1 //===-- X86ShuffleDecode.h - 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"
17 //===----------------------------------------------------------------------===//
18 // Vector Mask Decoding
19 //===----------------------------------------------------------------------===//
23 void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<unsigned> &ShuffleMask) {
24 // Defaults the copying the dest value.
25 ShuffleMask.push_back(0);
26 ShuffleMask.push_back(1);
27 ShuffleMask.push_back(2);
28 ShuffleMask.push_back(3);
30 // Decode the immediate.
31 unsigned ZMask = Imm & 15;
32 unsigned CountD = (Imm >> 4) & 3;
33 unsigned CountS = (Imm >> 6) & 3;
35 // CountS selects which input element to use.
36 unsigned InVal = 4+CountS;
37 // CountD specifies which element of destination to update.
38 ShuffleMask[CountD] = InVal;
39 // ZMask zaps values, potentially overriding the CountD elt.
40 if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero;
41 if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero;
42 if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero;
43 if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero;
47 void DecodeMOVHLPSMask(unsigned NElts,
48 SmallVectorImpl<unsigned> &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,
58 SmallVectorImpl<unsigned> &ShuffleMask) {
59 for (unsigned i = 0; i != NElts/2; ++i)
60 ShuffleMask.push_back(i);
62 for (unsigned i = 0; i != NElts/2; ++i)
63 ShuffleMask.push_back(NElts+i);
66 void DecodePSHUFMask(unsigned NElts, unsigned Imm,
67 SmallVectorImpl<unsigned> &ShuffleMask) {
68 for (unsigned i = 0; i != NElts; ++i) {
69 ShuffleMask.push_back(Imm % NElts);
74 void DecodePSHUFHWMask(unsigned Imm,
75 SmallVectorImpl<unsigned> &ShuffleMask) {
76 ShuffleMask.push_back(0);
77 ShuffleMask.push_back(1);
78 ShuffleMask.push_back(2);
79 ShuffleMask.push_back(3);
80 for (unsigned i = 0; i != 4; ++i) {
81 ShuffleMask.push_back(4+(Imm & 3));
86 void DecodePSHUFLWMask(unsigned Imm,
87 SmallVectorImpl<unsigned> &ShuffleMask) {
88 for (unsigned i = 0; i != 4; ++i) {
89 ShuffleMask.push_back((Imm & 3));
92 ShuffleMask.push_back(4);
93 ShuffleMask.push_back(5);
94 ShuffleMask.push_back(6);
95 ShuffleMask.push_back(7);
98 void DecodePUNPCKLMask(unsigned NElts,
99 SmallVectorImpl<unsigned> &ShuffleMask) {
100 for (unsigned i = 0; i != NElts/2; ++i) {
101 ShuffleMask.push_back(i);
102 ShuffleMask.push_back(i+NElts);
106 void DecodePUNPCKHMask(unsigned NElts,
107 SmallVectorImpl<unsigned> &ShuffleMask) {
108 for (unsigned i = 0; i != NElts/2; ++i) {
109 ShuffleMask.push_back(i+NElts/2);
110 ShuffleMask.push_back(i+NElts+NElts/2);
114 void DecodeSHUFPSMask(unsigned NElts, unsigned Imm,
115 SmallVectorImpl<unsigned> &ShuffleMask) {
116 // Part that reads from dest.
117 for (unsigned i = 0; i != NElts/2; ++i) {
118 ShuffleMask.push_back(Imm % NElts);
121 // Part that reads from src.
122 for (unsigned i = 0; i != NElts/2; ++i) {
123 ShuffleMask.push_back(Imm % NElts + NElts);
128 void DecodeUNPCKHPMask(unsigned NElts,
129 SmallVectorImpl<unsigned> &ShuffleMask) {
130 for (unsigned i = 0; i != NElts/2; ++i) {
131 ShuffleMask.push_back(i+NElts/2); // Reads from dest
132 ShuffleMask.push_back(i+NElts+NElts/2); // Reads from src
137 /// DecodeUNPCKLPMask - This decodes the shuffle masks for unpcklps/unpcklpd
138 /// etc. NElts indicates the number of elements in the vector allowing it to
139 /// handle different datatypes and vector widths.
140 void DecodeUNPCKLPMask(unsigned NElts,
141 SmallVectorImpl<unsigned> &ShuffleMask) {
142 for (unsigned i = 0; i != NElts/2; ++i) {
143 ShuffleMask.push_back(i); // Reads from dest
144 ShuffleMask.push_back(i+NElts); // Reads from src