1 | //===- X86FixupSetCC.cpp - fix zero-extension of setcc patterns -----------===// |
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 | // This file defines a pass that fixes zero-extension of setcc patterns. |
10 | // X86 setcc instructions are modeled to have no input arguments, and a single |
11 | // GR8 output argument. This is consistent with other similar instructions |
12 | // (e.g. movb), but means it is impossible to directly generate a setcc into |
13 | // the lower GR8 of a specified GR32. |
14 | // This means that ISel must select (zext (setcc)) into something like |
15 | // seta %al; movzbl %al, %eax. |
16 | // Unfortunately, this can cause a stall due to the partial register write |
17 | // performed by the setcc. Instead, we can use: |
18 | // xor %eax, %eax; seta %al |
19 | // This both avoids the stall, and encodes shorter. |
20 | // |
21 | // Furthurmore, we can use: |
22 | // setzua %al |
23 | // if feature zero-upper is available. It's faster than the xor+setcc sequence. |
24 | // When r16-r31 is used, it even encodes shorter. |
25 | //===----------------------------------------------------------------------===// |
26 | |
27 | #include "X86.h" |
28 | #include "X86InstrInfo.h" |
29 | #include "X86Subtarget.h" |
30 | #include "llvm/ADT/Statistic.h" |
31 | #include "llvm/CodeGen/MachineFunctionPass.h" |
32 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
33 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
34 | |
35 | using namespace llvm; |
36 | |
37 | #define DEBUG_TYPE "x86-fixup-setcc" |
38 | |
39 | STATISTIC(NumSubstZexts, "Number of setcc + zext pairs substituted" ); |
40 | |
41 | namespace { |
42 | class X86FixupSetCCPass : public MachineFunctionPass { |
43 | public: |
44 | static char ID; |
45 | |
46 | X86FixupSetCCPass() : MachineFunctionPass(ID) {} |
47 | |
48 | StringRef getPassName() const override { return "X86 Fixup SetCC" ; } |
49 | |
50 | bool runOnMachineFunction(MachineFunction &MF) override; |
51 | |
52 | private: |
53 | MachineRegisterInfo *MRI = nullptr; |
54 | const X86Subtarget *ST = nullptr; |
55 | const X86InstrInfo *TII = nullptr; |
56 | |
57 | enum { SearchBound = 16 }; |
58 | }; |
59 | } // end anonymous namespace |
60 | |
61 | char X86FixupSetCCPass::ID = 0; |
62 | |
63 | INITIALIZE_PASS(X86FixupSetCCPass, DEBUG_TYPE, DEBUG_TYPE, false, false) |
64 | |
65 | FunctionPass *llvm::createX86FixupSetCC() { return new X86FixupSetCCPass(); } |
66 | |
67 | bool X86FixupSetCCPass::runOnMachineFunction(MachineFunction &MF) { |
68 | bool Changed = false; |
69 | MRI = &MF.getRegInfo(); |
70 | ST = &MF.getSubtarget<X86Subtarget>(); |
71 | TII = ST->getInstrInfo(); |
72 | |
73 | SmallVector<MachineInstr*, 4> ToErase; |
74 | |
75 | for (auto &MBB : MF) { |
76 | MachineInstr *FlagsDefMI = nullptr; |
77 | for (auto &MI : MBB) { |
78 | // Remember the most recent preceding eflags defining instruction. |
79 | if (MI.definesRegister(Reg: X86::EFLAGS, /*TRI=*/nullptr)) |
80 | FlagsDefMI = &MI; |
81 | |
82 | // Find a setcc that is used by a zext. |
83 | // This doesn't have to be the only use, the transformation is safe |
84 | // regardless. |
85 | if (MI.getOpcode() != X86::SETCCr) |
86 | continue; |
87 | |
88 | MachineInstr *ZExt = nullptr; |
89 | Register Reg0 = MI.getOperand(i: 0).getReg(); |
90 | for (auto &Use : MRI->use_instructions(Reg: Reg0)) |
91 | if (Use.getOpcode() == X86::MOVZX32rr8) |
92 | ZExt = &Use; |
93 | |
94 | if (!ZExt) |
95 | continue; |
96 | |
97 | if (!FlagsDefMI) |
98 | continue; |
99 | |
100 | // We'd like to put something that clobbers eflags directly before |
101 | // FlagsDefMI. This can't hurt anything after FlagsDefMI, because |
102 | // it, itself, by definition, clobbers eflags. But it may happen that |
103 | // FlagsDefMI also *uses* eflags, in which case the transformation is |
104 | // invalid. |
105 | if (FlagsDefMI->readsRegister(Reg: X86::EFLAGS, /*TRI=*/nullptr)) |
106 | continue; |
107 | |
108 | // On 32-bit, we need to be careful to force an ABCD register. |
109 | const TargetRegisterClass *RC = |
110 | ST->is64Bit() ? &X86::GR32RegClass : &X86::GR32_ABCDRegClass; |
111 | if (!MRI->constrainRegClass(Reg: ZExt->getOperand(i: 0).getReg(), RC)) { |
112 | // If we cannot constrain the register, we would need an additional copy |
113 | // and are better off keeping the MOVZX32rr8 we have now. |
114 | continue; |
115 | } |
116 | |
117 | ++NumSubstZexts; |
118 | Changed = true; |
119 | |
120 | // X86 setcc/setzucc only takes an output GR8, so fake a GR32 input by |
121 | // inserting the setcc/setzucc result into the low byte of the zeroed |
122 | // register. |
123 | Register ZeroReg = MRI->createVirtualRegister(RegClass: RC); |
124 | if (ST->hasZU()) { |
125 | MI.setDesc(TII->get(Opcode: X86::SETZUCCr)); |
126 | BuildMI(BB&: *ZExt->getParent(), I: ZExt, MIMD: ZExt->getDebugLoc(), |
127 | MCID: TII->get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: ZeroReg); |
128 | } else { |
129 | // Initialize a register with 0. This must go before the eflags def |
130 | BuildMI(BB&: MBB, I: FlagsDefMI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: X86::MOV32r0), |
131 | DestReg: ZeroReg); |
132 | } |
133 | |
134 | BuildMI(BB&: *ZExt->getParent(), I: ZExt, MIMD: ZExt->getDebugLoc(), |
135 | MCID: TII->get(Opcode: X86::INSERT_SUBREG), DestReg: ZExt->getOperand(i: 0).getReg()) |
136 | .addReg(RegNo: ZeroReg) |
137 | .addReg(RegNo: Reg0) |
138 | .addImm(Val: X86::sub_8bit); |
139 | ToErase.push_back(Elt: ZExt); |
140 | } |
141 | } |
142 | |
143 | for (auto &I : ToErase) |
144 | I->eraseFromParent(); |
145 | |
146 | return Changed; |
147 | } |
148 | |