[llvm] [AMDGPU] Identify vector idiom to unlock SROA (PR #156791)

[Yaxun Liu via llvm-commits]

================
@@ -0,0 +1,317 @@
+//===- AMDGPUVectorIdiom.cpp ------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+// AMDGPU-specific vector idiom canonicalizations to unblock SROA and
+// subsequent scalarization/vectorization.
+//
+// Motivation:
+// - HIP vector types are often modeled as structs and copied with memcpy.
+//   Address-level selects on such copies block SROA. Converting to value-level
+//   operations or splitting the CFG enables SROA to break aggregates, which
+//   unlocks scalarization/vectorization on AMDGPU.
+//
+// Example pattern:
+//   %src = select i1 %c, ptr %A, ptr %B
+//   call void @llvm.memcpy(ptr %dst, ptr %src, i32 16, i1 false)
+//
+// Objectives:
+// - Canonicalize small memcpy patterns where source or destination is a select
+// of pointers.
+// - Prefer value-level selects (on loaded values) over address-level selects
+// when safe.
+// - When speculation is unsafe, split the CFG to isolate each arm.
+//
+// Assumptions:
+// - Only handles non-volatile memcpy with constant length N where 0 < N <=
+// MaxBytes (default 32).
+// - Source and destination must be in the same address space.
+// - Speculative loads are allowed only if a conservative alignment check
+// passes.
+// - No speculative stores are introduced.
+//
+// Transformations:
+// - Source-select memcpy: attempt speculative loads -> value select -> single
+// store.
+//   Fallback is CFG split with two memcpy calls.
+// - Destination-select memcpy: always CFG split to avoid speculative stores.
+//
+// Run this pass early, before SROA.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUVectorIdiom.h"
+#include "AMDGPU.h"
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AssumptionCache.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/InstIterator.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/IR/Type.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Local.h"
+
+using namespace llvm;
+using namespace PatternMatch;
+
+#define DEBUG_TYPE "amdgpu-vector-idiom"
+
+namespace {
+
+// Default to 32 bytes since the largest HIP vector types are double4 or long4.
+static cl::opt<unsigned> AMDGPUVectorIdiomMaxBytes(
+    "amdgpu-vector-idiom-max-bytes",
+    cl::desc("Max memcpy size (in bytes) to transform in AMDGPUVectorIdiom "
+             "(default 32)"),
+    cl::init(32));
----------------
yxsamliu wrote:

will do

https://github.com/llvm/llvm-project/pull/156791