blob: 4d24c839f548a66673b263759a220e91718ac7aa [file] [log] [blame]
Bruno Cardoso Lopes6b1d0a32010-09-02 18:40:13 +00001//===-- X86ShuffleDecode.h - X86 shuffle decode logic ---------------------===//
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// Define several functions to decode x86 specific shuffle semantics into a
11// generic vector mask.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef X86_SHUFFLE_DECODE_H
16#define X86_SHUFFLE_DECODE_H
17
18#include "llvm/ADT/SmallVector.h"
19using namespace llvm;
20
21//===----------------------------------------------------------------------===//
22// Vector Mask Decoding
23//===----------------------------------------------------------------------===//
24
25enum {
26 SM_SentinelZero = ~0U
27};
28
29static 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
53static void DecodeMOVHLPSMask(SmallVectorImpl<unsigned> &ShuffleMask) {
54 ShuffleMask.push_back(3);
55 ShuffleMask.push_back(1);
56}
57
58static void DecodeMOVLHPSMask(SmallVectorImpl<unsigned> &ShuffleMask) {
59 ShuffleMask.push_back(0);
60 ShuffleMask.push_back(2);
61}
62
63static void DecodePSHUFMask(unsigned NElts, unsigned Imm,
64 SmallVectorImpl<unsigned> &ShuffleMask) {
65 for (unsigned i = 0; i != NElts; ++i) {
66 ShuffleMask.push_back(Imm % NElts);
67 Imm /= NElts;
68 }
69}
70
71static void DecodePSHUFHWMask(unsigned Imm,
72 SmallVectorImpl<unsigned> &ShuffleMask) {
73 ShuffleMask.push_back(0);
74 ShuffleMask.push_back(1);
75 ShuffleMask.push_back(2);
76 ShuffleMask.push_back(3);
77 for (unsigned i = 0; i != 4; ++i) {
78 ShuffleMask.push_back(4+(Imm & 3));
79 Imm >>= 2;
80 }
81}
82
83static void DecodePSHUFLWMask(unsigned Imm,
84 SmallVectorImpl<unsigned> &ShuffleMask) {
85 for (unsigned i = 0; i != 4; ++i) {
86 ShuffleMask.push_back((Imm & 3));
87 Imm >>= 2;
88 }
89 ShuffleMask.push_back(4);
90 ShuffleMask.push_back(5);
91 ShuffleMask.push_back(6);
92 ShuffleMask.push_back(7);
93}
94
95static void DecodePUNPCKLMask(unsigned NElts,
96 SmallVectorImpl<unsigned> &ShuffleMask) {
97 for (unsigned i = 0; i != NElts/2; ++i) {
98 ShuffleMask.push_back(i);
99 ShuffleMask.push_back(i+NElts);
100 }
101}
102
103static void DecodePUNPCKHMask(unsigned NElts,
104 SmallVectorImpl<unsigned> &ShuffleMask) {
105 for (unsigned i = 0; i != NElts/2; ++i) {
106 ShuffleMask.push_back(i+NElts/2);
107 ShuffleMask.push_back(i+NElts+NElts/2);
108 }
109}
110
111static void DecodeSHUFPSMask(unsigned NElts, unsigned Imm,
112 SmallVectorImpl<unsigned> &ShuffleMask) {
113 // Part that reads from dest.
114 for (unsigned i = 0; i != NElts/2; ++i) {
115 ShuffleMask.push_back(Imm % NElts);
116 Imm /= NElts;
117 }
118 // Part that reads from src.
119 for (unsigned i = 0; i != NElts/2; ++i) {
120 ShuffleMask.push_back(Imm % NElts + NElts);
121 Imm /= NElts;
122 }
123}
124
125static void DecodeUNPCKHPMask(unsigned NElts,
126 SmallVectorImpl<unsigned> &ShuffleMask) {
127 for (unsigned i = 0; i != NElts/2; ++i) {
128 ShuffleMask.push_back(i+NElts/2); // Reads from dest
129 ShuffleMask.push_back(i+NElts+NElts/2); // Reads from src
130 }
131}
132
133
134/// DecodeUNPCKLPMask - This decodes the shuffle masks for unpcklps/unpcklpd
135/// etc. NElts indicates the number of elements in the vector allowing it to
136/// handle different datatypes and vector widths.
137static void DecodeUNPCKLPMask(unsigned NElts,
138 SmallVectorImpl<unsigned> &ShuffleMask) {
139 for (unsigned i = 0; i != NElts/2; ++i) {
140 ShuffleMask.push_back(i); // Reads from dest
141 ShuffleMask.push_back(i+NElts); // Reads from src
142 }
143}
144
145#endif