[Backend] Bump to llvm/llvm-project@a992f294 (#8919)

This pulls in the following fixes:

* llvm/llvm-project#168801
* llvm/llvm-project#168805
* llvm/llvm-project#169851

Along the way, we need to adjust `TMEMLoadOpPattern` and
`RelayoutTritonGPU` due to

* llvm/llvm-project#169606

Author: antiagainst

cmake/llvm-hash.txt

@@ -1 +1 @@
-ebf5d9ef7de29b55fd9e9d504f83689b4013e0de
+a992f29451b9e140424f35ac5e20177db4afbdc0

lib/Conversion/TritonToTritonGPU/RelayoutTritonGPU.cpp

@@ -34,14 +34,20 @@ struct TMEMLoadOpPattern : public OpConversionPattern<ttng::TMEMLoadOp> {
   LogicalResult
   matchAndRewrite(ttng::TMEMLoadOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
+    Type resultType = getTypeConverter()->convertType(op.getType());
     RankedTensorType type = getTMEMTensorLayout(
         typeConverter, op.getType(), op.getSrc().getType(), lookupNumWarps(op));
     rewriter.modifyOpInPlace(op, [&] { op.getResult().setType(type); });
-    Type resultType = getTypeConverter()->convertType(op.getType());
+    if (type == resultType)
+      return success();
+
     rewriter.setInsertionPointAfter(op);
     auto cvt = ConvertLayoutOp::create(rewriter, op.getLoc(), resultType,
                                        op.getResult());
-    rewriter.replaceAllUsesExcept(op.getResult(), cvt, cvt);
+    // Bypass the rewriter to avoid issues with the conversion framework's
+    // tracking of conditional replacements.
+    // See https://github.com/llvm/llvm-project/commit/504b50789602
+    op.getResult().replaceAllUsesExcept(cvt, cvt);
     return success();
   }
 };
@@ -115,7 +121,10 @@ class RelayoutTritonGPU
         // clang-format on
         >(typeConverter, context);
 
-    if (failed(applyPartialConversion(mod, target, std::move(patterns))))
+    ConversionConfig config;
+    config.allowPatternRollback = false;
+    if (failed(
+            applyPartialConversion(mod, target, std::move(patterns), config)))
       return signalPassFailure();
   }
 };

third_party/amd/backend/compiler.py

@@ -451,11 +451,6 @@ def make_amdgcn(src, metadata, options):
                             dump_file_id)
         amdgcn = llvm.translate_to_asm(src, amd.TARGET_TRIPLE, options.arch, features, flags, options.enable_fp_fusion,
                                        False)
-        # TODO: Remove the following workaround once LLVM is bumped to include: https://github.com/llvm/llvm-project/pull/169851
-        # Workaround for LLVM ERROR: cannot evaluate equated symbol 'amdgcn.device.init.num_named_barrier'
-        if knobs.compilation.enable_asan and 'gfx1250' not in options.arch:
-            amdgcn = amdgcn.replace('.amdgpu_metadata',
-                                    '\t.set\tamdgcn.device.init.num_named_barrier, 0\n.amdgpu_metadata')
         if knobs.amd.dump_amdgcn:
             print("// -----// AMDGCN Dump //----- //")
             print(amdgcn)