llvm/test/CodeGen/X86/AMX/amx-low-intrinsics-no-amx-bitcast.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
   2 ; RUN: opt -mtriple=x86_64 -lower-amx-intrinsics -enable-x86-scalar-amx=true %s -S | FileCheck %s
   3
   4 define dso_local void @test_no_bitcast(i32* %A_mem, i32* %B_mem, i32* %C_mem) local_unnamed_addr #0 {
   5 ; CHECK-LABEL: @test_no_bitcast(
   6 ; CHECK-NEXT:  entry:
   7 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast i32* [[C_MEM:%.*]] to i8*
   8 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_HEADER:%.*]]
   9 ; CHECK:       tileload.scalarize.rows.header:
  10 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_IV:%.*]] = phi i16 [ 0, [[ENTRY:%.*]] ], [ [[TILELOAD_SCALARIZE_ROWS_STEP:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH:%.*]] ]
  11 ; CHECK-NEXT:    [[VEC_PHI_ROW:%.*]] = phi <256 x i32> [ zeroinitializer, [[ENTRY]] ], [ [[TMP10:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH]] ]
  12 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_BODY:%.*]]
  13 ; CHECK:       tileload.scalarize.rows.body:
  14 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_HEADER:%.*]]
  15 ; CHECK:       tileload.scalarize.cols.header:
  16 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_IV:%.*]] = phi i16 [ 0, [[TILELOAD_SCALARIZE_ROWS_BODY]] ], [ [[TILELOAD_SCALARIZE_COLS_STEP:%.*]], [[TILELOAD_SCALARIZE_COLS_LATCH:%.*]] ]
  17 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <256 x i32> [ [[VEC_PHI_ROW]], [[TILELOAD_SCALARIZE_ROWS_BODY]] ], [ [[TMP10]], [[TILELOAD_SCALARIZE_COLS_LATCH]] ]
  18 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_BODY:%.*]]
  19 ; CHECK:       tileload.scalarize.cols.body:
  20 ; CHECK-NEXT:    [[TMP1:%.*]] = zext i16 [[TILELOAD_SCALARIZE_ROWS_IV]] to i64
  21 ; CHECK-NEXT:    [[TMP2:%.*]] = zext i16 [[TILELOAD_SCALARIZE_COLS_IV]] to i64
  22 ; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP1]], 4
  23 ; CHECK-NEXT:    [[TMP4:%.*]] = add i64 [[TMP3]], [[TMP2]]
  24 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[TMP0]] to i32*
  25 ; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr i32, i32* [[TMP5]], i64 [[TMP4]]
  26 ; CHECK-NEXT:    [[TMP7:%.*]] = mul i16 [[TILELOAD_SCALARIZE_ROWS_IV]], 16
  27 ; CHECK-NEXT:    [[TMP8:%.*]] = add i16 [[TMP7]], [[TILELOAD_SCALARIZE_COLS_IV]]
  28 ; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[TMP6]], align 4
  29 ; CHECK-NEXT:    [[TMP10]] = insertelement <256 x i32> [[VEC_PHI]], i32 [[TMP9]], i16 [[TMP8]]
  30 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_LATCH]]
  31 ; CHECK:       tileload.scalarize.cols.latch:
  32 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_STEP]] = add i16 [[TILELOAD_SCALARIZE_COLS_IV]], 1
  33 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_COND:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_COLS_STEP]], 4
  34 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_COLS_COND]], label [[TILELOAD_SCALARIZE_COLS_HEADER]], label [[TILELOAD_SCALARIZE_ROWS_LATCH]]
  35 ; CHECK:       tileload.scalarize.rows.latch:
  36 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_STEP]] = add i16 [[TILELOAD_SCALARIZE_ROWS_IV]], 1
  37 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_COND:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_ROWS_STEP]], 4
  38 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_ROWS_COND]], label [[TILELOAD_SCALARIZE_ROWS_HEADER]], label [[CONTINUE:%.*]]
  39 ; CHECK:       continue:
  40 ; CHECK-NEXT:    [[TMP11:%.*]] = bitcast <256 x i32> [[TMP10]] to x86_amx
  41 ; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[A_MEM:%.*]] to i8*
  42 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_HEADER2:%.*]]
  43 ; CHECK:       tileload.scalarize.rows.header2:
  44 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_IV5:%.*]] = phi i16 [ 0, [[CONTINUE]] ], [ [[TILELOAD_SCALARIZE_ROWS_STEP6:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH4:%.*]] ]
  45 ; CHECK-NEXT:    [[VEC_PHI_ROW14:%.*]] = phi <256 x i32> [ zeroinitializer, [[CONTINUE]] ], [ [[TMP22:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH4]] ]
  46 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_BODY3:%.*]]
  47 ; CHECK:       tileload.scalarize.rows.body3:
  48 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_HEADER8:%.*]]
  49 ; CHECK:       tileload.scalarize.cols.header8:
  50 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_IV11:%.*]] = phi i16 [ 0, [[TILELOAD_SCALARIZE_ROWS_BODY3]] ], [ [[TILELOAD_SCALARIZE_COLS_STEP12:%.*]], [[TILELOAD_SCALARIZE_COLS_LATCH10:%.*]] ]
  51 ; CHECK-NEXT:    [[VEC_PHI15:%.*]] = phi <256 x i32> [ [[VEC_PHI_ROW14]], [[TILELOAD_SCALARIZE_ROWS_BODY3]] ], [ [[TMP22]], [[TILELOAD_SCALARIZE_COLS_LATCH10]] ]
  52 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_BODY9:%.*]]
  53 ; CHECK:       tileload.scalarize.cols.body9:
  54 ; CHECK-NEXT:    [[TMP13:%.*]] = zext i16 [[TILELOAD_SCALARIZE_ROWS_IV5]] to i64
  55 ; CHECK-NEXT:    [[TMP14:%.*]] = zext i16 [[TILELOAD_SCALARIZE_COLS_IV11]] to i64
  56 ; CHECK-NEXT:    [[TMP15:%.*]] = mul i64 [[TMP13]], 4
  57 ; CHECK-NEXT:    [[TMP16:%.*]] = add i64 [[TMP15]], [[TMP14]]
  58 ; CHECK-NEXT:    [[TMP17:%.*]] = bitcast i8* [[TMP12]] to i32*
  59 ; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i64 [[TMP16]]
  60 ; CHECK-NEXT:    [[TMP19:%.*]] = mul i16 [[TILELOAD_SCALARIZE_ROWS_IV5]], 16
  61 ; CHECK-NEXT:    [[TMP20:%.*]] = add i16 [[TMP19]], [[TILELOAD_SCALARIZE_COLS_IV11]]
  62 ; CHECK-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TMP18]], align 4
  63 ; CHECK-NEXT:    [[TMP22]] = insertelement <256 x i32> [[VEC_PHI15]], i32 [[TMP21]], i16 [[TMP20]]
  64 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_LATCH10]]
  65 ; CHECK:       tileload.scalarize.cols.latch10:
  66 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_STEP12]] = add i16 [[TILELOAD_SCALARIZE_COLS_IV11]], 1
  67 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_COND13:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_COLS_STEP12]], 4
  68 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_COLS_COND13]], label [[TILELOAD_SCALARIZE_COLS_HEADER8]], label [[TILELOAD_SCALARIZE_ROWS_LATCH4]]
  69 ; CHECK:       tileload.scalarize.rows.latch4:
  70 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_STEP6]] = add i16 [[TILELOAD_SCALARIZE_ROWS_IV5]], 1
  71 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_COND7:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_ROWS_STEP6]], 4
  72 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_ROWS_COND7]], label [[TILELOAD_SCALARIZE_ROWS_HEADER2]], label [[CONTINUE1:%.*]]
  73 ; CHECK:       continue1:
  74 ; CHECK-NEXT:    [[TMP23:%.*]] = bitcast <256 x i32> [[TMP22]] to x86_amx
  75 ; CHECK-NEXT:    [[TMP24:%.*]] = bitcast i32* [[B_MEM:%.*]] to i8*
  76 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_HEADER17:%.*]]
  77 ; CHECK:       tileload.scalarize.rows.header17:
  78 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_IV20:%.*]] = phi i16 [ 0, [[CONTINUE1]] ], [ [[TILELOAD_SCALARIZE_ROWS_STEP21:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH19:%.*]] ]
  79 ; CHECK-NEXT:    [[VEC_PHI_ROW29:%.*]] = phi <256 x i32> [ zeroinitializer, [[CONTINUE1]] ], [ [[TMP34:%.*]], [[TILELOAD_SCALARIZE_ROWS_LATCH19]] ]
  80 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_ROWS_BODY18:%.*]]
  81 ; CHECK:       tileload.scalarize.rows.body18:
  82 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_HEADER23:%.*]]
  83 ; CHECK:       tileload.scalarize.cols.header23:
  84 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_IV26:%.*]] = phi i16 [ 0, [[TILELOAD_SCALARIZE_ROWS_BODY18]] ], [ [[TILELOAD_SCALARIZE_COLS_STEP27:%.*]], [[TILELOAD_SCALARIZE_COLS_LATCH25:%.*]] ]
  85 ; CHECK-NEXT:    [[VEC_PHI30:%.*]] = phi <256 x i32> [ [[VEC_PHI_ROW29]], [[TILELOAD_SCALARIZE_ROWS_BODY18]] ], [ [[TMP34]], [[TILELOAD_SCALARIZE_COLS_LATCH25]] ]
  86 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_BODY24:%.*]]
  87 ; CHECK:       tileload.scalarize.cols.body24:
  88 ; CHECK-NEXT:    [[TMP25:%.*]] = zext i16 [[TILELOAD_SCALARIZE_ROWS_IV20]] to i64
  89 ; CHECK-NEXT:    [[TMP26:%.*]] = zext i16 [[TILELOAD_SCALARIZE_COLS_IV26]] to i64
  90 ; CHECK-NEXT:    [[TMP27:%.*]] = mul i64 [[TMP25]], 4
  91 ; CHECK-NEXT:    [[TMP28:%.*]] = add i64 [[TMP27]], [[TMP26]]
  92 ; CHECK-NEXT:    [[TMP29:%.*]] = bitcast i8* [[TMP24]] to i32*
  93 ; CHECK-NEXT:    [[TMP30:%.*]] = getelementptr i32, i32* [[TMP29]], i64 [[TMP28]]
  94 ; CHECK-NEXT:    [[TMP31:%.*]] = mul i16 [[TILELOAD_SCALARIZE_ROWS_IV20]], 16
  95 ; CHECK-NEXT:    [[TMP32:%.*]] = add i16 [[TMP31]], [[TILELOAD_SCALARIZE_COLS_IV26]]
  96 ; CHECK-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP30]], align 4
  97 ; CHECK-NEXT:    [[TMP34]] = insertelement <256 x i32> [[VEC_PHI30]], i32 [[TMP33]], i16 [[TMP32]]
  98 ; CHECK-NEXT:    br label [[TILELOAD_SCALARIZE_COLS_LATCH25]]
  99 ; CHECK:       tileload.scalarize.cols.latch25:
 100 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_STEP27]] = add i16 [[TILELOAD_SCALARIZE_COLS_IV26]], 1
 101 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_COLS_COND28:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_COLS_STEP27]], 4
 102 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_COLS_COND28]], label [[TILELOAD_SCALARIZE_COLS_HEADER23]], label [[TILELOAD_SCALARIZE_ROWS_LATCH19]]
 103 ; CHECK:       tileload.scalarize.rows.latch19:
 104 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_STEP21]] = add i16 [[TILELOAD_SCALARIZE_ROWS_IV20]], 1
 105 ; CHECK-NEXT:    [[TILELOAD_SCALARIZE_ROWS_COND22:%.*]] = icmp ne i16 [[TILELOAD_SCALARIZE_ROWS_STEP21]], 4
 106 ; CHECK-NEXT:    br i1 [[TILELOAD_SCALARIZE_ROWS_COND22]], label [[TILELOAD_SCALARIZE_ROWS_HEADER17]], label [[CONTINUE16:%.*]]
 107 ; CHECK:       continue16:
 108 ; CHECK-NEXT:    [[TMP35:%.*]] = bitcast <256 x i32> [[TMP34]] to x86_amx
 109 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_ROWS_HEADER:%.*]]
 110 ; CHECK:       tiledpbssd.scalarize.rows.header:
 111 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_ROWS_IV:%.*]] = phi i16 [ 0, [[CONTINUE16]] ], [ [[TILEDPBSSD_SCALARIZE_ROWS_STEP:%.*]], [[TILEDPBSSD_SCALARIZE_ROWS_LATCH:%.*]] ]
 112 ; CHECK-NEXT:    [[VEC_C_PHI_ROW:%.*]] = phi <256 x i32> [ [[TMP10]], [[CONTINUE16]] ], [ [[TMP52:%.*]], [[TILEDPBSSD_SCALARIZE_ROWS_LATCH]] ]
 113 ; CHECK-NEXT:    [[VEC_D_PHI_ROW:%.*]] = phi <256 x i32> [ zeroinitializer, [[CONTINUE16]] ], [ [[TMP54:%.*]], [[TILEDPBSSD_SCALARIZE_ROWS_LATCH]] ]
 114 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_ROWS_BODY:%.*]]
 115 ; CHECK:       tiledpbssd.scalarize.rows.body:
 116 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_COLS_HEADER:%.*]]
 117 ; CHECK:       tiledpbssd.scalarize.cols.header:
 118 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_COLS_IV:%.*]] = phi i16 [ 0, [[TILEDPBSSD_SCALARIZE_ROWS_BODY]] ], [ [[TILEDPBSSD_SCALARIZE_COLS_STEP:%.*]], [[TILEDPBSSD_SCALARIZE_COLS_LATCH:%.*]] ]
 119 ; CHECK-NEXT:    [[VEC_C_PHI_COL:%.*]] = phi <256 x i32> [ [[VEC_C_PHI_ROW]], [[TILEDPBSSD_SCALARIZE_ROWS_BODY]] ], [ [[TMP52]], [[TILEDPBSSD_SCALARIZE_COLS_LATCH]] ]
 120 ; CHECK-NEXT:    [[VEC_D_PHI_COL:%.*]] = phi <256 x i32> [ [[VEC_D_PHI_ROW]], [[TILEDPBSSD_SCALARIZE_ROWS_BODY]] ], [ [[TMP54]], [[TILEDPBSSD_SCALARIZE_COLS_LATCH]] ]
 121 ; CHECK-NEXT:    [[TMP36:%.*]] = mul i16 [[TILEDPBSSD_SCALARIZE_ROWS_IV]], 16
 122 ; CHECK-NEXT:    [[TMP37:%.*]] = add i16 [[TMP36]], [[TILEDPBSSD_SCALARIZE_COLS_IV]]
 123 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_COLS_BODY:%.*]]
 124 ; CHECK:       tiledpbssd.scalarize.cols.body:
 125 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_INNER_HEADER:%.*]]
 126 ; CHECK:       tiledpbssd.scalarize.inner.header:
 127 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_INNER_IV:%.*]] = phi i16 [ 0, [[TILEDPBSSD_SCALARIZE_COLS_BODY]] ], [ [[TILEDPBSSD_SCALARIZE_INNER_STEP:%.*]], [[TILEDPBSSD_SCALARIZE_INNER_LATCH:%.*]] ]
 128 ; CHECK-NEXT:    [[VEC_C_INNER_PHI:%.*]] = phi <256 x i32> [ [[VEC_C_PHI_COL]], [[TILEDPBSSD_SCALARIZE_COLS_BODY]] ], [ [[TMP52]], [[TILEDPBSSD_SCALARIZE_INNER_LATCH]] ]
 129 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_INNER_BODY:%.*]]
 130 ; CHECK:       tiledpbssd.scalarize.inner.body:
 131 ; CHECK-NEXT:    [[TMP38:%.*]] = mul i16 [[TILEDPBSSD_SCALARIZE_ROWS_IV]], 16
 132 ; CHECK-NEXT:    [[TMP39:%.*]] = add i16 [[TMP38]], [[TILEDPBSSD_SCALARIZE_INNER_IV]]
 133 ; CHECK-NEXT:    [[TMP40:%.*]] = mul i16 [[TILEDPBSSD_SCALARIZE_INNER_IV]], 16
 134 ; CHECK-NEXT:    [[TMP41:%.*]] = add i16 [[TMP40]], [[TILEDPBSSD_SCALARIZE_COLS_IV]]
 135 ; CHECK-NEXT:    [[TMP42:%.*]] = extractelement <256 x i32> [[VEC_C_INNER_PHI]], i16 [[TMP37]]
 136 ; CHECK-NEXT:    [[TMP43:%.*]] = extractelement <256 x i32> [[TMP22]], i16 [[TMP39]]
 137 ; CHECK-NEXT:    [[TMP44:%.*]] = bitcast i32 [[TMP43]] to <4 x i8>
 138 ; CHECK-NEXT:    [[TMP45:%.*]] = extractelement <256 x i32> [[TMP34]], i16 [[TMP41]]
 139 ; CHECK-NEXT:    [[TMP46:%.*]] = bitcast i32 [[TMP45]] to <4 x i8>
 140 ; CHECK-NEXT:    [[TMP47:%.*]] = sext <4 x i8> [[TMP46]] to <4 x i32>
 141 ; CHECK-NEXT:    [[TMP48:%.*]] = sext <4 x i8> [[TMP44]] to <4 x i32>
 142 ; CHECK-NEXT:    [[TMP49:%.*]] = mul <4 x i32> [[TMP48]], [[TMP47]]
 143 ; CHECK-NEXT:    [[TMP50:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP49]])
 144 ; CHECK-NEXT:    [[TMP51:%.*]] = add i32 [[TMP42]], [[TMP50]]
 145 ; CHECK-NEXT:    [[TMP52]] = insertelement <256 x i32> [[VEC_C_INNER_PHI]], i32 [[TMP51]], i16 [[TMP37]]
 146 ; CHECK-NEXT:    br label [[TILEDPBSSD_SCALARIZE_INNER_LATCH]]
 147 ; CHECK:       tiledpbssd.scalarize.inner.latch:
 148 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_INNER_STEP]] = add i16 [[TILEDPBSSD_SCALARIZE_INNER_IV]], 1
 149 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_INNER_COND:%.*]] = icmp ne i16 [[TILEDPBSSD_SCALARIZE_INNER_STEP]], 4
 150 ; CHECK-NEXT:    br i1 [[TILEDPBSSD_SCALARIZE_INNER_COND]], label [[TILEDPBSSD_SCALARIZE_INNER_HEADER]], label [[TILEDPBSSD_SCALARIZE_COLS_LATCH]]
 151 ; CHECK:       tiledpbssd.scalarize.cols.latch:
 152 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_COLS_STEP]] = add i16 [[TILEDPBSSD_SCALARIZE_COLS_IV]], 1
 153 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_COLS_COND:%.*]] = icmp ne i16 [[TILEDPBSSD_SCALARIZE_COLS_STEP]], 4
 154 ; CHECK-NEXT:    [[TMP53:%.*]] = extractelement <256 x i32> [[TMP52]], i16 [[TMP37]]
 155 ; CHECK-NEXT:    [[TMP54]] = insertelement <256 x i32> [[VEC_D_PHI_COL]], i32 [[TMP53]], i16 [[TMP37]]
 156 ; CHECK-NEXT:    br i1 [[TILEDPBSSD_SCALARIZE_COLS_COND]], label [[TILEDPBSSD_SCALARIZE_COLS_HEADER]], label [[TILEDPBSSD_SCALARIZE_ROWS_LATCH]]
 157 ; CHECK:       tiledpbssd.scalarize.rows.latch:
 158 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_ROWS_STEP]] = add i16 [[TILEDPBSSD_SCALARIZE_ROWS_IV]], 1
 159 ; CHECK-NEXT:    [[TILEDPBSSD_SCALARIZE_ROWS_COND:%.*]] = icmp ne i16 [[TILEDPBSSD_SCALARIZE_ROWS_STEP]], 4
 160 ; CHECK-NEXT:    br i1 [[TILEDPBSSD_SCALARIZE_ROWS_COND]], label [[TILEDPBSSD_SCALARIZE_ROWS_HEADER]], label [[CONTINUE31:%.*]]
 161 ; CHECK:       continue31:
 162 ; CHECK-NEXT:    [[TMP55:%.*]] = bitcast <256 x i32> [[TMP54]] to x86_amx
 163 ; CHECK-NEXT:    br label [[TILESTORE_SCALARIZE_ROWS_HEADER:%.*]]
 164 ; CHECK:       tilestore.scalarize.rows.header:
 165 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_ROWS_IV:%.*]] = phi i16 [ 0, [[CONTINUE31]] ], [ [[TILESTORE_SCALARIZE_ROWS_STEP:%.*]], [[TILESTORE_SCALARIZE_ROWS_LATCH:%.*]] ]
 166 ; CHECK-NEXT:    br label [[TILESTORE_SCALARIZE_ROWS_BODY:%.*]]
 167 ; CHECK:       tilestore.scalarize.rows.body:
 168 ; CHECK-NEXT:    br label [[TILESTORE_SCALARIZE_COLS_HEADER:%.*]]
 169 ; CHECK:       tilestore.scalarize.cols.header:
 170 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_COLS_IV:%.*]] = phi i16 [ 0, [[TILESTORE_SCALARIZE_ROWS_BODY]] ], [ [[TILESTORE_SCALARIZE_COLS_STEP:%.*]], [[TILESTORE_SCALARIZE_COLS_LATCH:%.*]] ]
 171 ; CHECK-NEXT:    br label [[TILESTORE_SCALARIZE_COLS_BODY:%.*]]
 172 ; CHECK:       tilestore.scalarize.cols.body:
 173 ; CHECK-NEXT:    [[TMP56:%.*]] = zext i16 [[TILESTORE_SCALARIZE_ROWS_IV]] to i64
 174 ; CHECK-NEXT:    [[TMP57:%.*]] = zext i16 [[TILESTORE_SCALARIZE_COLS_IV]] to i64
 175 ; CHECK-NEXT:    [[TMP58:%.*]] = mul i64 [[TMP56]], 4
 176 ; CHECK-NEXT:    [[TMP59:%.*]] = add i64 [[TMP58]], [[TMP57]]
 177 ; CHECK-NEXT:    [[TMP60:%.*]] = bitcast i8* [[TMP0]] to i32*
 178 ; CHECK-NEXT:    [[TMP61:%.*]] = getelementptr i32, i32* [[TMP60]], i64 [[TMP59]]
 179 ; CHECK-NEXT:    [[TMP62:%.*]] = mul i16 [[TILESTORE_SCALARIZE_ROWS_IV]], 16
 180 ; CHECK-NEXT:    [[TMP63:%.*]] = add i16 [[TMP62]], [[TILESTORE_SCALARIZE_COLS_IV]]
 181 ; CHECK-NEXT:    [[TMP64:%.*]] = extractelement <256 x i32> [[TMP54]], i16 [[TMP63]]
 182 ; CHECK-NEXT:    store i32 [[TMP64]], i32* [[TMP61]], align 4
 183 ; CHECK-NEXT:    br label [[TILESTORE_SCALARIZE_COLS_LATCH]]
 184 ; CHECK:       tilestore.scalarize.cols.latch:
 185 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_COLS_STEP]] = add i16 [[TILESTORE_SCALARIZE_COLS_IV]], 1
 186 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_COLS_COND:%.*]] = icmp ne i16 [[TILESTORE_SCALARIZE_COLS_STEP]], 4
 187 ; CHECK-NEXT:    br i1 [[TILESTORE_SCALARIZE_COLS_COND]], label [[TILESTORE_SCALARIZE_COLS_HEADER]], label [[TILESTORE_SCALARIZE_ROWS_LATCH]]
 188 ; CHECK:       tilestore.scalarize.rows.latch:
 189 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_ROWS_STEP]] = add i16 [[TILESTORE_SCALARIZE_ROWS_IV]], 1
 190 ; CHECK-NEXT:    [[TILESTORE_SCALARIZE_ROWS_COND:%.*]] = icmp ne i16 [[TILESTORE_SCALARIZE_ROWS_STEP]], 4
 191 ; CHECK-NEXT:    br i1 [[TILESTORE_SCALARIZE_ROWS_COND]], label [[TILESTORE_SCALARIZE_ROWS_HEADER]], label [[CONTINUE32:%.*]]
 192 ; CHECK:       continue32:
 193 ; CHECK-NEXT:    ret void
 194 ;
 195 entry:
 196   %0 = bitcast i32* %C_mem to i8*
 197   %1 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 4, i16 16, i8* %0, i64 16)
 198   %2 = bitcast i32* %A_mem to i8*
 199   %3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 4, i16 16, i8* %2, i64 16)
 200   %4 = bitcast i32* %B_mem to i8*
 201   %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 4, i16 16, i8* %4, i64 16)
 202   %6 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 4, i16 16, i16 16, x86_amx %1, x86_amx %3, x86_amx %5)
 203   tail call void @llvm.x86.tilestored64.internal(i16 4, i16 16, i8* %0, i64 16, x86_amx %6)
 204   ret void
 205 }
 206
 207 declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
 208 declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
 209 declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
 210
 211 attributes #0 = { noinline nounwind optnone }