[LLVM][SROA] Teach SROA how to "bitcast" between fixed and scalable vectors. (#130973)

For function whose vscale_range is limited to a single value we can size
scalable vectors. This aids SROA by allowing scalable vector load and
store operations to be considered for replacement whereby bitcasts
through memory can be replaced by vector insert or extract operations.

Author: paulwalker-arm

clang/test/CodeGen/attr-arm-sve-vector-bits-cast.c

@@ -62,10 +62,7 @@ fixed_bool_t from_svbool_t(svbool_t type) {
 
 // CHECK-LABEL: @lax_cast(
 // CHECK-NEXT:  entry:
-// CHECK-NEXT:    [[SAVED_VALUE:%.*]] = alloca <16 x i32>, align 64
-// CHECK-NEXT:    [[TYPE:%.*]] = tail call <16 x i32> @llvm.vector.extract.v16i32.nxv4i32(<vscale x 4 x i32> [[TYPE_COERCE:%.*]], i64 0)
-// CHECK-NEXT:    store <16 x i32> [[TYPE]], ptr [[SAVED_VALUE]], align 64, !tbaa [[TBAA6:![0-9]+]]
-// CHECK-NEXT:    [[TMP0:%.*]] = load <vscale x 2 x i64>, ptr [[SAVED_VALUE]], align 64, !tbaa [[TBAA6]]
+// CHECK-NEXT:    [[TMP0:%.*]] = bitcast <vscale x 4 x i32> [[TYPE_COERCE:%.*]] to <vscale x 2 x i64>
 // CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
 //
 svint64_t lax_cast(fixed_int32_t type) {
@@ -74,18 +71,18 @@ svint64_t lax_cast(fixed_int32_t type) {
 
 // CHECK-LABEL: @to_svint32_t__from_gnu_int32_t(
 // CHECK-NEXT:  entry:
-// CHECK-NEXT:    [[TYPE:%.*]] = load <16 x i32>, ptr [[TMP0:%.*]], align 16, !tbaa [[TBAA6]]
-// CHECK-NEXT:    [[CASTSCALABLESVE:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.insert.nxv4i32.v16i32(<vscale x 4 x i32> poison, <16 x i32> [[TYPE]], i64 0)
-// CHECK-NEXT:    ret <vscale x 4 x i32> [[CASTSCALABLESVE]]
+// CHECK-NEXT:    [[TYPE:%.*]] = load <16 x i32>, ptr [[TMP0:%.*]], align 16, !tbaa [[TBAA2:![0-9]+]]
+// CHECK-NEXT:    [[CAST_SCALABLE:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.insert.nxv4i32.v16i32(<vscale x 4 x i32> poison, <16 x i32> [[TYPE]], i64 0)
+// CHECK-NEXT:    ret <vscale x 4 x i32> [[CAST_SCALABLE]]
 //
 svint32_t to_svint32_t__from_gnu_int32_t(gnu_int32_t type) {
   return type;
 }
 
 // CHECK-LABEL: @from_svint32_t__to_gnu_int32_t(
 // CHECK-NEXT:  entry:
-// CHECK-NEXT:    [[CASTFIXEDSVE:%.*]] = tail call <16 x i32> @llvm.vector.extract.v16i32.nxv4i32(<vscale x 4 x i32> [[TYPE:%.*]], i64 0)
-// CHECK-NEXT:    store <16 x i32> [[CASTFIXEDSVE]], ptr [[AGG_RESULT:%.*]], align 16, !tbaa [[TBAA6]]
+// CHECK-NEXT:    [[CAST_FIXED:%.*]] = tail call <16 x i32> @llvm.vector.extract.v16i32.nxv4i32(<vscale x 4 x i32> [[TYPE:%.*]], i64 0)
+// CHECK-NEXT:    store <16 x i32> [[CAST_FIXED]], ptr [[AGG_RESULT:%.*]], align 16, !tbaa [[TBAA2]]
 // CHECK-NEXT:    ret void
 //
 gnu_int32_t from_svint32_t__to_gnu_int32_t(svint32_t type) {
@@ -94,9 +91,9 @@ gnu_int32_t from_svint32_t__to_gnu_int32_t(svint32_t type) {
 
 // CHECK-LABEL: @to_fixed_int32_t__from_gnu_int32_t(
 // CHECK-NEXT:  entry:
-// CHECK-NEXT:    [[TYPE:%.*]] = load <16 x i32>, ptr [[TMP0:%.*]], align 16, !tbaa [[TBAA6]]
-// CHECK-NEXT:    [[CASTSCALABLESVE:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.insert.nxv4i32.v16i32(<vscale x 4 x i32> poison, <16 x i32> [[TYPE]], i64 0)
-// CHECK-NEXT:    ret <vscale x 4 x i32> [[CASTSCALABLESVE]]
+// CHECK-NEXT:    [[TYPE:%.*]] = load <16 x i32>, ptr [[TMP0:%.*]], align 16, !tbaa [[TBAA2]]
+// CHECK-NEXT:    [[CAST_SCALABLE:%.*]] = tail call <vscale x 4 x i32> @llvm.vector.insert.nxv4i32.v16i32(<vscale x 4 x i32> poison, <16 x i32> [[TYPE]], i64 0)
+// CHECK-NEXT:    ret <vscale x 4 x i32> [[CAST_SCALABLE]]
 //
 fixed_int32_t to_fixed_int32_t__from_gnu_int32_t(gnu_int32_t type) {
   return type;
@@ -105,7 +102,7 @@ fixed_int32_t to_fixed_int32_t__from_gnu_int32_t(gnu_int32_t type) {
 // CHECK-LABEL: @from_fixed_int32_t__to_gnu_int32_t(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[TYPE:%.*]] = tail call <16 x i32> @llvm.vector.extract.v16i32.nxv4i32(<vscale x 4 x i32> [[TYPE_COERCE:%.*]], i64 0)
-// CHECK-NEXT:    store <16 x i32> [[TYPE]], ptr [[AGG_RESULT:%.*]], align 16, !tbaa [[TBAA6]]
+// CHECK-NEXT:    store <16 x i32> [[TYPE]], ptr [[AGG_RESULT:%.*]], align 16, !tbaa [[TBAA2]]
 // CHECK-NEXT:    ret void
 //
 gnu_int32_t from_fixed_int32_t__to_gnu_int32_t(fixed_int32_t type) {

llvm/include/llvm/IR/Function.h

@@ -1053,6 +1053,10 @@ class LLVM_ABI Function : public GlobalObject, public ilist_node<Function> {
   /// defined.
   void setAlignment(MaybeAlign Align) { GlobalObject::setAlignment(Align); }
 
+  /// Return the value for vscale based on the vscale_range attribute or 0 when
+  /// unknown.
+  unsigned getVScaleValue() const;
+
 private:
   void allocHungoffUselist();
   template<int Idx> void setHungoffOperand(Constant *C);

llvm/lib/IR/Function.cpp

@@ -1165,6 +1165,18 @@ bool Function::nullPointerIsDefined() const {
   return hasFnAttribute(Attribute::NullPointerIsValid);
 }
 
+unsigned Function::getVScaleValue() const {
+  Attribute Attr = getFnAttribute(Attribute::VScaleRange);
+  if (!Attr.isValid())
+    return 0;
+
+  unsigned VScale = Attr.getVScaleRangeMin();
+  if (VScale && VScale == Attr.getVScaleRangeMax())
+    return VScale;
+
+  return 0;
+}
+
 bool llvm::NullPointerIsDefined(const Function *F, unsigned AS) {
   if (F && F->nullPointerIsDefined())
     return true;