[llvm] [SandboxVec][BottomUpVec] Implement packing of vectors (PR #116447)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: vporpo (vporpo)

<details>
<summary>Changes</summary>

Up until now we could only support packing of scalar elements. This patch fixes this by implementing packing of vector elements, by generating extractelement and insertelement instruction pairs.

---
Full diff: https://github.com/llvm/llvm-project/pull/116447.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp (+23-3) 
- (modified) llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll (+23) 


``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
index 4b82fea2aa2b22..a2ea11be59b8ed 100644
--- a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
+++ b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
@@ -181,12 +181,32 @@ Value *BottomUpVec::createPack(ArrayRef<Value *> ToPack) {
     // An element can be either scalar or vector. We need to generate different
     // IR for each case.
     if (Elm->getType()->isVectorTy()) {
-      llvm_unreachable("Unimplemented");
+      unsigned NumElms =
+          cast<FixedVectorType>(Elm->getType())->getNumElements();
+      for (auto ExtrLane : seq<int>(0, NumElms)) {
+        // We generate extract-insert pairs, for each lane in `Elm`.
+        Constant *ExtrLaneC =
+            ConstantInt::getSigned(Type::getInt32Ty(Ctx), ExtrLane);
+        // This may return a Constant if Elm is a Constant.
+        auto *ExtrI =
+            ExtractElementInst::create(Elm, ExtrLaneC, WhereIt, Ctx, "VPack");
+        if (!isa<Constant>(ExtrI))
+          WhereIt = std::next(cast<Instruction>(ExtrI)->getIterator());
+        Constant *InsertLaneC =
+            ConstantInt::getSigned(Type::getInt32Ty(Ctx), InsertIdx++);
+        // This may also return a Constant if ExtrI is a Constant.
+        auto *InsertI = InsertElementInst::create(
+            LastInsert, ExtrI, InsertLaneC, WhereIt, Ctx, "VPack");
+        if (!isa<Constant>(InsertI)) {
+          LastInsert = InsertI;
+          WhereIt = std::next(cast<Instruction>(LastInsert)->getIterator());
+        }
+      }
     } else {
       Constant *InsertLaneC =
           ConstantInt::getSigned(Type::getInt32Ty(Ctx), InsertIdx++);
-      // This may be folded into a Constant if LastInsert is a Constant. In that
-      // case we only collect the last constant.
+      // This may be folded into a Constant if LastInsert is a Constant. In
+      // that case we only collect the last constant.
       LastInsert = InsertElementInst::create(LastInsert, Elm, InsertLaneC,
                                              WhereIt, Ctx, "Pack");
       if (auto *NewI = dyn_cast<Instruction>(LastInsert))
diff --git a/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll b/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
index dff27d06d8ed22..7422d287ff3e2a 100644
--- a/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
+++ b/llvm/test/Transforms/SandboxVectorizer/bottomup_basic.ll
@@ -187,3 +187,26 @@ define void @cant_vectorize_seeds(ptr %ptr) {
   ret void
 }
 
+define void @pack_vectors(ptr %ptr, ptr %ptr2) {
+; CHECK-LABEL: define void @pack_vectors(
+; CHECK-SAME: ptr [[PTR:%.*]], ptr [[PTR2:%.*]]) {
+; CHECK-NEXT:    [[PTR0:%.*]] = getelementptr <2 x float>, ptr [[PTR]], i32 0
+; CHECK-NEXT:    [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 2
+; CHECK-NEXT:    [[LD0:%.*]] = load <2 x float>, ptr [[PTR0]], align 8
+; CHECK-NEXT:    [[LD1:%.*]] = load float, ptr [[PTR2]], align 4
+; CHECK-NEXT:    [[VPACK:%.*]] = extractelement <2 x float> [[LD0]], i32 0
+; CHECK-NEXT:    [[VPACK1:%.*]] = insertelement <3 x float> poison, float [[VPACK]], i32 0
+; CHECK-NEXT:    [[VPACK2:%.*]] = extractelement <2 x float> [[LD0]], i32 1
+; CHECK-NEXT:    [[VPACK3:%.*]] = insertelement <3 x float> [[VPACK1]], float [[VPACK2]], i32 1
+; CHECK-NEXT:    [[PACK:%.*]] = insertelement <3 x float> [[VPACK3]], float [[LD1]], i32 2
+; CHECK-NEXT:    store <3 x float> [[PACK]], ptr [[PTR0]], align 8
+; CHECK-NEXT:    ret void
+;
+  %ptr0 = getelementptr <2 x float>, ptr %ptr, i32 0
+  %ptr1 = getelementptr float, ptr %ptr, i32 2
+  %ld0 = load <2 x float>, ptr %ptr0
+  %ld1 = load float, ptr %ptr2
+  store <2 x float> %ld0, ptr %ptr0
+  store float %ld1, ptr %ptr1
+  ret void
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/116447