[CodeGen][SVE] CopyToReg: Split scalable EVTs that are not powers of 2

[sdesmalen]

Summary

Scalable vectors cannot use 'BUILD_VECTOR', so it is necessary to
properly split and widen scalable vectors when passing them
to CopyToReg/CopyFromReg.

This functionality is added to TargetLoweringBase::getVectorTypeBreakdown().

This patch only adds support for 'splitting' scalable vectors that
are a multiple of some legal type, e.g.

<vscale x 6 x i64> -> 3 x <vscale x 2 x i64>

[sdesmalen]

llvm/lib/CodeGen/TargetLoweringBase.cpp

955

The reason for not implementing this yet is because I don't have a way to test it.

1425

note: Widening of scalable vectors requires a bit more thought.

For example, I wasn't sure how a <vscale x 3 x i8> should be widened at this level. Is it widened to a <vscale x 4 x i8> and then promoted to a <vscale x 4 x i32>, or do we keep it simple and always insert it into an UNDEF of the first legal type (<vscale x 16 x i8> for SVE).

It would also require changes to legalise e.g. <vscale x 9 x i16>, that needs both widening to <vscale x 16 x i16>, and then splitting to two <vscale x 8 x i16>.

The case handled by this patch allows it to support the tuple types for the structured load/store instructions, such as svint32x3_t (implemented as <vscale x 12 x i32>).

[efriedma]

llvm/lib/CodeGen/TargetLoweringBase.cpp

955

I think the only way to trigger this is by adding more types to llvm/include/llvm/Support/MachineValueType.h; it's fine to ignore that for now.

1425

For widening values that fit in a single register, I think we should do whatever getTypeToTransformTo() would do. If you look a few lines up, there's a call to getTypeAction(); that code is supposed to handle cases where the type would fit into a single vector register. It probably makes sense to enhance it to be a bit more comprehensive in cases where there's more than one legalization step involved.

If we have a value with a legal element type that has more elements than the legal vector type, I guess we have to widen, then split: so <vscale x 9 x i16> would end up using <vscale x 8 x i16> components.

1437

Is flooring division correct here?

[c-rhodes]

llvm/test/CodeGen/AArch64/sve-breakdown-scalable-vectortype.ll

51

do you mean <vscale x 16 x i8> %legal? (and below)

[sdesmalen]

• Reuse getTypeConversion to break down the vector into legal parts.
• Added an assert that the illegal vector can be broken into some integer multiple of part-vectors.
• Fixed up <vscale x 14 x .. types in tests.

[efriedma]

llvm/lib/CodeGen/TargetLoweringBase.cpp

1439

I'm pretty sure this assert doesn't hold in general. If you don't want to think about it right now, I guess you can leave a FIXME.

[sdesmalen]

Added FIXME describing limitations of vector type breakdown.

[efriedma]

LGTM

llvm/lib/CodeGen/TargetLoweringBase.cpp

static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
                                           MVT &RegisterVT,
                                           TargetLoweringBase *TLI) {
   // Figure out the right, legal destination reg to copy into.
   ElementCount EC = VT.getVectorElementCount();
   MVT EltTy = VT.getVectorElementType();
 
   unsigned NumVectorRegs = 1;
 
+  // Scalable vectors cannot be scalarized, so splitting or widening is
+  // required.
+  if (VT.isScalableVector() && !isPowerOf2_32(EC.Min))
+    llvm_unreachable(
+        "Splitting or widening of non-power-of-2 MVTs is not implemented.");
+
   // FIXME: We don't support non-power-of-2-sized vectors for now.
   // Ideally we could break down into LHS/RHS like LegalizeDAG does.
   if (!isPowerOf2_32(EC.Min)) {
     // Split EC to unit size (scalable property is preserved).
     NumVectorRegs = EC.Min;
     EC = EC / NumVectorRegs;
   }
 
if (EltCnt.Min != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
     }
   }
 
   // Figure out the right, legal destination reg to copy into.
   EVT EltTy = VT.getVectorElementType();
 
   unsigned NumVectorRegs = 1;
 
-  // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally we
-  // could break down into LHS/RHS like LegalizeDAG does.
+  // Scalable vectors cannot be scalarized, so splitting is required.
+  // FIXME: Implement widening of scalable vectors.
+  if (VT.isScalableVector() && !isPowerOf2_32(EltCnt.Min)) {
+    EVT WidenedVT = VT.getPow2VectorType(Context);
+    EVT SubPartVT = WidenedVT;
+    while (SubPartVT.getVectorElementCount().Min > 1) {
+      // See if the power-of-two vector can be broken down into
+      // legal power-of-two subvectors.
+      SubPartVT = SubPartVT.getHalfNumVectorElementsVT(Context);
+      if (!isTypeLegal(SubPartVT))
+        continue;
+
+      NumIntermediates = VT.getVectorElementCount().Min /
+                         SubPartVT.getVectorElementCount().Min;
+      IntermediateVT = SubPartVT;
+      RegisterVT = getRegisterType(Context, IntermediateVT);
+      return NumIntermediates;
+    }
+
+    llvm_unreachable("Widening of scalable vectors not yet supported.");
+  }
+
+  // FIXME: We don't support non-power-of-2-sized vectors for now.  Ideally
+  // we could break down into LHS/RHS like LegalizeDAG does.
   if (!isPowerOf2_32(EltCnt.Min)) {
     NumVectorRegs = EltCnt.Min;
     EltCnt.Min = 1;
   }
 
   // Divide the input until we get to a supported size.  This will always
   // end with a scalar if the target doesn't support vectors.
   while (EltCnt.Min > 1 &&

llvm/test/CodeGen/AArch64/sve-breakdown-scalable-vectortype.ll

+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=+sve -asm-verbose=0 < %s | FileCheck %s
+
+; Test that scalable vectors that are a multiple of the legal vector size
+; can be properly broken down into part vectors.
+
+declare void @bar()
+
+;
+; Vectors twice the size
+;
+
+define <vscale x 32 x i8> @wide_32i8(i1 %b, <vscale x 16 x i8> %legal, <vscale x 32 x i8> %illegal) nounwind {
+; CHECK-LABEL: wide_32i8
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 32 x i8> undef
+L2:
+  ret <vscale x 32 x i8> %illegal
+}
+
+define <vscale x 16 x i16> @wide_16i16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 16 x i16> %illegal) nounwind {
+; CHECK-LABEL: wide_16i16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 16 x i16> undef
+L2:
+  ret <vscale x 16 x i16> %illegal
+}
+
+define <vscale x 8 x i32> @wide_8i32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 8 x i32> %illegal) nounwind {
+; CHECK-LABEL: wide_8i32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 8 x i32> undef
+L2:
+  ret <vscale x 8 x i32> %illegal
+}
+
+define <vscale x 4 x i64> @wide_4i64(i1 %b, <vscale x 14 x i8> %legal, <vscale x 4 x i64> %illegal) nounwind {
+; CHECK-LABEL: wide_4i64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 4 x i64> undef
+L2:
+  ret <vscale x 4 x i64> %illegal
+}
+
+define <vscale x 16 x half> @wide_16f16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 16 x half> %illegal) nounwind {
+; CHECK-LABEL: wide_16f16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 16 x half> undef
+L2:
+  ret <vscale x 16 x half> %illegal
+}
+
+define <vscale x 8 x float> @wide_8f32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 8 x float> %illegal) nounwind {
+; CHECK-LABEL: wide_8f32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 8 x float> undef
+L2:
+  ret <vscale x 8 x float> %illegal
+}
+
+define <vscale x 4 x double> @wide_4f64(i1 %b, <vscale x 14 x i8> %legal, <vscale x 4 x double> %illegal) nounwind {
+; CHECK-LABEL: wide_4f64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 4 x double> undef
+L2:
+  ret <vscale x 4 x double> %illegal
+}
+
+;
+; Vectors three times the size
+;
+
+define <vscale x 48 x i8> @wide_48i8(i1 %b, <vscale x 16 x i8> %legal, <vscale x 48 x i8> %illegal) nounwind {
+; CHECK-LABEL: wide_48i8
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 48 x i8> undef
+L2:
+  ret <vscale x 48 x i8> %illegal
+}
+
+define <vscale x 24 x i16> @wide_24i16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 24 x i16> %illegal) nounwind {
+; CHECK-LABEL: wide_24i16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 24 x i16> undef
+L2:
+  ret <vscale x 24 x i16> %illegal
+}
+
+define <vscale x 12 x i32> @wide_12i32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 12 x i32> %illegal) nounwind {
+; CHECK-LABEL: wide_12i32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 12 x i32> undef
+L2:
+  ret <vscale x 12 x i32> %illegal
+}
+
+define <vscale x 6 x i64> @wide_6i64(i1 %b, <vscale x 16 x i8> %legal, <vscale x 6 x i64> %illegal) nounwind {
+; CHECK-LABEL: wide_6i64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 6 x i64> undef
+L2:
+  ret <vscale x 6 x i64> %illegal
+}
+
+define <vscale x 24 x half> @wide_24f16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 24 x half> %illegal) nounwind {
+; CHECK-LABEL: wide_24f16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 24 x half> undef
+L2:
+  ret <vscale x 24 x half> %illegal
+}
+
+define <vscale x 12 x float> @wide_12f32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 12 x float> %illegal) nounwind {
+; CHECK-LABEL: wide_12f32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 12 x float> undef
+L2:
+  ret <vscale x 12 x float> %illegal
+}
+
+define <vscale x 6 x double> @wide_6f64(i1 %b, <vscale x 16 x i8> %legal, <vscale x 6 x double> %illegal) nounwind {
+; CHECK-LABEL: wide_6f64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 6 x double> undef
+L2:
+  ret <vscale x 6 x double> %illegal
+}
+
+;
+; Vectors four times the size
+;
+
+define <vscale x 64 x i8> @wide_64i8(i1 %b, <vscale x 16 x i8> %legal, <vscale x 64 x i8> %illegal) nounwind {
+; CHECK-LABEL: wide_64i8
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 64 x i8> undef
+L2:
+  ret <vscale x 64 x i8> %illegal
+}
+
+define <vscale x 32 x i16> @wide_32i16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 32 x i16> %illegal) nounwind {
+; CHECK-LABEL: wide_32i16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 32 x i16> undef
+L2:
+  ret <vscale x 32 x i16> %illegal
+}
+
+define <vscale x 16 x i32> @wide_16i32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 16 x i32> %illegal) nounwind {
+; CHECK-LABEL: wide_16i32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 16 x i32> undef
+L2:
+  ret <vscale x 16 x i32> %illegal
+}
+
+define <vscale x 8 x i64> @wide_8i64(i1 %b, <vscale x 16 x i8> %legal, <vscale x 8 x i64> %illegal) nounwind {
+; CHECK-LABEL: wide_8i64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 8 x i64> undef
+L2:
+  ret <vscale x 8 x i64> %illegal
+}
+
+define <vscale x 32 x half> @wide_32f16(i1 %b, <vscale x 16 x i8> %legal, <vscale x 32 x half> %illegal) nounwind {
+; CHECK-LABEL: wide_32f16
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 32 x half> undef
+L2:
+  ret <vscale x 32 x half> %illegal
+}
+
+define <vscale x 16 x float> @wide_16f32(i1 %b, <vscale x 16 x i8> %legal, <vscale x 16 x float> %illegal) nounwind {
+; CHECK-LABEL: wide_16f32
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 16 x float> undef
+L2:
+  ret <vscale x 16 x float> %illegal
+}
+
+define <vscale x 8 x double> @wide_8f64(i1 %b, <vscale x 16 x i8> %legal, <vscale x 8 x double> %illegal) nounwind {
+; CHECK-LABEL: wide_8f64
+; CHECK:      mov     z0.d, z1.d
+; CHECK-NEXT: mov     z1.d, z2.d
+; CHECK-NEXT: mov     z2.d, z3.d
+; CHECK-NEXT: mov     z3.d, z4.d
+; CHECK-NEXT: ret
+  br i1 %b, label %L1, label %L2
+L1:
+  call void @bar()
+  ret <vscale x 8 x double> undef
+L2:
+  ret <vscale x 8 x double> %illegal
+}