[llvm] [LV] Convert scatter w/uniform addr and mask being header mask to scalar store. (PR #172799)

[Luke Lau via llvm-commits]

================
@@ -549,26 +549,26 @@ define void @uniform_store_of_loop_varying(ptr noalias nocapture %a, ptr noalias
 ; SCALABLE-NEXT:  [[ENTRY:.*:]]
 ; SCALABLE-NEXT:    br label %[[VECTOR_PH:.*]]
 ; SCALABLE:       [[VECTOR_PH]]:
-; SCALABLE-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[B]], i64 0
-; SCALABLE-NEXT:    [[BROADCAST_SPLAT1:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
 ; SCALABLE-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[V]], i64 0
 ; SCALABLE-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
-; SCALABLE-NEXT:    [[TMP6:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
+; SCALABLE-NEXT:    [[TMP0:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
 ; SCALABLE-NEXT:    br label %[[VECTOR_BODY:.*]]
 ; SCALABLE:       [[VECTOR_BODY]]:
 ; SCALABLE-NEXT:    [[TMP10:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_EVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; SCALABLE-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP6]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SCALABLE-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; SCALABLE-NEXT:    [[AVL:%.*]] = phi i64 [ 1025, %[[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; SCALABLE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 2, i1 true)
 ; SCALABLE-NEXT:    [[TMP8:%.*]] = zext i32 [[TMP7]] to i64
-; SCALABLE-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP8]], i64 0
-; SCALABLE-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[DOTSPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
-; SCALABLE-NEXT:    call void @llvm.vp.scatter.nxv2i64.nxv2p0(<vscale x 2 x i64> [[VEC_IND]], <vscale x 2 x ptr> align 8 [[BROADCAST_SPLAT1]], <vscale x 2 x i1> splat (i1 true), i32 [[TMP7]])
+; SCALABLE-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP8]], i64 0
+; SCALABLE-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT:    [[TMP3:%.*]] = sub i64 [[TMP8]], 1
+; SCALABLE-NEXT:    [[TMP4:%.*]] = extractelement <vscale x 2 x i64> [[VEC_IND]], i64 [[TMP3]]
----------------
lukel97 wrote:

Not for this PR, but I think this can be optimized further to narrow the value being stored. I.e. VEC_IND can just become a scalar VPDerivedIVRecipe. Then we don't have the cost of the extractelement.

https://github.com/llvm/llvm-project/pull/172799