Chris Lattner | 6aa928d | 2010-08-28 20:42:31 +0000 | [diff] [blame^] | 1 | //===-- X86InstComments.cpp - Generate verbose-asm comments for instrs ----===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This defines functionality used to emit comments about X86 instructions to |
| 11 | // an output stream for -fverbose-asm. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "X86InstComments.h" |
| 16 | #include "X86GenInstrNames.inc" |
| 17 | #include "llvm/MC/MCInst.h" |
| 18 | #include "llvm/Support/raw_ostream.h" |
| 19 | using namespace llvm; |
| 20 | |
| 21 | //===----------------------------------------------------------------------===// |
| 22 | // Vector Mask Decoding |
| 23 | //===----------------------------------------------------------------------===// |
| 24 | |
| 25 | enum { |
| 26 | SM_SentinelZero = ~0U |
| 27 | }; |
| 28 | |
| 29 | static void DecodeINSERTPSMask(unsigned Imm, |
| 30 | SmallVectorImpl<unsigned> &ShuffleMask) { |
| 31 | // Defaults the copying the dest value. |
| 32 | ShuffleMask.push_back(0); |
| 33 | ShuffleMask.push_back(1); |
| 34 | ShuffleMask.push_back(2); |
| 35 | ShuffleMask.push_back(3); |
| 36 | |
| 37 | // Decode the immediate. |
| 38 | unsigned ZMask = Imm & 15; |
| 39 | unsigned CountD = (Imm >> 4) & 3; |
| 40 | unsigned CountS = (Imm >> 6) & 3; |
| 41 | |
| 42 | // CountS selects which input element to use. |
| 43 | unsigned InVal = 4+CountS; |
| 44 | // CountD specifies which element of destination to update. |
| 45 | ShuffleMask[CountD] = InVal; |
| 46 | // ZMask zaps values, potentially overriding the CountD elt. |
| 47 | if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero; |
| 48 | if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero; |
| 49 | if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero; |
| 50 | if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero; |
| 51 | } |
| 52 | |
| 53 | static void DecodeSHUFPSMask(unsigned NElts, unsigned Imm, |
| 54 | SmallVectorImpl<unsigned> &ShuffleMask) { |
| 55 | // Part that reads from dest. |
| 56 | for (unsigned i = 0; i != NElts/2; ++i) { |
| 57 | ShuffleMask.push_back(Imm % NElts); |
| 58 | Imm /= NElts; |
| 59 | } |
| 60 | // Part that reads from src. |
| 61 | for (unsigned i = 0; i != NElts/2; ++i) { |
| 62 | ShuffleMask.push_back(Imm % NElts + NElts); |
| 63 | Imm /= NElts; |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | static void DecodePSHUFMask(unsigned NElts, unsigned Imm, |
| 68 | SmallVectorImpl<unsigned> &ShuffleMask) { |
| 69 | for (unsigned i = 0; i != NElts; ++i) { |
| 70 | ShuffleMask.push_back(Imm % NElts); |
| 71 | Imm /= NElts; |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | /// DecodeUNPCKLPMask - This decodes the shuffle masks for unpcklps/unpcklpd |
| 76 | /// etc. NElts indicates the number of elements in the vector allowing it to |
| 77 | /// handle different datatypes and vector widths. |
| 78 | static void DecodeUNPCKLPMask(unsigned NElts, |
| 79 | SmallVectorImpl<unsigned> &ShuffleMask) { |
| 80 | for (unsigned i = 0; i != NElts/2; ++i) { |
| 81 | ShuffleMask.push_back(i); // Reads from dest |
| 82 | ShuffleMask.push_back(i+NElts); // Reads from src |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | //===----------------------------------------------------------------------===// |
| 87 | // Top Level Entrypoint |
| 88 | //===----------------------------------------------------------------------===// |
| 89 | |
| 90 | |
| 91 | /// EmitAnyX86InstComments - This function decodes x86 instructions and prints |
| 92 | /// newline terminated strings to the specified string if desired. This |
| 93 | /// information is shown in disassembly dumps when verbose assembly is enabled. |
| 94 | void llvm::EmitAnyX86InstComments(const MCInst *MI, raw_ostream &OS, |
| 95 | const char *(*getRegName)(unsigned)) { |
| 96 | // If this is a shuffle operation, the switch should fill in this state. |
| 97 | SmallVector<unsigned, 8> ShuffleMask; |
| 98 | const char *DestName = 0, *Src1Name = 0, *Src2Name = 0; |
| 99 | |
| 100 | switch (MI->getOpcode()) { |
| 101 | case X86::INSERTPSrr: |
| 102 | Src1Name = getRegName(MI->getOperand(1).getReg()); |
| 103 | Src2Name = getRegName(MI->getOperand(2).getReg()); |
| 104 | DecodeINSERTPSMask(MI->getOperand(3).getImm(), ShuffleMask); |
| 105 | break; |
| 106 | case X86::PSHUFDri: |
| 107 | Src1Name = getRegName(MI->getOperand(1).getReg()); |
| 108 | // FALL THROUGH. |
| 109 | case X86::PSHUFDmi: |
| 110 | DestName = getRegName(MI->getOperand(0).getReg()); |
| 111 | DecodePSHUFMask(4, MI->getOperand(MI->getNumOperands()-1).getImm(), |
| 112 | ShuffleMask); |
| 113 | break; |
| 114 | |
| 115 | case X86::SHUFPDrri: |
| 116 | DecodeSHUFPSMask(2, MI->getOperand(3).getImm(), ShuffleMask); |
| 117 | Src1Name = getRegName(MI->getOperand(0).getReg()); |
| 118 | Src2Name = getRegName(MI->getOperand(2).getReg()); |
| 119 | break; |
| 120 | case X86::SHUFPSrri: |
| 121 | DecodeSHUFPSMask(4, MI->getOperand(3).getImm(), ShuffleMask); |
| 122 | Src1Name = getRegName(MI->getOperand(0).getReg()); |
| 123 | Src2Name = getRegName(MI->getOperand(2).getReg()); |
| 124 | break; |
| 125 | case X86::UNPCKLPSrr: |
| 126 | Src2Name = getRegName(MI->getOperand(2).getReg()); |
| 127 | // FALL THROUGH. |
| 128 | case X86::UNPCKLPSrm: |
| 129 | DecodeUNPCKLPMask(4, ShuffleMask); |
| 130 | Src1Name = getRegName(MI->getOperand(0).getReg()); |
| 131 | break; |
| 132 | } |
| 133 | |
| 134 | |
| 135 | // If this was a shuffle operation, print the shuffle mask. |
| 136 | if (!ShuffleMask.empty()) { |
| 137 | if (DestName == 0) DestName = Src1Name; |
| 138 | OS << (DestName ? DestName : "mem") << " = "; |
| 139 | |
| 140 | // If the two sources are the same, canonicalize the input elements to be |
| 141 | // from the first src so that we get larger element spans. |
| 142 | if (Src1Name == Src2Name) { |
| 143 | for (unsigned i = 0, e = ShuffleMask.size(); i != e; ++i) { |
| 144 | if ((int)ShuffleMask[i] >= 0 && // Not sentinel. |
| 145 | ShuffleMask[i] >= e) // From second mask. |
| 146 | ShuffleMask[i] -= e; |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | // The shuffle mask specifies which elements of the src1/src2 fill in the |
| 151 | // destination, with a few sentinel values. Loop through and print them |
| 152 | // out. |
| 153 | for (unsigned i = 0, e = ShuffleMask.size(); i != e; ++i) { |
| 154 | if (i != 0) |
| 155 | OS << ','; |
| 156 | if (ShuffleMask[i] == SM_SentinelZero) { |
| 157 | OS << "zero"; |
| 158 | continue; |
| 159 | } |
| 160 | |
| 161 | // Otherwise, it must come from src1 or src2. Print the span of elements |
| 162 | // that comes from this src. |
| 163 | bool isSrc1 = ShuffleMask[i] < ShuffleMask.size(); |
| 164 | const char *SrcName = isSrc1 ? Src1Name : Src2Name; |
| 165 | OS << (SrcName ? SrcName : "mem") << '['; |
| 166 | bool IsFirst = true; |
| 167 | while (i != e && |
| 168 | (int)ShuffleMask[i] >= 0 && |
| 169 | (ShuffleMask[i] < ShuffleMask.size()) == isSrc1) { |
| 170 | if (!IsFirst) |
| 171 | OS << ','; |
| 172 | else |
| 173 | IsFirst = false; |
| 174 | OS << ShuffleMask[i] % ShuffleMask.size(); |
| 175 | ++i; |
| 176 | } |
| 177 | OS << ']'; |
| 178 | --i; // For loop increments element #. |
| 179 | } |
| 180 | //MI->print(OS, 0); |
| 181 | OS << "\n"; |
| 182 | } |
| 183 | |
| 184 | } |