shader_recompiler: split resource tracking and flatten load from buffer for sharp source by LNDF · Pull Request #4782 · shadps4-emu/shadPS4

LNDF · 2026-07-31T16:58:09Z

This splits resource tracking pass into resource discovery and patching passes. Then we put flattening pass in the middle.

That way:

We discover all resources first, accepting ReadConstBuffer as source and marking it for flattening.
Flattening pass flattens all ReadConsts + marked ReadConstBuffers
Patch resource instructions, also accepting ReadConstBuffer as source (since it was flattened)

LNDF · 2026-07-31T17:04:29Z

@@ -185,7 +202,7 @@ static void VisitPointer(u32 off_dw, IR::Inst* subtree, PassInfo& pass_info,
        c.mov(r10d, ptr[rdi + (src_off_dw << 2)]);


Do i need to handle buffer stride etc here if ReadConstBuffer?

brad0demx · 2026-07-31T18:24:00Z

On this PR, Call of Duty Cold War crashes in roughly the same point with this new critical
[Debug] <Critical> (shadPS4:GpuComm) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code! Unable to find sharp sources pc=0x70a8
CUSA24267.log

brad0demx · 2026-07-31T18:28:51Z

Killzone Shadow Fall still crashes while loading into offline match, now crashing with

[Debug] <Critical> (shadPS4:GpuComm) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code!
Unable to find sharp sources pc=0x1484

CUSA00008.log

bigol83 · 2026-07-31T21:49:00Z

Skyrim immediately crashes with this PR
CUSA05486.log

[Debug] <Critical> (shadPS4:GpuCommandProcessor) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code!
Unable to find sharp sources pc=0x80

StevenMiller123 · 2026-08-01T01:23:22Z

Current push fixes the Skyrim regression. Now the game behaves like main again.

LNDF · 2026-08-01T01:26:27Z

Skyrim immediately crashes with this PR CUSA05486.log

[Debug] <Critical> (shadPS4:GpuCommandProcessor) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code!
Unable to find sharp sources pc=0x80

Should be fixed

Killzone Shadow Fall still crashes while loading into offline match, now crashing with
[Debug] <Critical> (shadPS4:GpuComm) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code!
Unable to find sharp sources pc=0x1484
CUSA00008.log

Try latest commit

On this PR, Call of Duty Cold War crashes in roughly the same point with this new critical [Debug] <Critical> (shadPS4:GpuComm) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code! Unable to find sharp sources pc=0x70a8 CUSA24267.log

Try latest commit

bigol83 · 2026-08-03T01:13:43Z

Skyrim immediately crashes with this PR CUSA05486.log
[Debug] <Critical> (shadPS4:GpuCommandProcessor) resource_discover_pass.cpp:141 FindSharpSource: Unreachable code!
Unable to find sharp sources pc=0x80
Should be fixed

It is fixed for me

StevenMiller123 · 2026-08-07T20:00:59Z

This PR causes regressions in Uncharted: The Nathan Drake Collection and Bloodborne. I've reported these directly to @LNDF, just mentioning them here so people know these issues are already known.

bigol83 · 2026-08-08T00:54:36Z

Far Cry 5 shows this on this PR

[Render.Vulkan] <Info> (shadPS4:GpuCommandProcessor) vk_pipeline_cache.cpp:602 CompileModule: Compiling fs shader 0xc42b4153
[Debug] <Critical> (shadPS4:GpuCommandProcessor) breadth_first_search.h:139 DominatingBreadthFirstSearch: Assertion Failed!
Unable to deduce correct finding

Main build

[Render.Vulkan] <Info> (shadPS4:GpuCommandProcessor) vk_pipeline_cache.cpp:602 CompileModule: Compiling fs shader 0xc42b4153
[Debug] <Critical> (shadPS4:GpuCommandProcessor) resource_tracking_pass.cpp:489 TrackSharp: Assertion Failed!
Unable to deduce sharp source

crashes on both cases without even showing a logo

Final Fantasy XV crashes with

[Render.Vulkan] <Info> (shadPS4:GpuCommandProcessor) vk_pipeline_cache.cpp:602 CompileModule: Compiling cs shader 0x66f5202
[Debug] <Critical> (shadPS4:GpuCommandProcessor) ir_emitter.cpp:770 CompositeConstruct: Unreachable code!
Composite construct with 0 elements, only 2-4 are supported

Main

[Kernel.Vmm] <Info> (LuminousTask1) memory.cpp:246 sceKernelMapNamedDirectMemory: out_addr = 0x33e1e0000
[Debug] <Critical> (shadPS4:GpuCommandProcessor) resource_tracking_pass.cpp:411 FindSharpSources: Unreachable code!
Bindless sharp access detected pc=0x584

brad0demx · 2026-08-10T18:06:47Z

COD Cold War now hangs rather than crashing.

[Render.Recompiler] <Error> (shadPS4:GpuComm) flatten_extended_userdata_pass.cpp:546 ComputeOffset: Unexpected instruction for offset computation, BitCastU32F32
[Render.Recompiler] <Error> (shadPS4:GpuComm) flatten_extended_userdata_pass.cpp:589 VisitPointer: Failed to compute offset for SRT walker
[Lib.Http] <Info> (bdHTTPWorker(PSN)) http.cpp:1915 sceHttpWaitRequest: epoll id=2 returned 0 events after wait
[Lib.Http] <Info> (bdHTTPWorker(PSN)) http.cpp:1915 sceHttpWaitRequest: epoll id=2 returned 0 events after wait
[Lib.Http] <Info> (bdHTTPWorker(PSN)) http.cpp:1915 sceHttpWaitRequest: epoll id=2 returned 0 events after wait

CUSA24267.log

raphaelthegreat · 2026-08-12T09:35:45Z

+
+static bool EmitComputeOffsetIAdd32(Xbyak::CodeGenerator& c, Xbyak::Reg32 reg, PassInfo& pass_info,
+                                    IR::Inst* inst) {
+    if (inst->AreAllArgsImmediates()) {


I feel this check might not be needed as constant propagation has already run before this and should have constant folded the operation, have you noticed this in some code?

Done.

Tho min and max afaik are not constant folded so i left them

raphaelthegreat · 2026-08-12T09:50:50Z

        u32 vn;

        switch (inst->GetOpcode()) {
        case IR::Opcode::Phi: {


I feel this also suffers from the logic bug in main. If a shader has something like that

addr = ... read_const(addr) ... addr += 4 read_const(addr)

then the GVN might consider them identical since it skips over the intermediate add sees same addr. I have not confirmed this so worth checking. Another question is if we really need GVN pass, how much de-duplication actually happens in practice

For other insts that are not GetUserData, CompositeConstructU32x2, ReadConst and ReadConstBuffer GVN will associate an incrementing number to the instruction itself. For those 4 instructions it will insteed assign an incrementing number to a hash of the numbers given to the argument instructions.

Because of this, if you have

%10 = ReadConst %9 ... %11 = IAdd32 %10 ... %12 = ReadConst %11 ...

Both readconsts should have different number since the argument is different.

I figured out this by reading the gvn table .h file so i might have understood something wrong. Is there a game that is affected by this in main?

I think the example given was incorrect, what about this

addr = user_data() if (cond0) addr0 = addr + 4; else addr0 = addr; read_const(addr0); if (cond1) addr1 = addr + 8 else addr1 = addr; read_const(addr1);

Each ReadConst is gonna have a phi as address where 1 argument is the same SSA value directly from user data and another is an IAdd. The bfs is gonna return the addr SSA in both cases resulting in same value number

Done. We now handle phis as any other inst (so assugn a new gvn value for each phi)

Actualy, just updated it to use the same logic as in flattening pass

raphaelthegreat · 2026-08-12T09:52:31Z

+
+    // Perform dominance analysis on findings and eliminate ones that don't pass
+    // If a finding is dominated by another, the former can be eliminated.
+    size_t num_findings = findings.size();


I think this should be a separate function in dominance.h that takes the findings array by reference and prunes it

Yes. But is dominance.h the correct place? That is for frontend and what we are trying to separate is ir stuff.

Well this code prunes sources based on dominance analysis I feel it matches better in dominance than a what is supposed to be a generic BFS algorithm header impl

raphaelthegreat · 2026-08-12T09:55:50Z

 }

+template <typename Instruction, typename Pred>
+auto DominatingBreadthFirstSearch(Instruction* inst, const IR::Block& current_parent, bool deep,


The BFS header is getting crowded so worth doing a more holistic review.
The normal BFS

Only finds 1 instruction its looking for and stops

Keeps searching until it finds it without any way to stop it

This BFS

Can find multiple sources if presents

Is more selective where it searches and deep option stops it from going thru phis

A review of all BFS usages in recompiler could be useful to deduce if its possible to make a unified function that handles everything (also could use std::expected as predicate return to allow user to specify if they want to stop search on the current path or entirely etc)

raphaelthegreat · 2026-08-12T10:06:11Z

+    }
+
+    boost::container::small_vector<IR::Inst*, 8> visited, findings;
+    std::queue<IR::Inst*> queue;


Small optimization; since we spam this function a lot in resource discover might be useful to either pass this as argument in function so resource discover can maintain a queue or make this static and clear it on function entry, to avoid reconstructing the object on every call and associated heap allocations

raphaelthegreat · 2026-08-12T10:10:08Z

+        // For inmediates, add a sharp usage with null sharp source.
+        sharp_usages.emplace_back(ResourceDiscovery{&inst, &block, nullptr});
+    } else {
+        IR::Inst* buffer_handle = handle->Arg(0).InstRecursive();


For buffers handle is a composite that contains all 4 dwords of V#, can just call FindSharpSource for each, store them in an array (as mentioned in other review comment) and in resource patching verify offsets of flattened instructions are contiguous. If they are not, additional logic could be introduced (for example using gather intrinsic like _mm_i32gather_epi32 with a slim abstraction could be suitable)

Do you think this is in scope for this PR or out of scope?

IMO its a better idea to do this in another PR where we also include the refactor of image instructions to include both the hish and low vectors.

Sure it could be possible. You should add an assertion at least that the ReadConstBuffer insts you find have a use in the composite handle of the buffer opcode

raphaelthegreat · 2026-08-12T10:13:16Z

+
+    if (sharp_source && sharp_source->GetOpcode() == IR::Opcode::ReadConstBuffer) {
+        const auto texture_flags = inst.Flags<IR::TextureInstInfo>();
+        const auto is_r128 = texture_flags.is_r128.Value();


Mentioning here too, image sample ir op should be updated to hold 2 composites for the low and high part of the T# and do the same as buffers. The r128 check should be entirely a frontend thing and leave the high composite null

Answered in the previous comment

raphaelthegreat · 2026-08-12T10:19:05Z

+        auto inst_info = inst->Flags<IR::BufferInstInfo>();
+        return inst_info.flatbuf_off_dw;
+    }
+    return inst->Flags<u16>();


for good measure check if its IR::Opcode::ReadConst and add an UNREACHABLE at the end

raphaelthegreat · 2026-08-12T10:23:47Z

-                if (!inst.Arg(1).IsImmediate()) {
-                    LOG_WARNING(Render_Recompiler, "ReadConst has non-immediate offset");
-                    continue;
+            if (inst.GetOpcode() == IR::Opcode::ReadConst ||


condition can be inverted to reduce nesting

raphaelthegreat · 2026-08-12T10:33:57Z

+                    IR::Inst* inst{queue.front()};
+                    queue.pop();
+
+                    if (inst->GetOpcode() == IR::Opcode::ReadConstBuffer) {


Should this stop when it encounters a Phi? iirc the dynamic offset emit cannot handle them atm though not sure what the plan is to add them in the future. Also this is another BFS variant that searches the entire tree without stopping, would be good to unify it with the rest

Currently will go through all possible previous blocks. Then this will log an error when generating the x64 code.

LNDF · 2026-08-14T16:19:54Z

Regression on Uncharted The Nathan Drake Collection. It doesn't matter whether DMA is enabled or not, it crashes at the shader compilation stage. But I decided to launch it without DMA. Main Uploading Main-log! Without DMA.txt… Freezes while compiling shaders PR PR log Without DMA.txt The game doesn't start AT ALL.

Thanks for the report. This regression is already know and will be delt with before merging. There are more games thatare affected.

First will apply the suggested changes on the review though.

LNDF added 3 commits July 29, 2026 22:29

split resource tracking into 2 parts

bee1833

flatten load from buffer if used as sharp source

42b7bd2

clang-format

aac91f0

LNDF commented Jul 31, 2026

View reviewed changes

LNDF requested a review from raphaelthegreat July 31, 2026 17:04

LNDF and others added 2 commits July 31, 2026 19:31

fix inmediate sample raw

048268b

Merge branch 'shadps4-emu:main' into buffer-flattening

3800d37

LNDF added 4 commits August 1, 2026 02:34

defer sharp not found asserts

dd46168

flip asserts

a68eedd

disable aniso pattern

4efec91

fix disable aniso

645f373

some fixes

2bcc838

LNDF added 11 commits August 4, 2026 22:14

non immediate

2f527a3

fix PushPtr

b932e75

fix PushPtr for non immediate

07450be

don't assert on offset computation failures

d6bc173

robust bfs

1f34d19

Merge remote-tracking branch 'origin/main' into buffer-flattening

dc69f65

some xbyak fixes

4c83593

add bitwise ops

4eb3e36

umin umax and bit extract

5640a7e

correct rshift lshift 0st immediate

d84e79b

fix and in memory

4b096b3

some minor code fixes

97abe21

LNDF added 7 commits August 8, 2026 10:05

consider ReadConstBuffer as valid finding

926bfa9

fix xbyak

8e25815

fix magic_enum for opcodes

e8eb5d5

fix building (yes, i didn't test)

2bdc50c

don't corrupt stack when bailing out

0f861dd

ReadConstBuffer compute offset

02e8676

use canonical inst for readconst non-imm offset

6cbc39d

LNDF added 4 commits August 10, 2026 22:43

mark all readconstbuffer sharp dwords

54d7bc3

consider r128 images

1dc5c32

FindSharpLocation

cfb141c

handle non-imm offset readconstbuffer

07602af

raphaelthegreat reviewed Aug 12, 2026

View reviewed changes

Merge branch 'shadps4-emu:main' into buffer-flattening

0624f6f

This comment was marked as abuse.

Sign in to view

LNDF added 10 commits August 14, 2026 20:25

don't handle immediates when constant folding

15ab3e9

proper check for ReadConst to get flatbuf offset

b849e31

clang-format

18c62e8

fix stupidity

40a5e85

invert condition

b7bd91a

separate search dominant function

df0d6b4

handle phis loke other insts in gvn

8643ae8

use same logic for phis in gvn than for flatten pass

722e2ac

apparently idk how logic works

00a0371

revert back to old logic

630c407

		@@ -185,7 +202,7 @@ static void VisitPointer(u32 off_dw, IR::Inst* subtree, PassInfo& pass_info,
		c.mov(r10d, ptr[rdi + (src_off_dw << 2)]);

Uh oh!

Uh oh!

Conversation

LNDF commented Jul 31, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

brad0demx commented Jul 31, 2026

Uh oh!

brad0demx commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bigol83 commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

StevenMiller123 commented Aug 1, 2026

Uh oh!

LNDF commented Aug 1, 2026

Uh oh!

bigol83 commented Aug 3, 2026

Uh oh!

StevenMiller123 commented Aug 7, 2026

Uh oh!

bigol83 commented Aug 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

brad0demx commented Aug 10, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

brad0demx commented Jul 31, 2026 •

edited

Loading

bigol83 commented Jul 31, 2026 •

edited

Loading

bigol83 commented Aug 8, 2026 •

edited

Loading