GLSL decompiler type system was broken. We converted all return values
to float except for some cases where returning we couldn't and
implicitly broke the rule of returning floats (e.g. for bools or bool
pairs).
Instead of doing this introduce class Expression that knows what type a
return value has and when a consumer wants to use the string it asks for
it with a required type, emitting a runtime error if types are
incompatible.
This has the disadvantage that there's more C++ code, but we can emit
better GLSL code that's easier to read.
Implement VOTE using Nvidia's intrinsics. Documentation about these can
be found here
https://developer.nvidia.com/reading-between-threads-shader-intrinsics
Instead of using portable ARB instructions I opted to use Nvidia
intrinsics because these are the closest we have to how Tegra X1
hardware renders.
To stub VOTE on non-Nvidia drivers (including nouveau) this commit
simulates a GPU with a warp size of one, returning what is meaningful
for the instruction being emulated:
* anyThreadNV(value) -> value
* allThreadsNV(value) -> value
* allThreadsEqualNV(value) -> true
ballotARB, also known as "uint64_t(activeThreadsNV())", emits
VOTE.ANY Rd, PT, PT;
on nouveau's compiler. This doesn't match exactly to Nvidia's code
VOTE.ALL Rd, PT, PT;
Which is emulated with activeThreadsNV() by this commit. In theory this
shouldn't really matter since .ANY, .ALL and .EQ affect the predicates
(set to PT on those cases) and not the registers.
This commit implements gl_ViewportIndex and gl_Layer in vertex and
geometry shaders. In the case it's used in a vertex shader, it requires
ARB_shader_viewport_layer_array. This extension is available on AMD and
Nvidia devices (mesa and proprietary drivers), but not available on
Intel on any platform. At the moment of writing this description I don't
know if this is a hardware limitation or a driver limitation.
In the case that ARB_shader_viewport_layer_array is not available,
writes to these registers on a vertex shader are ignored, with the
appropriate logging.
Hardware testing revealed that SSY and PBK push to a different stack,
allowing code like this:
SSY label1;
PBK label2;
SYNC;
label1: PBK;
label2: EXIT;
Instead of having a vector of unique_ptr stored in a vector and
returning star pointers to this, use shared_ptr. While changing
initialization code, move it to a separate file when possible.
This is a first step to allow code analysis and node generation beyond
the ShaderIR class.
"position" was being written but not read anywhere besides geometry
shaders, where it had the same value as gl_Position.
This commit replaces "position" with gl_Position, reducing the
complexity of our code and the emitted GLSL code.
The following code is broken on AMD's proprietary GLSL compiler:
```glsl
uint idx = ...;
vec4 values = ...;
float some_value = values[idx & 3];
```
It index the wrong components, to fix this the following pessimized code
is emitted when that bug is present:
```glsl
uint idx = ...;
vec4 values = ...;
float some_value;
if ((idx & 3) == 0) some_value = values.x;
if ((idx & 3) == 1) some_value = values.y;
if ((idx & 3) == 2) some_value = values.z;
if ((idx & 3) == 3) some_value = values.w;
```
Gets rid of the need to special-case brace handling depending on the
overload used, and makes it consistent across the board with how fmt
handles them.
Strings with compile-time deducible strings are directly forwarded to
std::string's constructor, so we don't need to worry about the
performance difference here, as it'll be identical.
In a lot of places throughout the decompiler, string concatenation via
operator+ is used quite heavily. This is usually fine, when not heavily
used, but when used extensively, can be a problem. operator+ creates an
entirely new heap allocated temporary string and given we perform
expressions like:
std::string thing = a + b + c + d;
this ends up with a lot of unnecessary temporary strings being created
and discarded, which kind of thrashes the heap more than we need to.
Given we utilize fmt in some AddLine calls, we can make this a part of
the ShaderWriter's API. We can make an overload that simply acts as a
passthrough to fmt.
This way, whenever things need to be appended to a string, the operation
can be done via a single string formatting operation instead of
discarding numerous temporary strings. This also has the benefit of
making the strings themselves look nicer and makes it easier to spot
errors in them.
Operations done before the main half float operation (like HAdd) were
managing a packed value instead of the unpacked one. Adding an unpacked
operation allows us to drop the per-operand MetaHalfArithmetic entry,
simplifying the code overall.
Temporal generally indicates a relation to time, but this is just
creating a temporary, so this isn't really an accurate name for what the
function is actually doing.
TXQ returns integer types. Shaders usually do:
R0 = TXQ(); // => int
R0 = static_cast<float>(R0);
If we don't treat it as an integer, it will cast a binary float value as
float - resulting in a corrupted number.
This was originally included because texture operations returned a vec4.
These operations now return a single float and the F4 prefix doesn't
mean anything.
Previous code relied on GLSL parameter order (something that's always
ill-formed on an IR design). This approach passes spatial coordiantes
through operation nodes and array and depth compare values in the the
texture metadata. It still contains an "extra" vector containing generic
nodes for bias and component index (for example) which is still a bit
ill-formed but it should be better than the previous approach.
Constant buffer values on the shader IR were using different offsets if
the access direct or indirect. cbuf34 has a non-multiplied offset while
cbuf36 does. On shader decoding this commit multiplies it by four on
cbuf34 queries.
Geometry shaders follow a pattern that results in out of bound reads.
This pattern is:
- VSETP to predicate
- Use that predicate to conditionally set a register a big number
- Use the register to access geometry shaders
At the time of writing this commit I don't know what's the intent of
this number. Some drivers argue about these out of bound reads. To avoid
this issue, input reads are guarded limiting reads to the highest
posible vertex input of the current topology (e.g. points to 1 and
triangles to 3).
