November upstream merge

[dhernandez0]

Motivation

Upstream merge pointing to this commit: ROCm/llvm-project@73f26e2

Technical Details

Changes in upstream merge:

• new amdgpu dialect wmma and mfma op format
• tosa changes, figured out a way to avoid running tosa-validate pass
• CopyInterface removed
• changes to buffer-results-to-out-params pass, I had to cherry-pick changes from [MLIR][BufferResultsToOutParamsPass] Add Option to Modify Public Function's Signature llvm/llvm-project#167248

External LIT Tests

• check-llvm
• check-mlir

Jenkins Internal CI

• Weekly (parameterSweeps + Tuning)
  • Link:
• Nightly CI
  • Link:
• PR CI

MIGraphX CI

• Update MIGraphX SHA and run through their CI
  • Results here: [ignore this PR] Bump rocmlir commit hash to test November upstream merge AMDMIGraphX#4437

Performance

• Compare tuning runtime on CI with one of the weekly runs to see if tuning time has regressed or not.

Navi2X

• parameterSweeps (manual running)
• Nightly E2E tests (w/ fixed data)
• Nightly E2E tests (w/ random data)
• PR E2E tests

List of files that are different

Diff files for review

Open these files in notepad++ or any DiffViewer with "diff" language
diff -rup upstream/llvm-project/llvm rocMLIR/external/llvm-project/llvm &> llvm_diffs.txt
llvm_diffs.txt

diff -rup upstream/llvm-project/mlir rocMLIR/external/llvm-project/mlir &> mlir_diffs.txt
mlir_diffs.txt

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[umangyadav]

diff -rup llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp rocMLIR-2122/external/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
--- llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp	2025-11-11 11:15:33.302593110 +0100
+++ rocMLIR-2122/external/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp	2025-11-11 13:14:15.569060089 +0100
@@ -66,11 +66,12 @@ static cl::opt<unsigned> PromoteAllocaTo
     cl::desc("Maximum byte size to consider promote alloca to vector"),
     cl::init(0));
 
+// TODO(rocmlir): temporary fix for backend bug
 static cl::opt<unsigned> PromoteAllocaToVectorMaxRegs(
     "amdgpu-promote-alloca-to-vector-max-regs",
     cl::desc(
         "Maximum vector size (in 32b registers) to use when promoting alloca"),
-    cl::init(32));
+    cl::init(16));
 
 // Use up to 1/4 of available register budget for vectorization.
 // FIXME: Increase the limit for whole function budgets? Perhaps x2?
Only in llvm-project/llvm/test/CodeGen/AMDGPU: amdgpu-num-agpr.ll
Only in llvm-project/llvm/test/CodeGen/AMDGPU: min-waves-per-eu-not-respected.ll
diff -rup llvm-project/llvm/test/CodeGen/AMDGPU/promote-alloca-max-regs.ll rocMLIR-2122/external/llvm-project/llvm/test/CodeGen/AMDGPU/promote-alloca-max-regs.ll

This TODO is this not fixed by upstream yet ?

Only in llvm-project/llvm/test/CodeGen/AMDGPU: amdgpu-num-agpr.ll

Why does amdgpu-num-agpr.ll not exist in rocMLIR

[umangyadav]

--- llvm-project/mlir/include/mlir/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.h	2025-02-27 12:36:13.465135726 +0100
+++ rocMLIR-2122/external/llvm-project/mlir/include/mlir/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.h	2025-11-11 11:14:16.535541433 +0100
@@ -27,7 +27,8 @@ class Pass;
 /// populateAMDGPUMemorySpaceAttributeConversions().
 void populateAMDGPUToROCDLConversionPatterns(LLVMTypeConverter &converter,
                                              RewritePatternSet &patterns,
-                                             amdgpu::Chipset chipset);
+                                             amdgpu::Chipset chipset,
+                                             bool hackForDirectToLDS);
 
 /// Remap AMDGPU memory spaces to LLVM address spaces
 /// by mapping amdgpu::AddressSpace::fat_raw_buffer to ptr addrspace(7),

Do we still require hack ?

-  module.walk([&](func::CallOp op) {
-    auto callee = symtab.lookup<func::FuncOp>(op.getCallee());
-    if (!callee) {
-      op.emitError() << "cannot find callee '" << op.getCallee() << "' in "
-                     << "symbol table";
-      didFail = true;
-      return;
-    }
-    if (!options.filterFn(&callee))
-      return;
-    if (callee.isExternal() || callee.isPublic())
-      return;
+  module.walk([&](CallOpInterface op) {
+    // TODO(rocmlir) CallOpInterface does not have a getCallee() method.
+    // We should enable this back once we add this method.
+    //
+    // auto callee = symtab.lookup<func::FuncOp>(op.getCallee());
+    // if (!callee) {
+    //   op.emitError() << "cannot find callee '" << op.getCallee() << "' in "
+    //                  << "symbol table";
+    //   didFail = true;
+    //   return;
+    // }
+    // if (!options.filterFn(&callee))
+    //   return;
+    // if (callee.isPublic() && !options.modifyPublicFunctions)
+    //   return;
+    // if (callee.isExternal())
+    //   return;
+    // FIXME validate callee in the symbol table.

Is this some new change ? There is getCallableForCallee for the callOpInterface.

You may want to clean this change :

diff -rup llvm-project/mlir/test/lit.site.cfg.py.in rocMLIR-2122/external/llvm-project/mlir/test/lit.site.cfg.py.in
--- llvm-project/mlir/test/lit.site.cfg.py.in	2025-11-11 11:15:35.834562546 +0100
+++ rocMLIR-2122/external/llvm-project/mlir/test/lit.site.cfg.py.in	2025-11-11 11:14:35.595303785 +0100
@@ -82,4 +82,4 @@ import lit.llvm
 lit.llvm.initialize(lit_config, config)
 
 # Let the main config do the real work.
-lit_config.load_config(config, "@MLIR_SOURCE_DIR@/test/lit.cfg.py")
+lit_config.load_config(config, "@MLIR_SOURCE_DIR@/test/lit.cfg.py")
\ No newline at end of file

What's the motivation behind this "modifyPublicFunctions" ?

 /// Replace buffers that are returned from a function with an out parameter.
diff -rup llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td rocMLIR-2122/external/llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td
--- llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td	2025-11-11 11:16:05.246203784 +0100
+++ rocMLIR-2122/external/llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td	2025-11-11 11:14:35.447305709 +0100
@@ -258,6 +258,9 @@ def BufferResultsToOutParamsPass
               /*default=*/"false", "Hoist static allocations to call sites.">,
        Option<"hoistDynamicAllocs", "hoist-dynamic-allocs", "bool",
               /*default=*/"false", "Hoist dynamic allocations to call sites.">,
+       Option<"modifyPublicFunctions", "modify-public-functions", "bool",
+              /*default=*/"false", "Modify function signatures of public "
+              "functions.">,
   ];
   let dependentDialects = ["memref::MemRefDialect"];
 }

[umangyadav]

I see lot of build warnings. Check if you can remove them in a subsequent PR to not add more diffs in this one.

[dhernandez0]

diff -rup llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp rocMLIR-2122/external/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
--- llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp	2025-11-11 11:15:33.302593110 +0100
+++ rocMLIR-2122/external/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp	2025-11-11 13:14:15.569060089 +0100
@@ -66,11 +66,12 @@ static cl::opt<unsigned> PromoteAllocaTo
     cl::desc("Maximum byte size to consider promote alloca to vector"),
     cl::init(0));
 
+// TODO(rocmlir): temporary fix for backend bug
 static cl::opt<unsigned> PromoteAllocaToVectorMaxRegs(
     "amdgpu-promote-alloca-to-vector-max-regs",
     cl::desc(
         "Maximum vector size (in 32b registers) to use when promoting alloca"),
-    cl::init(32));
+    cl::init(16));
 
 // Use up to 1/4 of available register budget for vectorization.
 // FIXME: Increase the limit for whole function budgets? Perhaps x2?
Only in llvm-project/llvm/test/CodeGen/AMDGPU: amdgpu-num-agpr.ll
Only in llvm-project/llvm/test/CodeGen/AMDGPU: min-waves-per-eu-not-respected.ll
diff -rup llvm-project/llvm/test/CodeGen/AMDGPU/promote-alloca-max-regs.ll rocMLIR-2122/external/llvm-project/llvm/test/CodeGen/AMDGPU/promote-alloca-max-regs.ll

This TODO is this not fixed by upstream yet ?

Only in llvm-project/llvm/test/CodeGen/AMDGPU: amdgpu-num-agpr.ll

Why does amdgpu-num-agpr.ll not exist in rocMLIR

This is the workaround we introduced here: #2028 AFAIK the fix is not done yet. This is the ticket: https://ontrack-internal.amd.com/browse/SWDEV-559837 They are actively working no it (see the recent comments).

The file amdgpu-num-agpr.ll expects the warnings related to waves-per-eu that we remove. So, instead of fixing the test I thought it was easier to remove it. I'm not even sure why we remove the warning, we could just let it be printed as well.

[dhernandez0]

Do we still require hack ?

Yes, we do, @pabloantoniom is working on a better workaround: https://github.com/ROCm/rocMLIR-internal/issues/2062
This is the backend ticket: https://ontrack-internal.amd.com/browse/SWDEV-521121

Is this some new change ? There is getCallableForCallee for the callOpInterface.

This is not new, you can see this in develop:

rocMLIR/external/llvm-project/mlir/lib/Dialect/Bufferization/Transforms/BufferResultsToOutParams.cpp

Line 150 in 2e0be81

// TODO CallOpInterface does not have a getCallee() method.

You may want to clean this change :

Sure, I will.

What's the motivation behind this "modifyPublicFunctions" ?

They changed the behavior of the pass to avoid modifying public functions. Someone fixed that with this new param modifyPublicFunctions the day after I took the commit hash, here's the PR: llvm/llvm-project#167248
So, I'm just partially taking their PR changes, so that we don't have to update Pipelines.cpp in the future upstream merge, because we already use modifyPublicFunctions.

 /// Replace buffers that are returned from a function with an out parameter.
diff -rup llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td rocMLIR-2122/external/llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td
--- llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td	2025-11-11 11:16:05.246203784 +0100
+++ rocMLIR-2122/external/llvm-project/mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td	2025-11-11 11:14:35.447305709 +0100
@@ -258,6 +258,9 @@ def BufferResultsToOutParamsPass
               /*default=*/"false", "Hoist static allocations to call sites.">,
        Option<"hoistDynamicAllocs", "hoist-dynamic-allocs", "bool",
               /*default=*/"false", "Hoist dynamic allocations to call sites.">,
+       Option<"modifyPublicFunctions", "modify-public-functions", "bool",
+              /*default=*/"false", "Modify function signatures of public "
+              "functions.">,
   ];
   let dependentDialects = ["memref::MemRefDialect"];
 }

[umangyadav]

I'm not even sure why we remove the warning, we could just let it be printed as well.

It is disabled because otherwise when MIGraphX compiles and tunes, it will dump lot of unnecessary warnings.

[umangyadav]

Sure, I will.

Can you update diff files in this PR description after you've made all the required chanes ?