1New IR, or NIR, is an IR for Mesa intended to sit below GLSL IR and Mesa IR.
2Its design inherits from the various IR's that Mesa has used in the past, as
3well as Direct3D assembly, and it includes a few new ideas as well. It is a
4flat (in terms of using instructions instead of expressions), typeless IR,
5similar to TGSI and Mesa IR.  It also supports SSA (although it doesn't require
6it).
7
8Variables
9=========
10
11NIR includes support for source-level GLSL variables through a structure mostly
12copied from GLSL IR. These will be used for linking and conversion from GLSL IR
13(and later, from an AST), but for the most part, they will be lowered to
14registers (see below) and loads/stores.
15
16Registers
17=========
18
19Registers are light-weight; they consist of a structure that only contains its
20size, its index for liveness analysis, and an optional name for debugging. In
21addition, registers can be local to a function or global to the entire shader;
22the latter will be used in ARB_shader_subroutine for passing parameters and
23getting return values from subroutines. Registers can also be an array, in which
24case they can be accessed indirectly. Each ALU instruction (add, subtract, etc.)
25works directly with registers or SSA values (see below).
26
27SSA
28========
29
30Everywhere a register can be loaded/stored, an SSA value can be used instead.
31The only exception is that arrays/indirect addressing are not supported with
32SSA; although research has been done on extensions of SSA to arrays before, it's
33usually for the purpose of parallelization (which we're not interested in), and
34adds some overhead in the form of adding copies or extra arrays (which is much
35more expensive than introducing copies between non-array registers). SSA uses
36point directly to their corresponding definition, which in turn points to the
37instruction it is part of. This creates an implicit use-def chain and avoids the
38need for an external structure for each SSA register.
39
40Functions
41=========
42
43Support for function calls is mostly similar to GLSL IR. Each shader contains a
44list of functions, and each function has a list of overloads. Each overload
45contains a list of parameters, and may contain an implementation which specifies
46the variables that correspond to the parameters and return value. Inlining a
47function, assuming it has a single return point, is as simple as copying its
48instructions, registers, and local variables into the target function and then
49inserting copies to and from the new parameters as appropriate. After functions
50are inlined and any non-subroutine functions are deleted, parameters and return
51variables will be converted to global variables and then global registers. We
52don't do this lowering earlier (i.e. the fortranizer idea) for a few reasons:
53
54- If we want to do optimizations before link time, we need to have the function
55signature available during link-time.
56
57- If we do any inlining before link time, then we might wind up with the
58inlined function and the non-inlined function using the same global
59variables/registers which would preclude optimization.
60
61Intrinsics
62=========
63
64Any operation (other than function calls and textures) which touches a variable
65or is not referentially transparent is represented by an intrinsic. Intrinsics
66are similar to the idea of a "builtin function," i.e. a function declaration
67whose implementation is provided by the backend, except they are more powerful
68in the following ways:
69
70- They can also load and store registers when appropriate, which limits the
71number of variables needed in later stages of the IR while obviating the need
72for a separate load/store variable instruction.
73
74- Intrinsics can be marked as side-effect free, which permits them to be
75treated like any other instruction when it comes to optimizations. This allows
76load intrinsics to be represented as intrinsics while still being optimized
77away by dead code elimination, common subexpression elimination, etc.
78
79Intrinsics are used for:
80
81- Atomic operations
82- Memory barriers
83- Subroutine calls
84- Geometry shader emitVertex and endPrimitive
85- Loading and storing variables (before lowering)
86- Loading and storing uniforms, shader inputs and outputs, etc (after lowering)
87- Copying variables (cases where in GLSL the destination is a structure or
88array)
89- The kitchen sink
90- ...
91
92Textures
93=========
94
95Unfortunately, there are far too many texture operations to represent each one
96of them with an intrinsic, so there's a special texture instruction similar to
97the GLSL IR one. The biggest difference is that, while the texture instruction
98has a sampler dereference field used just like in GLSL IR, this gets lowered to
99a texture unit index (with a possible indirect offset) while the type
100information of the original sampler is kept around for backends. Also, all the
101non-constant sources are stored in a single array to make it easier for
102optimization passes to iterate over all the sources.
103
104Control Flow
105=========
106
107Like in GLSL IR, control flow consists of a tree of "control flow nodes", which
108include if statements and loops, and jump instructions (break, continue, and
109return). Unlike GLSL IR, though, the leaves of the tree aren't statements but
110basic blocks. Each basic block also keeps track of its successors and
111predecessors, and function implementations keep track of the beginning basic
112block (the first basic block of the function) and the ending basic block (a fake
113basic block that every return statement points to). Together, these elements
114make up the control flow graph, in this case a redundant piece of information on
115top of the control flow tree that will be used by almost all the optimizations.
116There are helper functions to add and remove control flow nodes that also update
117the control flow graph, and so usually it doesn't need to be touched by passes
118that modify control flow nodes.
119