[llvm-branch-commits] [llvm] [LoopVectorizer] Bundle partial reductions inside VPMulAccumulateReductionRecipe (PR #136173)

[Sam Tebbs via llvm-branch-commits]

================
@@ -253,38 +253,38 @@ define i64 @not_dotp_i8_to_i64_has_neon_dotprod(ptr readonly %a, ptr readonly %b
 ; CHECK-MAXBW-SAME: ptr readonly [[A:%.*]], ptr readonly [[B:%.*]]) #[[ATTR1:[0-9]+]] {
 ; CHECK-MAXBW-NEXT:  entry:
 ; CHECK-MAXBW-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-MAXBW-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 8
+; CHECK-MAXBW-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 16
 ; CHECK-MAXBW-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK-MAXBW:       vector.ph:
 ; CHECK-MAXBW-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-MAXBW-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 8
+; CHECK-MAXBW-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 16
 ; CHECK-MAXBW-NEXT:    [[N_MOD_VF:%.*]] = urem i64 1024, [[TMP3]]
 ; CHECK-MAXBW-NEXT:    [[N_VEC:%.*]] = sub i64 1024, [[N_MOD_VF]]
 ; CHECK-MAXBW-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-MAXBW-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 8
+; CHECK-MAXBW-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 16
 ; CHECK-MAXBW-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[A]], i64 [[N_VEC]]
 ; CHECK-MAXBW-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[B]], i64 [[N_VEC]]
 ; CHECK-MAXBW-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK-MAXBW:       vector.body:
 ; CHECK-MAXBW-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-MAXBW-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 8 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
+; CHECK-MAXBW-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-MAXBW-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-MAXBW-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP8]]
 ; CHECK-MAXBW-NEXT:    [[TMP9:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-MAXBW-NEXT:    [[NEXT_GEP1:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP9]]
 ; CHECK-MAXBW-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i32 0
-; CHECK-MAXBW-NEXT:    [[WIDE_LOAD:%.*]] = load <vscale x 8 x i8>, ptr [[TMP10]], align 1
-; CHECK-MAXBW-NEXT:    [[TMP11:%.*]] = zext <vscale x 8 x i8> [[WIDE_LOAD]] to <vscale x 8 x i64>
+; CHECK-MAXBW-NEXT:    [[WIDE_LOAD:%.*]] = load <vscale x 16 x i8>, ptr [[TMP10]], align 1
 ; CHECK-MAXBW-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[NEXT_GEP1]], i32 0
-; CHECK-MAXBW-NEXT:    [[WIDE_LOAD2:%.*]] = load <vscale x 8 x i8>, ptr [[TMP12]], align 1
-; CHECK-MAXBW-NEXT:    [[TMP13:%.*]] = zext <vscale x 8 x i8> [[WIDE_LOAD2]] to <vscale x 8 x i64>
-; CHECK-MAXBW-NEXT:    [[TMP14:%.*]] = mul nuw nsw <vscale x 8 x i64> [[TMP13]], [[TMP11]]
-; CHECK-MAXBW-NEXT:    [[TMP15]] = add <vscale x 8 x i64> [[TMP14]], [[VEC_PHI]]
+; CHECK-MAXBW-NEXT:    [[WIDE_LOAD2:%.*]] = load <vscale x 16 x i8>, ptr [[TMP12]], align 1
+; CHECK-MAXBW-NEXT:    [[TMP15:%.*]] = zext <vscale x 16 x i8> [[WIDE_LOAD2]] to <vscale x 16 x i64>
+; CHECK-MAXBW-NEXT:    [[TMP13:%.*]] = zext <vscale x 16 x i8> [[WIDE_LOAD]] to <vscale x 16 x i64>
+; CHECK-MAXBW-NEXT:    [[TMP14:%.*]] = mul nuw nsw <vscale x 16 x i64> [[TMP15]], [[TMP13]]
+; CHECK-MAXBW-NEXT:    [[PARTIAL_REDUCE]] = call <vscale x 2 x i64> @llvm.experimental.vector.partial.reduce.add.nxv2i64.nxv16i64(<vscale x 2 x i64> [[VEC_PHI]], <vscale x 16 x i64> [[TMP14]])
----------------
SamTebbs33 wrote:

Ah it looks like what was previously too high a cost for it to choose a 16i8 -> 2i64 partial reduction isn't sufficiently high now that the extend cost is hidden. I've made this permutation invalid.

https://github.com/llvm/llvm-project/pull/136173