80 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 62; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-NO-PARTIAL-REDUCTION3 4target triple = "aarch64"5 6define i128 @add_reduc_i32_i128_unsupported(ptr %a, ptr %b) "target-features"="+dotprod" {7; CHECK-NO-PARTIAL-REDUCTION-LABEL: define i128 @add_reduc_i32_i128_unsupported(8; CHECK-NO-PARTIAL-REDUCTION-SAME: ptr [[A:%.*]], ptr [[B:%.*]]) #[[ATTR0:[0-9]+]] {9; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[ENTRY:.*:]]10; CHECK-NO-PARTIAL-REDUCTION-NEXT: br label %[[VECTOR_PH:.*]]11; CHECK-NO-PARTIAL-REDUCTION: [[VECTOR_PH]]:12; CHECK-NO-PARTIAL-REDUCTION-NEXT: br label %[[VECTOR_BODY:.*]]13; CHECK-NO-PARTIAL-REDUCTION: [[VECTOR_BODY]]:14; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]15; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[VEC_PHI:%.*]] = phi <4 x i128> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP7:%.*]], %[[VECTOR_BODY]] ]16; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP0:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]17; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 118; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP1:%.*]] = zext <4 x i32> [[WIDE_LOAD]] to <4 x i64>19; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]20; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i32>, ptr [[TMP2]], align 121; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP3:%.*]] = zext <4 x i32> [[WIDE_LOAD1]] to <4 x i64>22; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP4:%.*]] = mul nuw <4 x i64> [[TMP1]], [[TMP3]]23; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP5:%.*]] = zext <4 x i64> [[TMP4]] to <4 x i128>24; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP7]] = add <4 x i128> [[VEC_PHI]], [[TMP5]]25; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 426; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 402427; CHECK-NO-PARTIAL-REDUCTION-NEXT: br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]28; CHECK-NO-PARTIAL-REDUCTION: [[MIDDLE_BLOCK]]:29; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[TMP8:%.*]] = call i128 @llvm.vector.reduce.add.v4i128(<4 x i128> [[TMP7]])30; CHECK-NO-PARTIAL-REDUCTION-NEXT: br label %[[SCALAR_PH:.*]]31; CHECK-NO-PARTIAL-REDUCTION: [[SCALAR_PH]]:32; CHECK-NO-PARTIAL-REDUCTION-NEXT: br label %[[FOR_BODY:.*]]33; CHECK-NO-PARTIAL-REDUCTION: [[FOR_BODY]]:34; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[IV:%.*]] = phi i64 [ 4024, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[FOR_BODY]] ]35; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[ACCUM:%.*]] = phi i128 [ [[TMP8]], %[[SCALAR_PH]] ], [ [[ADD:%.*]], %[[FOR_BODY]] ]36; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[GEP_A:%.*]] = getelementptr i32, ptr [[A]], i64 [[IV]]37; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[LOAD_A:%.*]] = load i32, ptr [[GEP_A]], align 138; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[EXT_A:%.*]] = zext i32 [[LOAD_A]] to i6439; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[GEP_B:%.*]] = getelementptr i32, ptr [[B]], i64 [[IV]]40; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[LOAD_B:%.*]] = load i32, ptr [[GEP_B]], align 141; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[EXT_B:%.*]] = zext i32 [[LOAD_B]] to i6442; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[MUL:%.*]] = mul nuw i64 [[EXT_A]], [[EXT_B]]43; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[MUL_ZEXT:%.*]] = zext i64 [[MUL]] to i12844; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[ADD]] = add i128 [[ACCUM]], [[MUL_ZEXT]]45; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[IV_NEXT]] = add i64 [[IV]], 146; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], 402547; CHECK-NO-PARTIAL-REDUCTION-NEXT: br i1 [[EXITCOND_NOT]], label %[[FOR_EXIT:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]48; CHECK-NO-PARTIAL-REDUCTION: [[FOR_EXIT]]:49; CHECK-NO-PARTIAL-REDUCTION-NEXT: [[ADD_LCSSA:%.*]] = phi i128 [ [[ADD]], %[[FOR_BODY]] ]50; CHECK-NO-PARTIAL-REDUCTION-NEXT: ret i128 [[ADD_LCSSA]]51;52entry:53 br label %for.body54 55for.body:56 %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]57 %accum = phi i128 [ 0, %entry ], [ %add, %for.body ]58 %gep.a = getelementptr i32, ptr %a, i64 %iv59 %load.a = load i32, ptr %gep.a, align 160 %ext.a = zext i32 %load.a to i6461 %gep.b = getelementptr i32, ptr %b, i64 %iv62 %load.b = load i32, ptr %gep.b, align 163 %ext.b = zext i32 %load.b to i6464 %mul = mul nuw i64 %ext.a, %ext.b65 %mul.zext = zext i64 %mul to i12866 %add = add i128 %accum, %mul.zext67 %iv.next = add i64 %iv, 168 %exitcond.not = icmp eq i64 %iv.next, 402569 br i1 %exitcond.not, label %for.exit, label %for.body70 71for.exit:72 ret i128 %add73}74;.75; CHECK-NO-PARTIAL-REDUCTION: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}76; CHECK-NO-PARTIAL-REDUCTION: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}77; CHECK-NO-PARTIAL-REDUCTION: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}78; CHECK-NO-PARTIAL-REDUCTION: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}79;.80