[Membar] Fix non-trivial function smem offsets

Codex rightly identified that we were not considering the offsets of
functions in our membar analysis at #9318 (comment)

Codex then went on and fixed it and added a regression test.

Author: lezcano

include/triton/Analysis/Membar.h

@@ -55,6 +55,16 @@ struct AllocationSlice {
 
   Allocation::BufferId getBufferId() const { return bufferId; }
 
+  AllocationSlice translated(size_t offset,
+                             bool invalidateBufferId = false) const {
+    AllocationSlice shifted = *this;
+    shifted.allocationInterval = Interval<size_t>(
+        allocationInterval.start() + offset, allocationInterval.end() + offset);
+    if (invalidateBufferId)
+      shifted.bufferId = Allocation::InvalidBufferId;
+    return shifted;
+  }
+
   void print(raw_ostream &os) const;
 
 private:
@@ -167,6 +177,26 @@ struct BlockInfo {
   }
 };
 
+inline BlockInfo translateBlockInfoToCallsite(const BlockInfo &calleeBlockInfo,
+                                              size_t callOffset) {
+  BlockInfo translatedBlockInfo;
+  auto translateSlices = [&](const BlockInfo::SliceMapT &srcSlices,
+                             BlockInfo::SliceMapT &dstSlices) {
+    for (const auto &[slice, ops] : srcSlices) {
+      auto translatedSlice =
+          slice.translated(callOffset, /*invalidateBufferId=*/true);
+      auto &dstOps = dstSlices[translatedSlice];
+      dstOps.insert(ops.begin(), ops.end());
+    }
+  };
+
+  translateSlices(calleeBlockInfo.syncReadSlices,
+                  translatedBlockInfo.syncReadSlices);
+  translateSlices(calleeBlockInfo.syncWriteSlices,
+                  translatedBlockInfo.syncWriteSlices);
+  return translatedBlockInfo;
+}
+
 //===----------------------------------------------------------------------===//
 // Shared Memory Barrier Analysis
 //===----------------------------------------------------------------------===//

lib/Analysis/Membar.cpp

@@ -279,8 +279,14 @@ void MembarAnalysis::update(Operation *op, BlockInfo *blockInfo,
     // Inter-function dependencies
     auto callOpInterface = dyn_cast<CallOpInterface>(op);
     if (auto callee =
-            dyn_cast<FunctionOpInterface>(callOpInterface.resolveCallable()))
-      curBlockInfo = funcBlockInfoMap->lookup(callee);
+            dyn_cast<FunctionOpInterface>(callOpInterface.resolveCallable())) {
+      auto calleeBlockInfo = funcBlockInfoMap->lookup(callee);
+      auto callBufferId = allocation->getBufferId(op);
+      size_t callOffset = 0;
+      if (callBufferId != Allocation::InvalidBufferId)
+        callOffset = allocation->getAllocatedInterval(callBufferId).start();
+      curBlockInfo = translateBlockInfoToCallsite(calleeBlockInfo, callOffset);
+    }
   } else {
     // Intra-function dependencies
     if (auto memoryEffectOpInterface = dyn_cast<MemoryEffectOpInterface>(op)) {

lib/Dialect/TritonNvidiaGPU/Transforms/ClusterBarrierInsertion.cpp

@@ -92,8 +92,14 @@ void ClusterBarrierAnalysis::update(Operation *op, BlockInfo *blockInfo,
   if (isa<triton::CallOp>(op)) {
     auto callOpInterface = dyn_cast<CallOpInterface>(op);
     if (auto callee =
-            dyn_cast<FunctionOpInterface>(callOpInterface.resolveCallable()))
-      curBlockInfo = funcBlockInfoMap->lookup(callee);
+            dyn_cast<FunctionOpInterface>(callOpInterface.resolveCallable())) {
+      auto calleeBlockInfo = funcBlockInfoMap->lookup(callee);
+      auto callBufferId = allocation->getBufferId(op);
+      size_t callOffset = 0;
+      if (callBufferId != Allocation::InvalidBufferId)
+        callOffset = allocation->getAllocatedInterval(callBufferId).start();
+      curBlockInfo = translateBlockInfoToCallsite(calleeBlockInfo, callOffset);
+    }
   } else {
     if (auto memEffects = dyn_cast<MemoryEffectOpInterface>(op)) {
       SmallVector<SideEffects::EffectInstance<MemoryEffects::Effect>>

test/Analysis/test-membar.mlir

@@ -1232,3 +1232,42 @@ module attributes {ttg.target = "cuda:90", "ttg.num-warps" = 8 : i32} {
     tt.return
   }
 }
+
+// -----
+
+#blockedLarge = #ttg.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
+#sharedLarge = #ttg.swizzled_shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [0]}>
+#blockedCallSrc = #ttg.blocked<{sizePerThread = [1, 32], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1]}>
+#mmaCall = #ttg.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [4, 1], instrShape = [16, 8]}>
+#blockedCallDst = #ttg.dot_op<{opIdx = 0, parent = #mmaCall, kWidth = 2}>
+#smem = #ttg.shared_memory
+
+module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32} {
+  tt.func private @callee_call_offset_membar() -> tensor<128x32xf16, #blockedCallDst> {
+    %cst = arith.constant dense<0.0> : tensor<128x32xf16, #blockedCallSrc>
+    %cvt = ttg.convert_layout %cst : tensor<128x32xf16, #blockedCallSrc> -> tensor<128x32xf16, #blockedCallDst>
+    tt.return %cvt : tensor<128x32xf16, #blockedCallDst>
+  }
+
+  // The call's virtual buffer is offset by the large allocation. The
+  // subsequent scratch op should alias at the same offset and require a membar.
+  // CHECK-LABEL: @caller_call_offset_membar
+  // CHECK: tt.call @callee_call_offset_membar{{.*}}allocation.offset = [[CALL_OFFSET:[1-9][0-9]*]]
+  // CHECK: ttg.barrier local
+  // CHECK-NEXT: ttg.convert_layout{{.*}}allocation.offset = [[CALL_OFFSET]]
+  tt.func @caller_call_offset_membar() -> tensor<128x32xf16, #blockedCallDst> {
+    %large = arith.constant dense<0> : tensor<65536xi8, #blockedLarge>
+    %buf = ttg.local_alloc : () -> !ttg.memdesc<65536xi8, #sharedLarge, #smem, mutable>
+    ttg.local_store %large, %buf : tensor<65536xi8, #blockedLarge> -> !ttg.memdesc<65536xi8, #sharedLarge, #smem, mutable>
+
+    %call = tt.call @callee_call_offset_membar() : () -> tensor<128x32xf16, #blockedCallDst>
+
+    %cst = arith.constant dense<0.0> : tensor<128x32xf16, #blockedCallSrc>
+    %cvt = ttg.convert_layout %cst : tensor<128x32xf16, #blockedCallSrc> -> tensor<128x32xf16, #blockedCallDst>
+    %sum = arith.addf %call, %cvt : tensor<128x32xf16, #blockedCallDst>
+
+    %ld = ttg.local_load %buf : !ttg.memdesc<65536xi8, #sharedLarge, #smem, mutable> -> tensor<65536xi8, #blockedLarge>
+    ttg.local_dealloc %buf : !ttg.memdesc<65536xi8, #sharedLarge, #smem, mutable>
+    tt.return %sum : tensor<128x32xf16, #blockedCallDst>
+  }
+}