[Backend] bump to llvm/llvm-project@1cec5fffd8fd (#6458)

Pick up the following changes:
* llvm/llvm-project#135243
* llvm/llvm-project#135131

Author: makslevental

cmake/llvm-hash.txt

@@ -1 +1 @@
-adba14acea99cc6a17d837763a3248c9d4a2fadf
+1cec5fffd8fddd9d85b516f876093b0e3f0eec5f

third_party/amd/lib/TritonAMDGPUToLLVM/LoadStoreOpToLLVM.cpp

@@ -366,7 +366,7 @@ struct LoadOpConversion : public ConvertOpToLLVMPattern<triton::LoadOp>,
 
     auto cacheMod = op.getCache();
     SmallVector<Value> loadedVals;
-    Type vecTy = LLVM::getFixedVectorType(valueElemTy, vec);
+    Type vecTy = LLVM::getVectorType(valueElemTy, vec);
     for (size_t vecStart = 0; vecStart < numElems; vecStart += vec) {
       const size_t maxWordWidth = std::max<size_t>(32, valueElemNBits);
       const size_t totalWidth = valueElemNBits * vec;
@@ -466,7 +466,7 @@ struct BufferLoadOpConversion
     // Create the resource descriptor and then emit the buffer_load intrinsic(s)
     Value rsrcDesc = bufferEmitter.createResourceDescriptor(llPtr, llStride);
     SmallVector<Value> loadedVals;
-    Type vecTy = LLVM::getFixedVectorType(valueElemTy, vec);
+    Type vecTy = LLVM::getVectorType(valueElemTy, vec);
     for (size_t vecStart = 0; vecStart < numElems; vecStart += vec) {
       Value pred = mask ? maskElems[vecStart] : b.int_val(1, 1);
       Value falseVal = createZeroVector(rewriter, loc, cast<VectorType>(vecTy));
@@ -818,7 +818,7 @@ struct StoreOpConversion : public ConvertOpToLLVMPattern<triton::StoreOp>,
       Value pred =
           llMask ? b.and_(threadPred, maskElems[vecStart]) : threadPred;
 
-      auto vecTy = LLVM::getFixedVectorType(valueElemTy, vec);
+      auto vecTy = LLVM::getVectorType(valueElemTy, vec);
 
       const size_t maxWordWidth = std::max<size_t>(32, valueElemNBits);
       const size_t totalWidth = valueElemNBits * vec;
@@ -1038,7 +1038,7 @@ struct BufferAtomicRMWOpConversion
       Value pred =
           llMask ? b.and_(threadPred, maskElems[vecStart]) : threadPred;
 
-      Type vecTy = LLVM::getFixedVectorType(valueElemTy, vec);
+      Type vecTy = LLVM::getVectorType(valueElemTy, vec);
       Value falseVal = createZeroVector(rewriter, loc, cast<VectorType>(vecTy));
       // Create the store val
       Value storeVal = packElementRangeIntoVector(
@@ -1148,7 +1148,7 @@ struct BufferStoreOpConversion
       Value pred =
           llMask ? b.and_(threadPred, maskElems[vecStart]) : threadPred;
 
-      Type vecTy = LLVM::getFixedVectorType(valueElemTy, vec);
+      Type vecTy = LLVM::getVectorType(valueElemTy, vec);
       // Create the store val
       Value storeVal = packElementRangeIntoVector(
           rewriter, this->getTypeConverter(), loc, cast<VectorType>(vecTy),

third_party/amd/lib/TritonAMDGPUToLLVM/Utility.cpp

@@ -49,7 +49,7 @@ std::string mangleFunc(std::string name, Type type) {
 Value createVectorMaskFromPredicate(RewriterBase &rewriter, Location loc,
                                     Value pred, int64_t vecSize) {
   auto b = TritonLLVMOpBuilder(loc, rewriter);
-  auto vecMaskTy = LLVM::getFixedVectorType(rewriter.getI1Type(), vecSize);
+  auto vecMaskTy = LLVM::getVectorType(rewriter.getI1Type(), vecSize);
   Value maskVal = b.undef(vecMaskTy);
   for (size_t s = 0; s < vecSize; ++s) {
     Value indexVal =
@@ -70,7 +70,7 @@ int64_t getNumElements(Type ty) {
 Type castToVectorType(Type ty) {
   if (isa<VectorType>(ty))
     return ty;
-  return LLVM::getFixedVectorType(ty, 1);
+  return LLVM::getVectorType(ty, 1);
 }
 
 } // namespace

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/LoadStoreOpToLLVM.cpp

@@ -305,7 +305,7 @@ struct LoadOpConversion : public ConvertOpToLLVMPattern<triton::LoadOp>,
 
           size_t size = width / valueElemNBits;
 
-          auto vecTy = LLVM::getFixedVectorType(valueElemTy, size);
+          auto vecTy = LLVM::getVectorType(valueElemTy, size);
           Value v = b.undef(vecTy);
           for (size_t s = 0; s < size; ++s) {
             Value falseVal = otherElems[vecStart + ii * size + s];
@@ -376,8 +376,8 @@ struct LoadOpConversion : public ConvertOpToLLVMPattern<triton::LoadOp>,
         } else {
           curr = ret;
         }
-        curr = b.bitcast(curr, LLVM::getFixedVectorType(
-                                   valueElemTy, width / valueElemNBits));
+        curr = b.bitcast(
+            curr, LLVM::getVectorType(valueElemTy, width / valueElemNBits));
         rets.push_back(curr);
       }
       int tmp = width / valueElemNBits;

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TargetInfo.cpp

@@ -128,7 +128,8 @@ Value TargetInfo::ballot(RewriterBase &rewriter, Location loc, Type type,
                          Value cmp) const {
   auto b = TritonLLVMOpBuilder(loc, rewriter);
   Value threadMask = b.int_val(type.getIntOrFloatBitWidth(), -1);
-  return rewriter.create<NVVM::VoteBallotOp>(loc, type, threadMask, cmp);
+  return rewriter.create<NVVM::VoteSyncOp>(loc, type, threadMask, cmp,
+                                           NVVM::VoteSyncKind::ballot);
 }
 
 static Value mapa(RewriterBase &rewriter, Location loc, Value ptr, Value ctaid,