1 //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Define several functions to decode x86 specific shuffle semantics into a
11 // generic vector mask.
13 //===----------------------------------------------------------------------===//
15 #include "X86ShuffleDecode.h"
16 #include "llvm/CodeGen/MachineValueType.h"
18 //===----------------------------------------------------------------------===//
19 // Vector Mask Decoding
20 //===----------------------------------------------------------------------===//
24 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
25 // Defaults the copying the dest value.
26 ShuffleMask.push_back(0);
27 ShuffleMask.push_back(1);
28 ShuffleMask.push_back(2);
29 ShuffleMask.push_back(3);
31 // Decode the immediate.
32 unsigned ZMask = Imm & 15;
33 unsigned CountD = (Imm >> 4) & 3;
34 unsigned CountS = (Imm >> 6) & 3;
36 // CountS selects which input element to use.
37 unsigned InVal = 4+CountS;
38 // CountD specifies which element of destination to update.
39 ShuffleMask[CountD] = InVal;
40 // ZMask zaps values, potentially overriding the CountD elt.
41 if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero;
42 if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero;
43 if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero;
44 if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero;
48 void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
49 for (unsigned i = NElts/2; i != NElts; ++i)
50 ShuffleMask.push_back(NElts+i);
52 for (unsigned i = NElts/2; i != NElts; ++i)
53 ShuffleMask.push_back(i);
57 void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) {
58 for (unsigned i = 0; i != NElts/2; ++i)
59 ShuffleMask.push_back(i);
61 for (unsigned i = 0; i != NElts/2; ++i)
62 ShuffleMask.push_back(NElts+i);
65 void DecodePALIGNRMask(MVT VT, unsigned Imm,
66 SmallVectorImpl<int> &ShuffleMask) {
67 unsigned NumElts = VT.getVectorNumElements();
68 unsigned Offset = Imm * (VT.getVectorElementType().getSizeInBits() / 8);
70 unsigned NumLanes = VT.getSizeInBits() / 128;
71 unsigned NumLaneElts = NumElts / NumLanes;
73 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
74 for (unsigned i = 0; i != NumLaneElts; ++i) {
75 unsigned Base = i + Offset;
76 // if i+offset is out of this lane then we actually need the other source
77 if (Base >= NumLaneElts) Base += NumElts - NumLaneElts;
78 ShuffleMask.push_back(Base + l);
83 /// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
84 /// VT indicates the type of the vector allowing it to handle different
85 /// datatypes and vector widths.
86 void DecodePSHUFMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
87 unsigned NumElts = VT.getVectorNumElements();
89 unsigned NumLanes = VT.getSizeInBits() / 128;
90 unsigned NumLaneElts = NumElts / NumLanes;
92 unsigned NewImm = Imm;
93 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
94 for (unsigned i = 0; i != NumLaneElts; ++i) {
95 ShuffleMask.push_back(NewImm % NumLaneElts + l);
96 NewImm /= NumLaneElts;
98 if (NumLaneElts == 4) NewImm = Imm; // reload imm
102 void DecodePSHUFHWMask(MVT VT, unsigned Imm,
103 SmallVectorImpl<int> &ShuffleMask) {
104 unsigned NumElts = VT.getVectorNumElements();
106 for (unsigned l = 0; l != NumElts; l += 8) {
107 unsigned NewImm = Imm;
108 for (unsigned i = 0, e = 4; i != e; ++i) {
109 ShuffleMask.push_back(l + i);
111 for (unsigned i = 4, e = 8; i != e; ++i) {
112 ShuffleMask.push_back(l + 4 + (NewImm & 3));
118 void DecodePSHUFLWMask(MVT VT, unsigned Imm,
119 SmallVectorImpl<int> &ShuffleMask) {
120 unsigned NumElts = VT.getVectorNumElements();
122 for (unsigned l = 0; l != NumElts; l += 8) {
123 unsigned NewImm = Imm;
124 for (unsigned i = 0, e = 4; i != e; ++i) {
125 ShuffleMask.push_back(l + (NewImm & 3));
128 for (unsigned i = 4, e = 8; i != e; ++i) {
129 ShuffleMask.push_back(l + i);
134 /// DecodeSHUFPMask - This decodes the shuffle masks for shufp*. VT indicates
135 /// the type of the vector allowing it to handle different datatypes and vector
137 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
138 unsigned NumElts = VT.getVectorNumElements();
140 unsigned NumLanes = VT.getSizeInBits() / 128;
141 unsigned NumLaneElts = NumElts / NumLanes;
143 unsigned NewImm = Imm;
144 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
145 // each half of a lane comes from different source
146 for (unsigned s = 0; s != NumElts*2; s += NumElts) {
147 for (unsigned i = 0; i != NumLaneElts/2; ++i) {
148 ShuffleMask.push_back(NewImm % NumLaneElts + s + l);
149 NewImm /= NumLaneElts;
152 if (NumLaneElts == 4) NewImm = Imm; // reload imm
156 /// DecodeUNPCKHMask - This decodes the shuffle masks for unpckhps/unpckhpd
157 /// and punpckh*. VT indicates the type of the vector allowing it to handle
158 /// different datatypes and vector widths.
159 void DecodeUNPCKHMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
160 unsigned NumElts = VT.getVectorNumElements();
162 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
163 // independently on 128-bit lanes.
164 unsigned NumLanes = VT.getSizeInBits() / 128;
165 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
166 unsigned NumLaneElts = NumElts / NumLanes;
168 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
169 for (unsigned i = l + NumLaneElts/2, e = l + NumLaneElts; i != e; ++i) {
170 ShuffleMask.push_back(i); // Reads from dest/src1
171 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
176 /// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
177 /// and punpckl*. VT indicates the type of the vector allowing it to handle
178 /// different datatypes and vector widths.
179 void DecodeUNPCKLMask(MVT VT, SmallVectorImpl<int> &ShuffleMask) {
180 unsigned NumElts = VT.getVectorNumElements();
182 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
183 // independently on 128-bit lanes.
184 unsigned NumLanes = VT.getSizeInBits() / 128;
185 if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
186 unsigned NumLaneElts = NumElts / NumLanes;
188 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
189 for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
190 ShuffleMask.push_back(i); // Reads from dest/src1
191 ShuffleMask.push_back(i+NumElts); // Reads from src/src2
196 void DecodeVPERM2X128Mask(MVT VT, unsigned Imm,
197 SmallVectorImpl<int> &ShuffleMask) {
199 return; // Not a shuffle
201 unsigned HalfSize = VT.getVectorNumElements()/2;
203 for (unsigned l = 0; l != 2; ++l) {
204 unsigned HalfBegin = ((Imm >> (l*4)) & 0x3) * HalfSize;
205 for (unsigned i = HalfBegin, e = HalfBegin+HalfSize; i != e; ++i)
206 ShuffleMask.push_back(i);
210 /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
211 /// No VT provided since it only works on 256-bit, 4 element vectors.
212 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
213 for (unsigned i = 0; i != 4; ++i) {
214 ShuffleMask.push_back((Imm >> (2*i)) & 3);