138 | break; |
139 | } |
140 | return Reg; |
141 | } |
142 | |
143 | static uint64_t getComputePGMRSrc2Reg(const SIProgramInfo &ProgInfo) { |
144 | uint64_t Reg = S_00B84C_USER_SGPR(ProgInfo.UserSGPR) | |
145 | S_00B84C_TRAP_HANDLER(ProgInfo.TrapHandlerEnable) | |
146 | S_00B84C_TGID_X_EN(ProgInfo.TGIdXEnable) | |
147 | S_00B84C_TGID_Y_EN(ProgInfo.TGIdYEnable) | |
148 | S_00B84C_TGID_Z_EN(ProgInfo.TGIdZEnable) | |
149 | S_00B84C_TG_SIZE_EN(ProgInfo.TGSizeEnable) | |
150 | S_00B84C_TIDIG_COMP_CNT(ProgInfo.TIdIGCompCount) | |
151 | S_00B84C_EXCP_EN_MSB(ProgInfo.EXCPEnMSB) | |
152 | S_00B84C_LDS_SIZE(ProgInfo.LdsSize) | |
153 | S_00B84C_EXCP_EN(ProgInfo.EXCPEnable); |
154 | |
155 | return Reg; |
156 | } |
157 | |
158 | static const MCExpr *MaskShift(const MCExpr *Val, uint32_t Mask, uint32_t Shift, |
159 | MCContext &Ctx) { |
160 | if (Mask) { |
161 | const MCExpr *MaskExpr = MCConstantExpr::create(Value: Mask, Ctx); |
162 | Val = MCBinaryExpr::createAnd(LHS: Val, RHS: MaskExpr, Ctx); |
163 | } |
164 | if (Shift) { |
165 | const MCExpr *ShiftExpr = MCConstantExpr::create(Value: Shift, Ctx); |
166 | Val = MCBinaryExpr::createShl(LHS: Val, RHS: ShiftExpr, Ctx); |
167 | } |
168 | return Val; |
169 | } |
170 | |
171 | const MCExpr *SIProgramInfo::getComputePGMRSrc1(const GCNSubtarget &ST, |
172 | MCContext &Ctx) const { |
173 | uint64_t Reg = getComputePGMRSrc1Reg(ProgInfo: *this, ST); |
174 | const MCExpr *RegExpr = MCConstantExpr::create(Value: Reg, Ctx); |
175 | const MCExpr *Res = MCBinaryExpr::createOr( |
176 | LHS: MaskShift(Val: VGPRBlocks, /*Mask=*/0x3F, /*Shift=*/0, Ctx), |
177 | RHS: MaskShift(Val: SGPRBlocks, /*Mask=*/0xF, /*Shift=*/6, Ctx), Ctx); |
178 | return MCBinaryExpr::createOr(LHS: RegExpr, RHS: Res, Ctx); |
179 | } |
180 | |
181 | const MCExpr *SIProgramInfo::getPGMRSrc1(CallingConv::ID CC, |
182 | const GCNSubtarget &ST, |
183 | MCContext &Ctx) const { |
184 | if (AMDGPU::isCompute(CC)) { |
185 | return getComputePGMRSrc1(ST, Ctx); |
186 | } |
187 | |
188 | uint64_t Reg = getPGMRSrc1Reg(ProgInfo: *this, CC, ST); |
189 | const MCExpr *RegExpr = MCConstantExpr::create(Value: Reg, Ctx); |
190 | const MCExpr *Res = MCBinaryExpr::createOr( |
191 | LHS: MaskShift(Val: VGPRBlocks, /*Mask=*/0x3F, /*Shift=*/0, Ctx), |
192 | RHS: MaskShift(Val: SGPRBlocks, /*Mask=*/0xF, /*Shift=*/6, Ctx), Ctx); |
193 | return MCBinaryExpr::createOr(LHS: RegExpr, RHS: Res, Ctx); |
194 | } |
195 | |
196 | const MCExpr *SIProgramInfo::getComputePGMRSrc2(MCContext &Ctx) const { |
197 | uint64_t Reg = getComputePGMRSrc2Reg(ProgInfo: *this); |
198 | const MCExpr *RegExpr = MCConstantExpr::create(Value: Reg, Ctx); |
199 | return MCBinaryExpr::createOr(LHS: ScratchEnable, RHS: RegExpr, Ctx); |
200 | } |
201 | |
202 | const MCExpr *SIProgramInfo::getPGMRSrc2(CallingConv::ID CC, |
203 | MCContext &Ctx) const { |
204 | if (AMDGPU::isCompute(CC)) |
205 | return getComputePGMRSrc2(Ctx); |
206 | |
207 | return MCConstantExpr::create(Value: 0, Ctx); |
208 | } |
209 | |
210 | uint64_t SIProgramInfo::getFunctionCodeSize(const MachineFunction &MF, |
211 | bool IsLowerBound) { |
212 | if (!IsLowerBound && CodeSizeInBytes.has_value()) |
213 | return *CodeSizeInBytes; |
214 | |
215 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
216 | const SIInstrInfo *TII = STM.getInstrInfo(); |
217 | |
218 | uint64_t CodeSize = 0; |
219 | |
220 | for (const MachineBasicBlock &MBB : MF) { |
221 | // The amount of padding to align code can be both underestimated and |
222 | // overestimated. In case of inline asm used getInstSizeInBytes() will |
223 | // return a maximum size of a single instruction, where the real size may |
224 | // differ. At this point CodeSize may be already off. |
225 | if (!IsLowerBound) |
226 | CodeSize = alignTo(Size: CodeSize, A: MBB.getAlignment()); |
227 | |
228 | for (const MachineInstr &MI : MBB) { |
229 | // TODO: CodeSize should account for multiple functions. |
230 | |
231 | if (MI.isMetaInstruction()) |
232 | continue; |
233 | |
234 | // We cannot properly estimate inline asm size. It can be as small as zero |
235 | // if that is just a comment. |
236 | if (IsLowerBound && MI.isInlineAsm()) |
237 | continue; |
238 | |
239 | CodeSize += TII->getInstSizeInBytes(MI); |
240 | } |
241 | } |
242 | |
243 | CodeSizeInBytes = CodeSize; |
244 | return CodeSize; |
245 | } |
246 | |