Add support for bfloat in VInsert PreLegalizerCombiner

Xilinx · Nov 21, 2024 · d6afe1c · d6afe1c
1 parent 0b3a94c
commit d6afe1c
Show file tree

Hide file tree

Showing 2 changed files with 232 additions and 1 deletion.
diff --git a/llvm/lib/Target/AIE/AIE2PreLegalizerCombiner.cpp b/llvm/lib/Target/AIE/AIE2PreLegalizerCombiner.cpp
@@ -204,6 +204,7 @@ unsigned getVInsertScalarSize(unsigned IntrinsicID) {
   case Intrinsic::aie2_vinsert8_I512:
     return 8;
   case Intrinsic::aie2_vinsert16_I512:
+  case Intrinsic::aie2_vinsert16_bf512:
     return 16;
   case Intrinsic::aie2_vinsert32_I512:
     return 32;
@@ -226,7 +227,9 @@ AIE2PreLegalizerCombinerImpl::getVectorInsertIndices(
     return isa<GIntrinsic>(MI) && (cast<GIntrinsic>(*MI).getIntrinsicID() ==
                                        Intrinsic::aie2_set_I512_I128 ||
                                    cast<GIntrinsic>(*MI).getIntrinsicID() ==
-                                       Intrinsic::aie2_set_I512_I256);
+                                       Intrinsic::aie2_set_I512_I256 ||
+                                   cast<GIntrinsic>(*MI).getIntrinsicID() ==
+                                       Intrinsic::aie2_set_bf512_bf256);
   };
 
   while (IsVInsert(CurMI, SclSrcBits)) {
@@ -317,6 +320,7 @@ bool AIE2PreLegalizerCombinerImpl::tryToCombineIntrinsic(
   }
   case Intrinsic::aie2_vinsert8_I512:
   case Intrinsic::aie2_vinsert16_I512:
+  case Intrinsic::aie2_vinsert16_bf512:
   case Intrinsic::aie2_vinsert32_I512: {
     return tryToCombineVectorInserts(MI, getVInsertScalarSize(IntrinsicID));
   }

diff --git a/llvm/test/CodeGen/AIE/aie2/GlobalISel/combine-vinsert-sequence-prelegalizer.mir b/llvm/test/CodeGen/AIE/aie2/GlobalISel/combine-vinsert-sequence-prelegalizer.mir
@@ -707,3 +707,230 @@ body:             |
     %24:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.I512), %23(<16 x s32>), %8(s32), %2(s32)
     PseudoRET implicit $lr, implicit %24
 ...
+
+---
+name:            vinsert16-bf256
+legalized:       false
+body:             |
+  bb.1.entry:
+    ; CHECK-LABEL: name: vinsert16-bf256
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
+    ; CHECK-NEXT: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
+    ; CHECK-NEXT: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
+    ; CHECK-NEXT: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK-NEXT: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
+    ; CHECK-NEXT: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
+    ; CHECK-NEXT: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
+    ; CHECK-NEXT: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CHECK-NEXT: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
+    ; CHECK-NEXT: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
+    ; CHECK-NEXT: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
+    ; CHECK-NEXT: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<16 x s16>) = G_BUILD_VECTOR_TRUNC [[C15]](s32), [[C13]](s32), [[C11]](s32), [[C9]](s32), [[C7]](s32), [[C5]](s32), [[C3]](s32), [[C1]](s32), [[C]](s32), [[C2]](s32), [[C4]](s32), [[C6]](s32), [[C8]](s32), [[C10]](s32), [[C12]](s32), [[C14]](s32)
+    ; CHECK-NEXT: [[AIE_PAD_VECTOR_UNDEF:%[0-9]+]]:_(<32 x s16>) = G_AIE_PAD_VECTOR_UNDEF [[BUILD_VECTOR_TRUNC]](<16 x s16>)
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(<16 x s32>) = G_BITCAST [[AIE_PAD_VECTOR_UNDEF]](<32 x s16>)
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[BITCAST]](<16 x s32>)
+    %0:_(<16 x s16>) = G_INTRINSIC intrinsic(@llvm.aie2.v16bfloat16)
+    %100:_(<8 x s32>) = G_BITCAST %0(<16 x s16>)
+    %1:_(s32) = G_CONSTANT i32 0
+    %2:_(s32) = G_CONSTANT i32 1
+    %3:_(s32) = G_CONSTANT i32 2
+    %4:_(s32) = G_CONSTANT i32 3
+    %5:_(s32) = G_CONSTANT i32 4
+    %6:_(s32) = G_CONSTANT i32 5
+    %7:_(s32) = G_CONSTANT i32 6
+    %8:_(s32) = G_CONSTANT i32 7
+    %9:_(s32) = G_CONSTANT i32 8
+    %10:_(s32) = G_CONSTANT i32 9
+    %11:_(s32) = G_CONSTANT i32 10
+    %12:_(s32) = G_CONSTANT i32 11
+    %13:_(s32) = G_CONSTANT i32 12
+    %14:_(s32) = G_CONSTANT i32 13
+    %15:_(s32) = G_CONSTANT i32 14
+    %16:_(s32) = G_CONSTANT i32 15
+    %101:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.set.bf512.bf256), %100(<8 x s32>), %1(s32)
+    %17:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %101(<16 x s32>), %1(s32), %16(s32)
+    %18:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %17(<16 x s32>), %2(s32), %14(s32)
+    %19:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %18(<16 x s32>), %3(s32), %12(s32)
+    %20:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %19(<16 x s32>), %4(s32), %10(s32)
+    %21:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %20(<16 x s32>), %5(s32), %8(s32)
+    %22:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %21(<16 x s32>), %6(s32), %6(s32)
+    %23:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %22(<16 x s32>), %7(s32), %4(s32)
+    %24:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %23(<16 x s32>), %8(s32), %2(s32)
+    %25:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %24(<16 x s32>), %9(s32), %1(s32)
+    %26:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %25(<16 x s32>), %10(s32), %3(s32)
+    %27:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %26(<16 x s32>), %11(s32), %5(s32)
+    %28:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %27(<16 x s32>), %12(s32), %7(s32)
+    %29:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %28(<16 x s32>), %13(s32), %9(s32)
+    %30:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %29(<16 x s32>), %14(s32), %11(s32)
+    %31:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %30(<16 x s32>), %15(s32), %13(s32)
+    %32:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %31(<16 x s32>), %16(s32), %15(s32)
+    PseudoRET implicit $lr, implicit %32
+...
+
+# Negative Test Case: vinsert.16s did not combine into G_BUILD_VECTOR_TRUNC due to a missing vinsert.16 for an index
+---
+name:            vinsert16-bf256_idx_miss
+legalized:       false
+body:             |
+  bb.1.entry:
+
+    ; CHECK-LABEL: name: vinsert16-bf256_idx_miss
+    ; CHECK: [[INT:%[0-9]+]]:_(<16 x s16>) = G_INTRINSIC intrinsic(@llvm.aie2.v16bfloat16)
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(<8 x s32>) = G_BITCAST [[INT]](<16 x s16>)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
+    ; CHECK-NEXT: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
+    ; CHECK-NEXT: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
+    ; CHECK-NEXT: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK-NEXT: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
+    ; CHECK-NEXT: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
+    ; CHECK-NEXT: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
+    ; CHECK-NEXT: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CHECK-NEXT: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
+    ; CHECK-NEXT: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
+    ; CHECK-NEXT: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
+    ; CHECK-NEXT: [[INT1:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.set.bf512.bf256), [[BITCAST]](<8 x s32>), [[C]](s32)
+    ; CHECK-NEXT: [[INT2:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT1]](<16 x s32>), [[C]](s32), [[C15]](s32)
+    ; CHECK-NEXT: [[INT3:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT2]](<16 x s32>), [[C1]](s32), [[C13]](s32)
+    ; CHECK-NEXT: [[INT4:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT3]](<16 x s32>), [[C2]](s32), [[C11]](s32)
+    ; CHECK-NEXT: [[INT5:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT4]](<16 x s32>), [[C3]](s32), [[C9]](s32)
+    ; CHECK-NEXT: [[INT6:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT5]](<16 x s32>), [[C4]](s32), [[C7]](s32)
+    ; CHECK-NEXT: [[INT7:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT6]](<16 x s32>), [[C5]](s32), [[C5]](s32)
+    ; CHECK-NEXT: [[INT8:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT7]](<16 x s32>), [[C6]](s32), [[C3]](s32)
+    ; CHECK-NEXT: [[INT9:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT8]](<16 x s32>), [[C7]](s32), [[C1]](s32)
+    ; CHECK-NEXT: [[INT10:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT9]](<16 x s32>), [[C8]](s32), [[C]](s32)
+    ; CHECK-NEXT: [[INT11:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT10]](<16 x s32>), [[C9]](s32), [[C2]](s32)
+    ; CHECK-NEXT: [[INT12:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT11]](<16 x s32>), [[C10]](s32), [[C4]](s32)
+    ; CHECK-NEXT: [[INT13:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT12]](<16 x s32>), [[C12]](s32), [[C8]](s32)
+    ; CHECK-NEXT: [[INT14:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT13]](<16 x s32>), [[C13]](s32), [[C10]](s32)
+    ; CHECK-NEXT: [[INT15:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT14]](<16 x s32>), [[C14]](s32), [[C12]](s32)
+    ; CHECK-NEXT: [[INT16:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT15]](<16 x s32>), [[C15]](s32), [[C14]](s32)
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[INT16]](<16 x s32>)
+    %0:_(<16 x s16>) = G_INTRINSIC intrinsic(@llvm.aie2.v16bfloat16)
+    %100:_(<8 x s32>) = G_BITCAST %0(<16 x s16>)
+    %1:_(s32) = G_CONSTANT i32 0
+    %2:_(s32) = G_CONSTANT i32 1
+    %3:_(s32) = G_CONSTANT i32 2
+    %4:_(s32) = G_CONSTANT i32 3
+    %5:_(s32) = G_CONSTANT i32 4
+    %6:_(s32) = G_CONSTANT i32 5
+    %7:_(s32) = G_CONSTANT i32 6
+    %8:_(s32) = G_CONSTANT i32 7
+    %9:_(s32) = G_CONSTANT i32 8
+    %10:_(s32) = G_CONSTANT i32 9
+    %11:_(s32) = G_CONSTANT i32 10
+    %12:_(s32) = G_CONSTANT i32 11
+    %13:_(s32) = G_CONSTANT i32 12
+    %14:_(s32) = G_CONSTANT i32 13
+    %15:_(s32) = G_CONSTANT i32 14
+    %16:_(s32) = G_CONSTANT i32 15
+    %101:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.set.bf512.bf256), %100(<8 x s32>), %1(s32)
+    %17:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %101(<16 x s32>), %1(s32), %16(s32)
+    %18:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %17(<16 x s32>), %2(s32), %14(s32)
+    %19:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %18(<16 x s32>), %3(s32), %12(s32)
+    %20:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %19(<16 x s32>), %4(s32), %10(s32)
+    %21:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %20(<16 x s32>), %5(s32), %8(s32)
+    %22:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %21(<16 x s32>), %6(s32), %6(s32)
+    %23:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %22(<16 x s32>), %7(s32), %4(s32)
+    %24:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %23(<16 x s32>), %8(s32), %2(s32)
+    %25:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %24(<16 x s32>), %9(s32), %1(s32)
+    %26:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %25(<16 x s32>), %10(s32), %3(s32)
+    %27:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %26(<16 x s32>), %11(s32), %5(s32)
+    %28:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %27(<16 x s32>), %13(s32), %9(s32)
+    %29:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %28(<16 x s32>), %14(s32), %11(s32)
+    %30:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %29(<16 x s32>), %15(s32), %13(s32)
+    %31:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %30(<16 x s32>), %16(s32), %15(s32)
+    PseudoRET implicit $lr, implicit %31
+...
+
+# Negative Test Case: vinsert.16s did not combine into G_BUILD_VECTOR_TRUNC due to multiple vinsert.16 instructions for a single index
+---
+name:            vinsert16-bf256_idx_multiple
+legalized:       false
+body:             |
+  bb.1.entry:
+    ; CHECK-LABEL: name: vinsert16-bf256_idx_multiple
+    ; CHECK: [[INT:%[0-9]+]]:_(<16 x s16>) = G_INTRINSIC intrinsic(@llvm.aie2.v16bfloat16)
+    ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:_(<8 x s32>) = G_BITCAST [[INT]](<16 x s16>)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; CHECK-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+    ; CHECK-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
+    ; CHECK-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
+    ; CHECK-NEXT: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
+    ; CHECK-NEXT: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
+    ; CHECK-NEXT: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK-NEXT: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
+    ; CHECK-NEXT: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
+    ; CHECK-NEXT: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
+    ; CHECK-NEXT: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
+    ; CHECK-NEXT: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
+    ; CHECK-NEXT: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
+    ; CHECK-NEXT: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
+    ; CHECK-NEXT: [[INT1:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.set.bf512.bf256), [[BITCAST]](<8 x s32>), [[C]](s32)
+    ; CHECK-NEXT: [[INT2:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT1]](<16 x s32>), [[C]](s32), [[C15]](s32)
+    ; CHECK-NEXT: [[INT3:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT2]](<16 x s32>), [[C1]](s32), [[C13]](s32)
+    ; CHECK-NEXT: [[INT4:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT3]](<16 x s32>), [[C2]](s32), [[C11]](s32)
+    ; CHECK-NEXT: [[INT5:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT4]](<16 x s32>), [[C3]](s32), [[C9]](s32)
+    ; CHECK-NEXT: [[INT6:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT5]](<16 x s32>), [[C4]](s32), [[C7]](s32)
+    ; CHECK-NEXT: [[INT7:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT6]](<16 x s32>), [[C5]](s32), [[C5]](s32)
+    ; CHECK-NEXT: [[INT8:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT7]](<16 x s32>), [[C6]](s32), [[C3]](s32)
+    ; CHECK-NEXT: [[INT9:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT8]](<16 x s32>), [[C7]](s32), [[C1]](s32)
+    ; CHECK-NEXT: [[INT10:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT9]](<16 x s32>), [[C8]](s32), [[C]](s32)
+    ; CHECK-NEXT: [[INT11:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT10]](<16 x s32>), [[C9]](s32), [[C2]](s32)
+    ; CHECK-NEXT: [[INT12:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT11]](<16 x s32>), [[C10]](s32), [[C4]](s32)
+    ; CHECK-NEXT: [[INT13:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT12]](<16 x s32>), [[C11]](s32), [[C6]](s32)
+    ; CHECK-NEXT: [[INT14:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT13]](<16 x s32>), [[C12]](s32), [[C8]](s32)
+    ; CHECK-NEXT: [[INT15:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT14]](<16 x s32>), [[C13]](s32), [[C10]](s32)
+    ; CHECK-NEXT: [[INT16:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT15]](<16 x s32>), [[C13]](s32), [[C15]](s32)
+    ; CHECK-NEXT: [[INT17:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT16]](<16 x s32>), [[C14]](s32), [[C12]](s32)
+    ; CHECK-NEXT: [[INT18:%[0-9]+]]:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), [[INT17]](<16 x s32>), [[C15]](s32), [[C14]](s32)
+    ; CHECK-NEXT: PseudoRET implicit $lr, implicit [[INT18]](<16 x s32>)
+    %0:_(<16 x s16>) = G_INTRINSIC intrinsic(@llvm.aie2.v16bfloat16)
+    %100:_(<8 x s32>) = G_BITCAST %0(<16 x s16>)
+    %1:_(s32) = G_CONSTANT i32 0
+    %2:_(s32) = G_CONSTANT i32 1
+    %3:_(s32) = G_CONSTANT i32 2
+    %4:_(s32) = G_CONSTANT i32 3
+    %5:_(s32) = G_CONSTANT i32 4
+    %6:_(s32) = G_CONSTANT i32 5
+    %7:_(s32) = G_CONSTANT i32 6
+    %8:_(s32) = G_CONSTANT i32 7
+    %9:_(s32) = G_CONSTANT i32 8
+    %10:_(s32) = G_CONSTANT i32 9
+    %11:_(s32) = G_CONSTANT i32 10
+    %12:_(s32) = G_CONSTANT i32 11
+    %13:_(s32) = G_CONSTANT i32 12
+    %14:_(s32) = G_CONSTANT i32 13
+    %15:_(s32) = G_CONSTANT i32 14
+    %16:_(s32) = G_CONSTANT i32 15
+    %101:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.set.bf512.bf256), %100(<8 x s32>), %1(s32)
+    %17:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %101(<16 x s32>), %1(s32), %16(s32)
+    %18:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %17(<16 x s32>), %2(s32), %14(s32)
+    %19:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %18(<16 x s32>), %3(s32), %12(s32)
+    %20:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %19(<16 x s32>), %4(s32), %10(s32)
+    %21:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %20(<16 x s32>), %5(s32), %8(s32)
+    %22:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %21(<16 x s32>), %6(s32), %6(s32)
+    %23:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %22(<16 x s32>), %7(s32), %4(s32)
+    %24:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %23(<16 x s32>), %8(s32), %2(s32)
+    %25:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %24(<16 x s32>), %9(s32), %1(s32)
+    %26:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %25(<16 x s32>), %10(s32), %3(s32)
+    %27:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %26(<16 x s32>), %11(s32), %5(s32)
+    %28:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %27(<16 x s32>), %12(s32), %7(s32)
+    %29:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %28(<16 x s32>), %13(s32), %9(s32)
+    %30:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %29(<16 x s32>), %14(s32), %11(s32)
+    %31:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %30(<16 x s32>), %14(s32), %16(s32)
+    %32:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %31(<16 x s32>), %15(s32), %13(s32)
+    %33:_(<16 x s32>) = G_INTRINSIC intrinsic(@llvm.aie2.vinsert16.bf512), %32(<16 x s32>), %16(s32), %15(s32)
+    PseudoRET implicit $lr, implicit %33
+...