[llvm] [LV] Avoid bailing on SCEVUnknown in uniformity-rew (PR #160144)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 26 08:52:33 PDT 2025
================
@@ -41,105 +140,117 @@ define i32 @chained_recurrences(i32 %x, i64 %y, ptr %src.1, i32 %z, ptr %src.2)
; DEFAULT-NEXT: [[RED_1:%.*]] = or i32 [[TMP74]], [[SUM_RED]]
; DEFAULT-NEXT: [[RED_2]] = or i32 [[RED_1]], [[TMP75]]
; DEFAULT-NEXT: [[EC:%.*]] = icmp eq i64 [[IV]], [[Y]]
-; DEFAULT-NEXT: br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; DEFAULT-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
; DEFAULT: [[EXIT]]:
-; DEFAULT-NEXT: [[RED_2_LCSSA:%.*]] = phi i32 [ [[RED_2]], %[[LOOP]] ]
+; DEFAULT-NEXT: [[RED_2_LCSSA:%.*]] = phi i32 [ [[RED_2]], %[[LOOP]] ], [ [[TMP64]], %[[MIDDLE_BLOCK]] ]
; DEFAULT-NEXT: ret i32 [[RED_2_LCSSA]]
;
; VSCALEFORTUNING2-LABEL: define i32 @chained_recurrences(
; VSCALEFORTUNING2-SAME: i32 [[X:%.*]], i64 [[Y:%.*]], ptr [[SRC_1:%.*]], i32 [[Z:%.*]], ptr [[SRC_2:%.*]]) #[[ATTR0:[0-9]+]] {
; VSCALEFORTUNING2-NEXT: [[ENTRY:.*]]:
; VSCALEFORTUNING2-NEXT: [[TMP0:%.*]] = add i64 [[Y]], 1
-; VSCALEFORTUNING2-NEXT: [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
-; VSCALEFORTUNING2-NEXT: [[TMP2:%.*]] = shl nuw i64 [[TMP1]], 3
-; VSCALEFORTUNING2-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], [[TMP2]]
+; VSCALEFORTUNING2-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 8
; VSCALEFORTUNING2-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
; VSCALEFORTUNING2: [[VECTOR_PH]]:
-; VSCALEFORTUNING2-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64()
-; VSCALEFORTUNING2-NEXT: [[TMP4:%.*]] = mul nuw i64 [[TMP3]], 8
-; VSCALEFORTUNING2-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], [[TMP4]]
+; VSCALEFORTUNING2-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 8
; VSCALEFORTUNING2-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Z]], i64 0
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[X]], i64 0
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[Z]], i64 0
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <4 x i32> poison, i32 [[X]], i64 0
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT1]], <4 x i32> poison, <4 x i32> zeroinitializer
; VSCALEFORTUNING2-NEXT: [[TMP7:%.*]] = add i64 [[Y]], 1
; VSCALEFORTUNING2-NEXT: [[TMP8:%.*]] = getelementptr i32, ptr [[SRC_1]], i64 [[TMP7]]
-; VSCALEFORTUNING2-NEXT: [[TMP9:%.*]] = lshr <vscale x 4 x i32> [[BROADCAST_SPLAT]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP10:%.*]] = shl <vscale x 4 x i32> [[BROADCAST_SPLAT]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP11:%.*]] = or <vscale x 4 x i32> [[TMP9]], [[TMP10]]
-; VSCALEFORTUNING2-NEXT: [[TMP12:%.*]] = or <vscale x 4 x i32> [[BROADCAST_SPLAT2]], [[BROADCAST_SPLAT]]
-; VSCALEFORTUNING2-NEXT: [[TMP13:%.*]] = and <vscale x 4 x i32> [[TMP12]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP14:%.*]] = xor <vscale x 4 x i32> [[TMP13]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP15:%.*]] = zext <vscale x 4 x i32> [[TMP14]] to <vscale x 4 x i64>
-; VSCALEFORTUNING2-NEXT: [[DOTSPLAT:%.*]] = getelementptr i32, ptr [[SRC_2]], <vscale x 4 x i64> [[TMP15]]
-; VSCALEFORTUNING2-NEXT: [[TMP18:%.*]] = call i32 @llvm.vscale.i32()
-; VSCALEFORTUNING2-NEXT: [[TMP19:%.*]] = mul nuw i32 [[TMP18]], 4
-; VSCALEFORTUNING2-NEXT: [[TMP20:%.*]] = sub i32 [[TMP19]], 1
-; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR_INIT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 0, i32 [[TMP20]]
-; VSCALEFORTUNING2-NEXT: [[TMP21:%.*]] = call i32 @llvm.vscale.i32()
-; VSCALEFORTUNING2-NEXT: [[TMP22:%.*]] = mul nuw i32 [[TMP21]], 4
-; VSCALEFORTUNING2-NEXT: [[TMP23:%.*]] = sub i32 [[TMP22]], 1
-; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR_INIT3:%.*]] = insertelement <vscale x 4 x i32> poison, i32 0, i32 [[TMP23]]
+; VSCALEFORTUNING2-NEXT: [[TMP3:%.*]] = lshr i32 [[X]], 1
+; VSCALEFORTUNING2-NEXT: [[TMP4:%.*]] = shl i32 [[X]], 1
+; VSCALEFORTUNING2-NEXT: [[TMP5:%.*]] = or i32 [[TMP3]], [[TMP4]]
+; VSCALEFORTUNING2-NEXT: [[TMP6:%.*]] = or <4 x i32> [[BROADCAST_SPLAT]], [[BROADCAST_SPLAT2]]
+; VSCALEFORTUNING2-NEXT: [[TMP18:%.*]] = and <4 x i32> [[TMP6]], splat (i32 1)
+; VSCALEFORTUNING2-NEXT: [[TMP66:%.*]] = xor <4 x i32> [[TMP18]], splat (i32 1)
+; VSCALEFORTUNING2-NEXT: [[TMP9:%.*]] = zext <4 x i32> [[TMP66]] to <4 x i64>
+; VSCALEFORTUNING2-NEXT: [[TMP10:%.*]] = extractelement <4 x i64> [[TMP9]], i32 0
+; VSCALEFORTUNING2-NEXT: [[TMP11:%.*]] = getelementptr i32, ptr [[SRC_2]], i64 [[TMP10]]
+; VSCALEFORTUNING2-NEXT: [[TMP12:%.*]] = extractelement <4 x i64> [[TMP9]], i32 1
+; VSCALEFORTUNING2-NEXT: [[TMP13:%.*]] = getelementptr i32, ptr [[SRC_2]], i64 [[TMP12]]
+; VSCALEFORTUNING2-NEXT: [[TMP14:%.*]] = extractelement <4 x i64> [[TMP9]], i32 2
+; VSCALEFORTUNING2-NEXT: [[TMP15:%.*]] = getelementptr i32, ptr [[SRC_2]], i64 [[TMP14]]
+; VSCALEFORTUNING2-NEXT: [[TMP16:%.*]] = extractelement <4 x i64> [[TMP9]], i32 3
+; VSCALEFORTUNING2-NEXT: [[TMP17:%.*]] = getelementptr i32, ptr [[SRC_2]], i64 [[TMP16]]
; VSCALEFORTUNING2-NEXT: br label %[[VECTOR_BODY:.*]]
; VSCALEFORTUNING2: [[VECTOR_BODY]]:
; VSCALEFORTUNING2-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR:%.*]] = phi <vscale x 4 x i32> [ [[VECTOR_RECUR_INIT]], %[[VECTOR_PH]] ], [ [[BROADCAST_SPLAT7:%.*]], %[[VECTOR_BODY]] ]
-; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR4:%.*]] = phi <vscale x 4 x i32> [ [[VECTOR_RECUR_INIT3]], %[[VECTOR_PH]] ], [ [[TMP26:%.*]], %[[VECTOR_BODY]] ]
-; VSCALEFORTUNING2-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP47:%.*]], %[[VECTOR_BODY]] ]
-; VSCALEFORTUNING2-NEXT: [[VEC_PHI5:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP48:%.*]], %[[VECTOR_BODY]] ]
+; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, %[[VECTOR_PH]] ], [ [[BROADCAST_SPLAT6:%.*]], %[[VECTOR_BODY]] ]
+; VSCALEFORTUNING2-NEXT: [[VECTOR_RECUR3:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, %[[VECTOR_PH]] ], [ [[TMP20:%.*]], %[[VECTOR_BODY]] ]
+; VSCALEFORTUNING2-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP67:%.*]], %[[VECTOR_BODY]] ]
+; VSCALEFORTUNING2-NEXT: [[VEC_PHI4:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP68:%.*]], %[[VECTOR_BODY]] ]
; VSCALEFORTUNING2-NEXT: [[TMP24:%.*]] = load i32, ptr [[TMP8]], align 4
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT6:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP24]], i64 0
-; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT7]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT6]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
-; VSCALEFORTUNING2-NEXT: [[TMP25:%.*]] = call <vscale x 4 x i32> @llvm.vector.splice.nxv4i32(<vscale x 4 x i32> [[VECTOR_RECUR]], <vscale x 4 x i32> [[BROADCAST_SPLAT7]], i32 -1)
-; VSCALEFORTUNING2-NEXT: [[TMP26]] = call <vscale x 4 x i32> @llvm.vector.splice.nxv4i32(<vscale x 4 x i32> [[BROADCAST_SPLAT7]], <vscale x 4 x i32> [[BROADCAST_SPLAT7]], i32 -1)
-; VSCALEFORTUNING2-NEXT: [[TMP27:%.*]] = call <vscale x 4 x i32> @llvm.vector.splice.nxv4i32(<vscale x 4 x i32> [[VECTOR_RECUR4]], <vscale x 4 x i32> [[TMP25]], i32 -1)
-; VSCALEFORTUNING2-NEXT: [[TMP28:%.*]] = call <vscale x 4 x i32> @llvm.vector.splice.nxv4i32(<vscale x 4 x i32> [[TMP25]], <vscale x 4 x i32> [[TMP26]], i32 -1)
-; VSCALEFORTUNING2-NEXT: [[TMP29:%.*]] = or <vscale x 4 x i32> [[TMP27]], [[BROADCAST_SPLAT]]
-; VSCALEFORTUNING2-NEXT: [[TMP30:%.*]] = or <vscale x 4 x i32> [[TMP28]], [[BROADCAST_SPLAT]]
-; VSCALEFORTUNING2-NEXT: [[TMP31:%.*]] = shl <vscale x 4 x i32> [[TMP29]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP32:%.*]] = shl <vscale x 4 x i32> [[TMP30]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP33:%.*]] = or <vscale x 4 x i32> [[TMP31]], splat (i32 2)
-; VSCALEFORTUNING2-NEXT: [[TMP34:%.*]] = or <vscale x 4 x i32> [[TMP32]], splat (i32 2)
-; VSCALEFORTUNING2-NEXT: [[TMP35:%.*]] = or <vscale x 4 x i32> [[TMP11]], [[TMP33]]
-; VSCALEFORTUNING2-NEXT: [[TMP36:%.*]] = or <vscale x 4 x i32> [[TMP11]], [[TMP34]]
-; VSCALEFORTUNING2-NEXT: [[TMP37:%.*]] = or <vscale x 4 x i32> [[TMP35]], [[BROADCAST_SPLAT]]
-; VSCALEFORTUNING2-NEXT: [[TMP38:%.*]] = or <vscale x 4 x i32> [[TMP36]], [[BROADCAST_SPLAT]]
-; VSCALEFORTUNING2-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[DOTSPLAT]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
-; VSCALEFORTUNING2-NEXT: [[WIDE_MASKED_GATHER8:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[DOTSPLAT]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
-; VSCALEFORTUNING2-NEXT: [[TMP39:%.*]] = lshr <vscale x 4 x i32> [[TMP37]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP40:%.*]] = lshr <vscale x 4 x i32> [[TMP38]], splat (i32 1)
-; VSCALEFORTUNING2-NEXT: [[TMP41:%.*]] = zext <vscale x 4 x i32> [[TMP39]] to <vscale x 4 x i64>
-; VSCALEFORTUNING2-NEXT: [[TMP42:%.*]] = zext <vscale x 4 x i32> [[TMP40]] to <vscale x 4 x i64>
-; VSCALEFORTUNING2-NEXT: [[TMP43:%.*]] = getelementptr i32, ptr [[SRC_2]], <vscale x 4 x i64> [[TMP41]]
-; VSCALEFORTUNING2-NEXT: [[TMP44:%.*]] = getelementptr i32, ptr [[SRC_2]], <vscale x 4 x i64> [[TMP42]]
-; VSCALEFORTUNING2-NEXT: [[WIDE_MASKED_GATHER9:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP43]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
-; VSCALEFORTUNING2-NEXT: [[WIDE_MASKED_GATHER10:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP44]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
-; VSCALEFORTUNING2-NEXT: [[TMP45:%.*]] = or <vscale x 4 x i32> [[WIDE_MASKED_GATHER]], [[VEC_PHI]]
-; VSCALEFORTUNING2-NEXT: [[TMP46:%.*]] = or <vscale x 4 x i32> [[WIDE_MASKED_GATHER8]], [[VEC_PHI5]]
-; VSCALEFORTUNING2-NEXT: [[TMP47]] = or <vscale x 4 x i32> [[TMP45]], [[WIDE_MASKED_GATHER9]]
-; VSCALEFORTUNING2-NEXT: [[TMP48]] = or <vscale x 4 x i32> [[TMP46]], [[WIDE_MASKED_GATHER10]]
-; VSCALEFORTUNING2-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]]
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <4 x i32> poison, i32 [[TMP24]], i64 0
+; VSCALEFORTUNING2-NEXT: [[BROADCAST_SPLAT6]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT5]], <4 x i32> poison, <4 x i32> zeroinitializer
+; VSCALEFORTUNING2-NEXT: [[TMP19:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[BROADCAST_SPLAT6]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; VSCALEFORTUNING2-NEXT: [[TMP20]] = shufflevector <4 x i32> [[BROADCAST_SPLAT6]], <4 x i32> [[BROADCAST_SPLAT6]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; VSCALEFORTUNING2-NEXT: [[TMP21:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR3]], <4 x i32> [[TMP19]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; VSCALEFORTUNING2-NEXT: [[TMP22:%.*]] = shufflevector <4 x i32> [[TMP19]], <4 x i32> [[TMP20]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; VSCALEFORTUNING2-NEXT: [[TMP23:%.*]] = extractelement <4 x i32> [[TMP21]], i32 0
+; VSCALEFORTUNING2-NEXT: [[TMP69:%.*]] = or i32 [[TMP23]], [[X]]
+; VSCALEFORTUNING2-NEXT: [[TMP25:%.*]] = extractelement <4 x i32> [[TMP22]], i32 0
+; VSCALEFORTUNING2-NEXT: [[TMP26:%.*]] = or i32 [[TMP25]], [[X]]
+; VSCALEFORTUNING2-NEXT: [[TMP27:%.*]] = shl i32 [[TMP69]], 1
+; VSCALEFORTUNING2-NEXT: [[TMP28:%.*]] = shl i32 [[TMP26]], 1
+; VSCALEFORTUNING2-NEXT: [[TMP29:%.*]] = or i32 [[TMP27]], 2
+; VSCALEFORTUNING2-NEXT: [[TMP30:%.*]] = or i32 [[TMP28]], 2
+; VSCALEFORTUNING2-NEXT: [[TMP31:%.*]] = or i32 [[TMP5]], [[TMP29]]
+; VSCALEFORTUNING2-NEXT: [[TMP32:%.*]] = or i32 [[TMP5]], [[TMP30]]
+; VSCALEFORTUNING2-NEXT: [[TMP33:%.*]] = or i32 [[TMP31]], [[X]]
+; VSCALEFORTUNING2-NEXT: [[TMP34:%.*]] = or i32 [[TMP32]], [[X]]
+; VSCALEFORTUNING2-NEXT: [[TMP35:%.*]] = load i32, ptr [[TMP11]], align 4
+; VSCALEFORTUNING2-NEXT: [[TMP36:%.*]] = load i32, ptr [[TMP13]], align 4
+; VSCALEFORTUNING2-NEXT: [[TMP37:%.*]] = load i32, ptr [[TMP15]], align 4
+; VSCALEFORTUNING2-NEXT: [[TMP38:%.*]] = load i32, ptr [[TMP17]], align 4
+; VSCALEFORTUNING2-NEXT: [[TMP39:%.*]] = insertelement <4 x i32> poison, i32 [[TMP35]], i32 0
----------------
artagnon wrote:
The pointer operand of the GEP is an argument, which is an SCEVUnknown: this means that various instructions could be considered uniform in the chain and narrowed to scalar operations; I think the net result of vscale versus no-vscale vectorization was a cost-model decision? I'm not sure if the vectorization is worse overall, since many binops are operating on scalars?
https://github.com/llvm/llvm-project/pull/160144
More information about the llvm-commits
mailing list