New shader translator implementation (#654)
* Start implementing a new shader translator
* Fix shift instructions and a typo
* Small refactoring on StructuredProgram, move RemovePhis method to a separate class
* Initial geometry shader support
* Implement TLD4
* Fix -- There's no negation on FMUL32I
* Add constant folding and algebraic simplification optimizations, nits
* Some leftovers from constant folding
* Avoid cast for constant assignments
* Add a branch elimination pass, and misc small fixes
* Remove redundant branches, add expression propagation and other improvements on the code
* Small leftovers -- add missing break and continue, remove unused properties, other improvements
* Add null check to handle empty block cases on block visitor
* Add HADD2 and HMUL2 half float shader instructions
* Optimize pack/unpack sequences, some fixes related to half float instructions
* Add TXQ, TLD, TLDS and TLD4S shader texture instructions, and some support for bindless textures, some refactoring on codegen
* Fix copy paste mistake that caused RZ to be ignored on the AST instruction
* Add workaround for conditional exit, and fix half float instruction with constant buffer
* Add missing 0.0 source for TLDS.LZ variants
* Simplify the switch for TLDS.LZ
* Texture instructions related fixes
* Implement the HFMA instruction, and some misc. fixes
* Enable constant folding on UnpackHalf2x16 instructions
* Refactor HFMA to use OpCode* for opcode decoding rather than on the helper methods
* Remove the old shader translator
* Remove ShaderDeclInfo and other unused things
* Add dual vertex shader support
* Add ShaderConfig, used to pass shader type and maximum cbuffer size
* Move and rename some instruction enums
* Move texture instructions into a separate file
* Move operand GetExpression and locals management to OperandManager
* Optimize opcode decoding using a simple list and binary search
* Add missing condition for do-while on goto elimination
* Misc. fixes on texture instructions
* Simplify TLDS switch
* Address PR feedback, and a nit
2019-04-17 16:57:08 -07:00
|
|
|
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
|
|
|
|
using System.Collections.Generic;
|
|
|
|
|
|
|
|
namespace Ryujinx.Graphics.Shader.Translation
|
|
|
|
{
|
|
|
|
static class Dominance
|
|
|
|
{
|
2019-07-01 19:39:22 -07:00
|
|
|
// Those methods are an implementation of the algorithms on "A Simple, Fast Dominance Algorithm".
|
|
|
|
// https://www.cs.rice.edu/~keith/EMBED/dom.pdf
|
New shader translator implementation (#654)
* Start implementing a new shader translator
* Fix shift instructions and a typo
* Small refactoring on StructuredProgram, move RemovePhis method to a separate class
* Initial geometry shader support
* Implement TLD4
* Fix -- There's no negation on FMUL32I
* Add constant folding and algebraic simplification optimizations, nits
* Some leftovers from constant folding
* Avoid cast for constant assignments
* Add a branch elimination pass, and misc small fixes
* Remove redundant branches, add expression propagation and other improvements on the code
* Small leftovers -- add missing break and continue, remove unused properties, other improvements
* Add null check to handle empty block cases on block visitor
* Add HADD2 and HMUL2 half float shader instructions
* Optimize pack/unpack sequences, some fixes related to half float instructions
* Add TXQ, TLD, TLDS and TLD4S shader texture instructions, and some support for bindless textures, some refactoring on codegen
* Fix copy paste mistake that caused RZ to be ignored on the AST instruction
* Add workaround for conditional exit, and fix half float instruction with constant buffer
* Add missing 0.0 source for TLDS.LZ variants
* Simplify the switch for TLDS.LZ
* Texture instructions related fixes
* Implement the HFMA instruction, and some misc. fixes
* Enable constant folding on UnpackHalf2x16 instructions
* Refactor HFMA to use OpCode* for opcode decoding rather than on the helper methods
* Remove the old shader translator
* Remove ShaderDeclInfo and other unused things
* Add dual vertex shader support
* Add ShaderConfig, used to pass shader type and maximum cbuffer size
* Move and rename some instruction enums
* Move texture instructions into a separate file
* Move operand GetExpression and locals management to OperandManager
* Optimize opcode decoding using a simple list and binary search
* Add missing condition for do-while on goto elimination
* Misc. fixes on texture instructions
* Simplify TLDS switch
* Address PR feedback, and a nit
2019-04-17 16:57:08 -07:00
|
|
|
public static void FindDominators(BasicBlock entry, int blocksCount)
|
|
|
|
{
|
|
|
|
HashSet<BasicBlock> visited = new HashSet<BasicBlock>();
|
|
|
|
|
|
|
|
Stack<BasicBlock> blockStack = new Stack<BasicBlock>();
|
|
|
|
|
|
|
|
List<BasicBlock> postOrderBlocks = new List<BasicBlock>(blocksCount);
|
|
|
|
|
|
|
|
int[] postOrderMap = new int[blocksCount];
|
|
|
|
|
|
|
|
visited.Add(entry);
|
|
|
|
|
|
|
|
blockStack.Push(entry);
|
|
|
|
|
|
|
|
while (blockStack.TryPop(out BasicBlock block))
|
|
|
|
{
|
|
|
|
if (block.Next != null && visited.Add(block.Next))
|
|
|
|
{
|
|
|
|
blockStack.Push(block);
|
|
|
|
blockStack.Push(block.Next);
|
|
|
|
}
|
|
|
|
else if (block.Branch != null && visited.Add(block.Branch))
|
|
|
|
{
|
|
|
|
blockStack.Push(block);
|
|
|
|
blockStack.Push(block.Branch);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
postOrderMap[block.Index] = postOrderBlocks.Count;
|
|
|
|
|
|
|
|
postOrderBlocks.Add(block);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
BasicBlock Intersect(BasicBlock block1, BasicBlock block2)
|
|
|
|
{
|
|
|
|
while (block1 != block2)
|
|
|
|
{
|
|
|
|
while (postOrderMap[block1.Index] < postOrderMap[block2.Index])
|
|
|
|
{
|
|
|
|
block1 = block1.ImmediateDominator;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (postOrderMap[block2.Index] < postOrderMap[block1.Index])
|
|
|
|
{
|
|
|
|
block2 = block2.ImmediateDominator;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return block1;
|
|
|
|
}
|
|
|
|
|
|
|
|
entry.ImmediateDominator = entry;
|
|
|
|
|
|
|
|
bool modified;
|
|
|
|
|
|
|
|
do
|
|
|
|
{
|
|
|
|
modified = false;
|
|
|
|
|
|
|
|
for (int blkIndex = postOrderBlocks.Count - 2; blkIndex >= 0; blkIndex--)
|
|
|
|
{
|
|
|
|
BasicBlock block = postOrderBlocks[blkIndex];
|
|
|
|
|
|
|
|
BasicBlock newIDom = null;
|
|
|
|
|
|
|
|
foreach (BasicBlock predecessor in block.Predecessors)
|
|
|
|
{
|
|
|
|
if (predecessor.ImmediateDominator != null)
|
|
|
|
{
|
|
|
|
if (newIDom != null)
|
|
|
|
{
|
|
|
|
newIDom = Intersect(predecessor, newIDom);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
newIDom = predecessor;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (block.ImmediateDominator != newIDom)
|
|
|
|
{
|
|
|
|
block.ImmediateDominator = newIDom;
|
|
|
|
|
|
|
|
modified = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
while (modified);
|
|
|
|
}
|
|
|
|
|
|
|
|
public static void FindDominanceFrontiers(BasicBlock[] blocks)
|
|
|
|
{
|
|
|
|
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
|
|
|
|
{
|
|
|
|
BasicBlock block = blocks[blkIndex];
|
|
|
|
|
|
|
|
if (block.Predecessors.Count < 2)
|
|
|
|
{
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int pBlkIndex = 0; pBlkIndex < block.Predecessors.Count; pBlkIndex++)
|
|
|
|
{
|
|
|
|
BasicBlock current = block.Predecessors[pBlkIndex];
|
|
|
|
|
|
|
|
while (current != block.ImmediateDominator)
|
|
|
|
{
|
|
|
|
current.DominanceFrontiers.Add(block);
|
|
|
|
|
|
|
|
current = current.ImmediateDominator;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|