77
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1 ==============================================
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2 Using ARM NEON instructions in big endian mode
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3 ==============================================
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
4
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
5 .. contents::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
6 :local:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
7
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
8 Introduction
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
9 ============
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
10
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
11 Generating code for big endian ARM processors is for the most part straightforward. NEON loads and stores however have some interesting properties that make code generation decisions less obvious in big endian mode.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
12
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
13 The aim of this document is to explain the problem with NEON loads and stores, and the solution that has been implemented in LLVM.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
14
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
15 In this document the term "vector" refers to what the ARM ABI calls a "short vector", which is a sequence of items that can fit in a NEON register. This sequence can be 64 or 128 bits in length, and can constitute 8, 16, 32 or 64 bit items. This document refers to A64 instructions throughout, but is almost applicable to the A32/ARMv7 instruction sets also. The ABI format for passing vectors in A32 is sligtly different to A64. Apart from that, the same concepts apply.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
16
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
17 Example: C-level intrinsics -> assembly
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
18 ---------------------------------------
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
19
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
20 It may be helpful first to illustrate how C-level ARM NEON intrinsics are lowered to instructions.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
21
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
22 This trivial C function takes a vector of four ints and sets the zero'th lane to the value "42"::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
23
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
24 #include <arm_neon.h>
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
25 int32x4_t f(int32x4_t p) {
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
26 return vsetq_lane_s32(42, p, 0);
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
27 }
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
28
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
29 arm_neon.h intrinsics generate "generic" IR where possible (that is, normal IR instructions not ``llvm.arm.neon.*`` intrinsic calls). The above generates::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
30
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
31 define <4 x i32> @f(<4 x i32> %p) {
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
32 %vset_lane = insertelement <4 x i32> %p, i32 42, i32 0
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
33 ret <4 x i32> %vset_lane
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
34 }
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
35
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
36 Which then becomes the following trivial assembly::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
37
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
38 f: // @f
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
39 movz w8, #0x2a
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
40 ins v0.s[0], w8
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
41 ret
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
42
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
43 Problem
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
44 =======
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
45
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
46 The main problem is how vectors are represented in memory and in registers.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
47
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
48 First, a recap. The "endianness" of an item affects its representation in memory only. In a register, a number is just a sequence of bits - 64 bits in the case of AArch64 general purpose registers. Memory, however, is a sequence of addressable units of 8 bits in size. Any number greater than 8 bits must therefore be split up into 8-bit chunks, and endianness describes the order in which these chunks are laid out in memory.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
49
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
50 A "little endian" layout has the least significant byte first (lowest in memory address). A "big endian" layout has the *most* significant byte first. This means that when loading an item from big endian memory, the lowest 8-bits in memory must go in the most significant 8-bits, and so forth.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
51
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
52 ``LDR`` and ``LD1``
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
53 ===================
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
54
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
55 .. figure:: ARM-BE-ldr.png
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
56 :align: right
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
57
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
58 Big endian vector load using ``LDR``.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
59
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
60
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
61 A vector is a consecutive sequence of items that are operated on simultaneously. To load a 64-bit vector, 64 bits need to be read from memory. In little endian mode, we can do this by just performing a 64-bit load - ``LDR q0, [foo]``. However if we try this in big endian mode, because of the byte swapping the lane indices end up being swapped! The zero'th item as laid out in memory becomes the n'th lane in the vector.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
62
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
63 .. figure:: ARM-BE-ld1.png
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
64 :align: right
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
65
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
66 Big endian vector load using ``LD1``. Note that the lanes retain the correct ordering.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
67
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
68
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
69 Because of this, the instruction ``LD1`` performs a vector load but performs byte swapping not on the entire 64 bits, but on the individual items within the vector. This means that the register content is the same as it would have been on a little endian system.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
70
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
71 It may seem that ``LD1`` should suffice to peform vector loads on a big endian machine. However there are pros and cons to the two approaches that make it less than simple which register format to pick.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
72
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
73 There are two options:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
74
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
75 1. The content of a vector register is the same *as if* it had been loaded with an ``LDR`` instruction.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
76 2. The content of a vector register is the same *as if* it had been loaded with an ``LD1`` instruction.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
77
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
78 Because ``LD1 == LDR + REV`` and similarly ``LDR == LD1 + REV`` (on a big endian system), we can simulate either type of load with the other type of load plus a ``REV`` instruction. So we're not deciding which instructions to use, but which format to use (which will then influence which instruction is best to use).
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
79
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
80 .. The 'clearer' container is required to make the following section header come after the floated
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
81 images above.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
82 .. container:: clearer
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
83
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
84 Note that throughout this section we only mention loads. Stores have exactly the same problems as their associated loads, so have been skipped for brevity.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
85
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
86
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
87 Considerations
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
88 ==============
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
89
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
90 LLVM IR Lane ordering
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
91 ---------------------
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
92
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
93 LLVM IR has first class vector types. In LLVM IR, the zero'th element of a vector resides at the lowest memory address. The optimizer relies on this property in certain areas, for example when concatenating vectors together. The intention is for arrays and vectors to have identical memory layouts - ``[4 x i8]`` and ``<4 x i8>`` should be represented the same in memory. Without this property there would be many special cases that the optimizer would have to cleverly handle.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
94
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
95 Use of ``LDR`` would break this lane ordering property. This doesn't preclude the use of ``LDR``, but we would have to do one of two things:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
96
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
97 1. Insert a ``REV`` instruction to reverse the lane order after every ``LDR``.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
98 2. Disable all optimizations that rely on lane layout, and for every access to an individual lane (``insertelement``/``extractelement``/``shufflevector``) reverse the lane index.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
99
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
100 AAPCS
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
101 -----
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
102
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
103 The ARM procedure call standard (AAPCS) defines the ABI for passing vectors between functions in registers. It states:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
104
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
105 When a short vector is transferred between registers and memory it is treated as an opaque object. That is a short vector is stored in memory as if it were stored with a single ``STR`` of the entire register; a short vector is loaded from memory using the corresponding ``LDR`` instruction. On a little-endian system this means that element 0 will always contain the lowest addressed element of a short vector; on a big-endian system element 0 will contain the highest-addressed element of a short vector.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
106
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
107 -- Procedure Call Standard for the ARM 64-bit Architecture (AArch64), 4.1.2 Short Vectors
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
108
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
109 The use of ``LDR`` and ``STR`` as the ABI defines has at least one advantage over ``LD1`` and ``ST1``. ``LDR`` and ``STR`` are oblivious to the size of the individual lanes of a vector. ``LD1`` and ``ST1`` are not - the lane size is encoded within them. This is important across an ABI boundary, because it would become necessary to know the lane width the callee expects. Consider the following code:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
110
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
111 .. code-block:: c
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
112
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
113 <callee.c>
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
114 void callee(uint32x2_t v) {
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
115 ...
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
116 }
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
117
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
118 <caller.c>
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
119 extern void callee(uint32x2_t);
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
120 void caller() {
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
121 callee(...);
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
122 }
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
123
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
124 If ``callee`` changed its signature to ``uint16x4_t``, which is equivalent in register content, if we passed as ``LD1`` we'd break this code until ``caller`` was updated and recompiled.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
125
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
126 There is an argument that if the signatures of the two functions are different then the behaviour should be undefined. But there may be functions that are agnostic to the lane layout of the vector, and treating the vector as an opaque value (just loading it and storing it) would be impossible without a common format across ABI boundaries.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
127
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
128 So to preserve ABI compatibility, we need to use the ``LDR`` lane layout across function calls.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
129
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
130 Alignment
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
131 ---------
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
132
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
133 In strict alignment mode, ``LDR qX`` requires its address to be 128-bit aligned, whereas ``LD1`` only requires it to be as aligned as the lane size. If we canonicalised on using ``LDR``, we'd still need to use ``LD1`` in some places to avoid alignment faults (the result of the ``LD1`` would then need to be reversed with ``REV``).
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
134
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
135 Most operating systems however do not run with alignment faults enabled, so this is often not an issue.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
136
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
137 Summary
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
138 -------
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
139
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
140 The following table summarises the instructions that are required to be emitted for each property mentioned above for each of the two solutions.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
141
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
142 +-------------------------------+-------------------------------+---------------------+
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
143 | | ``LDR`` layout | ``LD1`` layout |
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
144 +===============================+===============================+=====================+
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
145 | Lane ordering | ``LDR + REV`` | ``LD1`` |
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
146 +-------------------------------+-------------------------------+---------------------+
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
147 | AAPCS | ``LDR`` | ``LD1 + REV`` |
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
148 +-------------------------------+-------------------------------+---------------------+
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
149 | Alignment for strict mode | ``LDR`` / ``LD1 + REV`` | ``LD1`` |
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
150 +-------------------------------+-------------------------------+---------------------+
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
151
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
152 Neither approach is perfect, and choosing one boils down to choosing the lesser of two evils. The issue with lane ordering, it was decided, would have to change target-agnostic compiler passes and would result in a strange IR in which lane indices were reversed. It was decided that this was worse than the changes that would have to be made to support ``LD1``, so ``LD1`` was chosen as the canonical vector load instruction (and by inference, ``ST1`` for vector stores).
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
153
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
154 Implementation
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
155 ==============
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
156
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
157 There are 3 parts to the implementation:
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
158
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
159 1. Predicate ``LDR`` and ``STR`` instructions so that they are never allowed to be selected to generate vector loads and stores. The exception is one-lane vectors [1]_ - these by definition cannot have lane ordering problems so are fine to use ``LDR``/``STR``.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
160
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
161 2. Create code generation patterns for bitconverts that create ``REV`` instructions.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
162
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
163 3. Make sure appropriate bitconverts are created so that vector values get passed over call boundaries as 1-element vectors (which is the same as if they were loaded with ``LDR``).
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
164
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
165 Bitconverts
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
166 -----------
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
167
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
168 .. image:: ARM-BE-bitcastfail.png
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
169 :align: right
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
170
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
171 The main problem with the ``LD1`` solution is dealing with bitconverts (or bitcasts, or reinterpret casts). These are pseudo instructions that only change the compiler's interpretation of data, not the underlying data itself. A requirement is that if data is loaded and then saved again (called a "round trip"), the memory contents should be the same after the store as before the load. If a vector is loaded and is then bitconverted to a different vector type before storing, the round trip will currently be broken.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
172
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
173 Take for example this code sequence::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
174
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
175 %0 = load <4 x i32> %x
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
176 %1 = bitcast <4 x i32> %0 to <2 x i64>
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
177 store <2 x i64> %1, <2 x i64>* %y
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
178
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
179 This would produce a code sequence such as that in the figure on the right. The mismatched ``LD1`` and ``ST1`` cause the stored data to differ from the loaded data.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
180
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
181 .. container:: clearer
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
182
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
183 When we see a bitcast from type ``X`` to type ``Y``, what we need to do is to change the in-register representation of the data to be *as if* it had just been loaded by a ``LD1`` of type ``Y``.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
184
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
185 .. image:: ARM-BE-bitcastsuccess.png
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
186 :align: right
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
187
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
188 Conceptually this is simple - we can insert a ``REV`` undoing the ``LD1`` of type ``X`` (converting the in-register representation to the same as if it had been loaded by ``LDR``) and then insert another ``REV`` to change the representation to be as if it had been loaded by an ``LD1`` of type ``Y``.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
189
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
190 For the previous example, this would be::
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
191
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
192 LD1 v0.4s, [x]
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
193
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
194 REV64 v0.4s, v0.4s // There is no REV128 instruction, so it must be synthesizedcd
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
195 EXT v0.16b, v0.16b, v0.16b, #8 // with a REV64 then an EXT to swap the two 64-bit elements.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
196
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
197 REV64 v0.2d, v0.2d
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
198 EXT v0.16b, v0.16b, v0.16b, #8
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
199
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
200 ST1 v0.2d, [y]
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
201
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
202 It turns out that these ``REV`` pairs can, in almost all cases, be squashed together into a single ``REV``. For the example above, a ``REV128 4s`` + ``REV128 2d`` is actually a ``REV64 4s``, as shown in the figure on the right.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
203
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
204 .. [1] One lane vectors may seem useless as a concept but they serve to distinguish between values held in general purpose registers and values held in NEON/VFP registers. For example, an ``i64`` would live in an ``x`` register, but ``<1 x i64>`` would live in a ``d`` register.
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
205
|
