[llvm] [RISCV] Improve getInterleavedMemoryOpCost for interleave groups with tail gaps. (PR #192074)
Mel Chen via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 16 06:54:39 PDT 2026
https://github.com/Mel-Chen updated https://github.com/llvm/llvm-project/pull/192074
>From db5bf8bf1ddfd00fe73a96753f4b3fb099497fee Mon Sep 17 00:00:00 2001
From: Mel Chen <mel.chen at sifive.com>
Date: Thu, 16 Apr 2026 06:12:31 -0700
Subject: [PATCH 1/2] nfc, fix the test
---
.../RISCV/tail-folding-interleave.ll | 87 +++++++++----------
1 file changed, 40 insertions(+), 47 deletions(-)
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll b/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
index aeab26ebc9f44..dc29e32d29c19 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
@@ -118,27 +118,26 @@ define i32 @load_factor_4_with_gap(i64 %n, ptr noalias %a) {
; IF-EVL-NEXT: entry:
; IF-EVL-NEXT: br label [[VECTOR_PH:%.*]]
; IF-EVL: vector.ph:
-; IF-EVL-NEXT: [[TMP2:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
; IF-EVL-NEXT: br label [[VECTOR_BODY:%.*]]
; IF-EVL: vector.body:
-; IF-EVL-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP2]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
+; IF-EVL-NEXT: [[CURRENT_ITERATION_IV:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[CURRENT_ITERATION_NEXT:%.*]], [[VECTOR_BODY]] ]
; IF-EVL-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP12:%.*]], [[VECTOR_BODY]] ]
; IF-EVL-NEXT: [[AVL:%.*]] = phi i64 [ [[N:%.*]], [[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], [[VECTOR_BODY]] ]
; IF-EVL-NEXT: [[TMP4:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 4, i1 true)
-; IF-EVL-NEXT: [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
-; IF-EVL-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP5]], i64 0
-; IF-EVL-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
-; IF-EVL-NEXT: [[TMP7:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
-; IF-EVL-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP7]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
+; IF-EVL-NEXT: [[TMP1:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], i64 [[CURRENT_ITERATION_IV]], i32 0
+; IF-EVL-NEXT: [[INTERLEAVE_EVL:%.*]] = mul nuw nsw i32 [[TMP4]], 4
+; IF-EVL-NEXT: [[WIDE_VP_LOAD:%.*]] = call <vscale x 16 x i32> @llvm.vp.load.nxv16i32.p0(ptr align 4 [[TMP1]], <vscale x 16 x i1> splat (i1 true), i32 [[INTERLEAVE_EVL]])
+; IF-EVL-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VP_LOAD]])
+; IF-EVL-NEXT: [[WIDE_MASKED_GATHER:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; IF-EVL-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; IF-EVL-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; IF-EVL-NEXT: [[TMP8:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; IF-EVL-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP7]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
; IF-EVL-NEXT: [[TMP9:%.*]] = add <vscale x 4 x i32> [[TMP8]], [[WIDE_MASKED_GATHER1]]
-; IF-EVL-NEXT: [[TMP10:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 3
-; IF-EVL-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP10]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
; IF-EVL-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i32> [[TMP9]], [[WIDE_MASKED_GATHER2]]
; IF-EVL-NEXT: [[TMP12]] = call <vscale x 4 x i32> @llvm.vp.merge.nxv4i32(<vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> [[TMP11]], <vscale x 4 x i32> [[VEC_PHI]], i32 [[TMP4]])
+; IF-EVL-NEXT: [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
+; IF-EVL-NEXT: [[CURRENT_ITERATION_NEXT]] = add i64 [[TMP5]], [[CURRENT_ITERATION_IV]]
; IF-EVL-NEXT: [[AVL_NEXT]] = sub nuw i64 [[AVL]], [[TMP5]]
-; IF-EVL-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
; IF-EVL-NEXT: [[TMP14:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
; IF-EVL-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
; IF-EVL: middle.block:
@@ -151,38 +150,33 @@ define i32 @load_factor_4_with_gap(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: entry:
; NO-VP-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
; NO-VP-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 2
-; NO-VP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 [[N:%.*]], [[TMP1]]
+; NO-VP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]
; NO-VP-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
; NO-VP: vector.ph:
; NO-VP-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
; NO-VP-NEXT: [[TMP3:%.*]] = shl nuw i64 [[TMP2]], 2
; NO-VP-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
-; NO-VP-NEXT: [[TMP4:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
-; NO-VP-NEXT: [[TMP5:%.*]] = select i1 [[TMP4]], i64 [[TMP3]], i64 [[N_MOD_VF]]
-; NO-VP-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[TMP5]]
-; NO-VP-NEXT: [[TMP6:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
-; NO-VP-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP3]], i64 0
-; NO-VP-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; NO-VP-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
; NO-VP-NEXT: br label [[VECTOR_BODY:%.*]]
; NO-VP: vector.body:
; NO-VP-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP6]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
; NO-VP-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[TMP9:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
+; NO-VP-NEXT: [[TMP4:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], i64 [[INDEX]], i32 0
+; NO-VP-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP4]], align 4
+; NO-VP-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 3
; NO-VP-NEXT: [[TMP10:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i32> [[TMP10]], [[WIDE_MASKED_GATHER1]]
-; NO-VP-NEXT: [[TMP12:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 3
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP12]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP13]] = add <vscale x 4 x i32> [[TMP11]], [[WIDE_MASKED_GATHER2]]
; NO-VP-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
-; NO-VP-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 4 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
; NO-VP-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; NO-VP-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; NO-VP: middle.block:
; NO-VP-NEXT: [[TMP15:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP13]])
-; NO-VP-NEXT: br label [[SCALAR_PH]]
+; NO-VP-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; NO-VP-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
; NO-VP: scalar.ph:
; NO-VP-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
; NO-VP-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP15]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
@@ -194,16 +188,16 @@ define i32 @load_factor_4_with_gap(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: [[TMP16:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; NO-VP-NEXT: [[ADD:%.*]] = add nsw i32 [[RDX]], [[TMP16]]
; NO-VP-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 1
-; NO-VP-NEXT: [[TMP17:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
+; NO-VP-NEXT: [[TMP17:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
; NO-VP-NEXT: [[ADD1:%.*]] = add nsw i32 [[ADD]], [[TMP17]]
; NO-VP-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 3
; NO-VP-NEXT: [[TMP18:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; NO-VP-NEXT: [[ADD2]] = add nsw i32 [[ADD1]], [[TMP18]]
; NO-VP-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
; NO-VP-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
-; NO-VP-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
+; NO-VP-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
; NO-VP: exit:
-; NO-VP-NEXT: [[ADD2_LCSSA:%.*]] = phi i32 [ [[ADD2]], [[FOR_BODY]] ]
+; NO-VP-NEXT: [[ADD2_LCSSA:%.*]] = phi i32 [ [[ADD2]], [[FOR_BODY]] ], [ [[TMP15]], [[MIDDLE_BLOCK]] ]
; NO-VP-NEXT: ret i32 [[ADD2_LCSSA]]
;
entry:
@@ -216,7 +210,7 @@ for.body:
%0 = load i32, ptr %arrayidx, align 4
%add = add nsw i32 %rdx, %0
%arrayidx1 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 1
- %1 = load i32, ptr %arrayidx, align 4
+ %1 = load i32, ptr %arrayidx1, align 4
%add1 = add nsw i32 %add, %1
%arrayidx2 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 3
%2 = load i32, ptr %arrayidx2, align 4
@@ -363,7 +357,8 @@ define i32 @load_factor_4_with_tail_gap(i64 %n, ptr noalias %a) {
; IF-EVL-NEXT: [[TMP7:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
; IF-EVL-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP7]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
; IF-EVL-NEXT: [[TMP8:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; IF-EVL-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP7]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
+; IF-EVL-NEXT: [[TMP6:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 1
+; IF-EVL-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP6]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
; IF-EVL-NEXT: [[TMP9:%.*]] = add <vscale x 4 x i32> [[TMP8]], [[WIDE_MASKED_GATHER1]]
; IF-EVL-NEXT: [[TMP10:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 2
; IF-EVL-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP10]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP4]])
@@ -392,24 +387,20 @@ define i32 @load_factor_4_with_tail_gap(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: [[TMP4:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
; NO-VP-NEXT: [[TMP5:%.*]] = select i1 [[TMP4]], i64 [[TMP3]], i64 [[N_MOD_VF]]
; NO-VP-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[TMP5]]
-; NO-VP-NEXT: [[TMP6:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
-; NO-VP-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP3]], i64 0
-; NO-VP-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
; NO-VP-NEXT: br label [[VECTOR_BODY:%.*]]
; NO-VP: vector.body:
; NO-VP-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP6]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
; NO-VP-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[TMP9:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
+; NO-VP-NEXT: [[TMP6:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], i64 [[INDEX]], i32 0
+; NO-VP-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 16 x i32>, ptr [[TMP6]], align 4
+; NO-VP-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave4.nxv16i32(<vscale x 16 x i32> [[WIDE_VEC]])
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 2
; NO-VP-NEXT: [[TMP10:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i32> [[TMP10]], [[WIDE_MASKED_GATHER1]]
-; NO-VP-NEXT: [[TMP12:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 2
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP12]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP13]] = add <vscale x 4 x i32> [[TMP11]], [[WIDE_MASKED_GATHER2]]
; NO-VP-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
-; NO-VP-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 4 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
; NO-VP-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; NO-VP-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
; NO-VP: middle.block:
@@ -426,7 +417,7 @@ define i32 @load_factor_4_with_tail_gap(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: [[TMP16:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; NO-VP-NEXT: [[ADD:%.*]] = add nsw i32 [[RDX]], [[TMP16]]
; NO-VP-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 1
-; NO-VP-NEXT: [[TMP17:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
+; NO-VP-NEXT: [[TMP17:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
; NO-VP-NEXT: [[ADD1:%.*]] = add nsw i32 [[ADD]], [[TMP17]]
; NO-VP-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 2
; NO-VP-NEXT: [[TMP18:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
@@ -448,7 +439,7 @@ for.body:
%0 = load i32, ptr %arrayidx, align 4
%add = add nsw i32 %rdx, %0
%arrayidx1 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 1
- %1 = load i32, ptr %arrayidx, align 4
+ %1 = load i32, ptr %arrayidx1, align 4
%add1 = add nsw i32 %add, %1
%arrayidx2 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 2
%2 = load i32, ptr %arrayidx2, align 4
@@ -593,7 +584,8 @@ define i32 @load_factor_4_reverse(i64 %n, ptr noalias %a) {
; IF-EVL-NEXT: [[TMP9:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
; IF-EVL-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP6]])
; IF-EVL-NEXT: [[TMP10:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; IF-EVL-NEXT: [[WIDE_MASKED_GATHER3:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP9]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP6]])
+; IF-EVL-NEXT: [[TMP17:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 1
+; IF-EVL-NEXT: [[WIDE_MASKED_GATHER3:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP17]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP6]])
; IF-EVL-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i32> [[TMP10]], [[WIDE_MASKED_GATHER3]]
; IF-EVL-NEXT: [[TMP12:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 2
; IF-EVL-NEXT: [[WIDE_MASKED_GATHER4:%.*]] = call <vscale x 4 x i32> @llvm.vp.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP12]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP6]])
@@ -643,7 +635,8 @@ define i32 @load_factor_4_reverse(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: [[TMP10:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
; NO-VP-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP10]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP11:%.*]] = add <vscale x 4 x i32> [[VEC_PHI]], [[WIDE_MASKED_GATHER]]
-; NO-VP-NEXT: [[WIDE_MASKED_GATHER3:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP10]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
+; NO-VP-NEXT: [[TMP23:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 1
+; NO-VP-NEXT: [[WIDE_MASKED_GATHER3:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP23]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
; NO-VP-NEXT: [[TMP12:%.*]] = add <vscale x 4 x i32> [[TMP11]], [[WIDE_MASKED_GATHER3]]
; NO-VP-NEXT: [[TMP13:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i64> [[VEC_IND]], i32 2
; NO-VP-NEXT: [[WIDE_MASKED_GATHER4:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> align 4 [[TMP13]], <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
@@ -670,7 +663,7 @@ define i32 @load_factor_4_reverse(i64 %n, ptr noalias %a) {
; NO-VP-NEXT: [[TMP19:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; NO-VP-NEXT: [[ADD:%.*]] = add nsw i32 [[RDX]], [[TMP19]]
; NO-VP-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 1
-; NO-VP-NEXT: [[TMP20:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
+; NO-VP-NEXT: [[TMP20:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
; NO-VP-NEXT: [[ADD1:%.*]] = add nsw i32 [[ADD]], [[TMP20]]
; NO-VP-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], i64 [[IV]], i32 2
; NO-VP-NEXT: [[TMP21:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
@@ -695,7 +688,7 @@ for.body:
%0 = load i32, ptr %arrayidx, align 4
%add = add nsw i32 %rdx, %0
%arrayidx1 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 1
- %1 = load i32, ptr %arrayidx, align 4
+ %1 = load i32, ptr %arrayidx1, align 4
%add1 = add nsw i32 %add, %1
%arrayidx2 = getelementptr inbounds [4 x i32], ptr %a, i64 %iv, i32 2
%2 = load i32, ptr %arrayidx2, align 4
>From 0af6b10a5aa23deb59478e66bf060de88024866e Mon Sep 17 00:00:00 2001
From: Mel Chen <mel.chen at sifive.com>
Date: Tue, 14 Apr 2026 07:33:17 -0700
Subject: [PATCH 2/2] strided seg store
---
.../Target/RISCV/RISCVTargetTransformInfo.cpp | 28 ++++++++++++++++++
.../RISCV/tail-folding-interleave.ll | 29 +++++++------------
2 files changed, 39 insertions(+), 18 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
index fd762f35124b5..9e20bd8d2b428 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
@@ -1143,6 +1143,34 @@ InstructionCost RISCVTTIImpl::getInterleavedMemoryOpCost(
return InstructionCost::getInvalid();
auto *FVTy = cast<FixedVectorType>(VecTy);
+ // When gaps are only at the tail, for interleaved load, we can emit a wide
+ // masked load and shufflevectors. For interleaved store, we can emit
+ // shufflevectors and a wide masked store. The interleaved memory access pass
+ // will lower them into vlsseg/vssseg intrinsics.
+ if (UseMaskForGaps) {
+ unsigned NumOfFields = Indices.size();
+ bool IsTailGapOnly = NumOfFields > 1 && (NumOfFields == Indices.back() + 1);
+ if (IsTailGapOnly &&
+ NumOfFields <= TLI->getMaxSupportedInterleaveFactor()) {
+ std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(FVTy);
+ if (LT.second.isVector() &&
+ FVTy->getElementCount().isKnownMultipleOf(Factor)) {
+ auto *SubVecTy = VectorType::get(
+ FVTy->getElementType(),
+ FVTy->getElementCount().divideCoefficientBy(Factor));
+ if (TLI->isLegalInterleavedAccessType(SubVecTy, NumOfFields, Alignment,
+ AddressSpace, DL)) {
+ // The cost is proportional to the total number of element accesses.
+ InstructionCost MemOpCost =
+ getMemoryOpCost(Opcode, FVTy->getElementType(), Alignment, 0,
+ CostKind, {TTI::OK_AnyValue, TTI::OP_None});
+ unsigned NumAccesses = getEstimatedVLFor(FVTy);
+ return NumAccesses * MemOpCost;
+ }
+ }
+ }
+ }
+
InstructionCost MemCost =
getMemoryOpCost(Opcode, VecTy, Alignment, AddressSpace, CostKind);
unsigned VF = FVTy->getNumElements() / Factor;
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll b/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
index dc29e32d29c19..5f9d8c885bbe3 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/tail-folding-interleave.ll
@@ -493,30 +493,23 @@ define void @store_factor_4_with_tail_gap(i32 %n, ptr noalias %a) {
;
; NO-VP-LABEL: @store_factor_4_with_tail_gap(
; NO-VP-NEXT: entry:
-; NO-VP-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()
-; NO-VP-NEXT: [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 2
-; NO-VP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], [[TMP1]]
+; NO-VP-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 8
; NO-VP-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
; NO-VP: vector.ph:
-; NO-VP-NEXT: [[TMP2:%.*]] = call i32 @llvm.vscale.i32()
-; NO-VP-NEXT: [[TMP9:%.*]] = shl nuw i32 [[TMP2]], 2
-; NO-VP-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], [[TMP9]]
+; NO-VP-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 8
; NO-VP-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
-; NO-VP-NEXT: [[TMP7:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()
-; NO-VP-NEXT: [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP9]], i64 0
-; NO-VP-NEXT: [[BROADCAST_SPLAT3:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT2]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
; NO-VP-NEXT: br label [[VECTOR_BODY:%.*]]
; NO-VP: vector.body:
; NO-VP-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[VEC_IND4:%.*]] = phi <vscale x 4 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT5:%.*]], [[VECTOR_BODY]] ]
-; NO-VP-NEXT: [[TMP11:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], <vscale x 4 x i32> [[VEC_IND4]], i32 0
-; NO-VP-NEXT: call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[VEC_IND4]], <vscale x 4 x ptr> align 4 [[TMP11]], <vscale x 4 x i1> splat (i1 true))
-; NO-VP-NEXT: [[TMP12:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i32> [[VEC_IND4]], i32 1
-; NO-VP-NEXT: call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[VEC_IND4]], <vscale x 4 x ptr> align 4 [[TMP12]], <vscale x 4 x i1> splat (i1 true))
-; NO-VP-NEXT: [[TMP13:%.*]] = getelementptr inbounds [4 x i32], ptr [[A]], <vscale x 4 x i32> [[VEC_IND4]], i32 2
-; NO-VP-NEXT: call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[VEC_IND4]], <vscale x 4 x ptr> align 4 [[TMP13]], <vscale x 4 x i1> splat (i1 true))
-; NO-VP-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP9]]
-; NO-VP-NEXT: [[VEC_IND_NEXT5]] = add nuw nsw <vscale x 4 x i32> [[VEC_IND4]], [[BROADCAST_SPLAT3]]
+; NO-VP-NEXT: [[VEC_IND:%.*]] = phi <8 x i32> [ <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
+; NO-VP-NEXT: [[TMP0:%.*]] = getelementptr inbounds [4 x i32], ptr [[A:%.*]], i32 [[INDEX]], i32 0
+; NO-VP-NEXT: [[TMP1:%.*]] = shufflevector <8 x i32> [[VEC_IND]], <8 x i32> [[VEC_IND]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+; NO-VP-NEXT: [[TMP2:%.*]] = shufflevector <8 x i32> [[VEC_IND]], <8 x i32> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+; NO-VP-NEXT: [[TMP3:%.*]] = shufflevector <16 x i32> [[TMP1]], <16 x i32> [[TMP2]], <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
+; NO-VP-NEXT: [[INTERLEAVED_VEC:%.*]] = shufflevector <32 x i32> [[TMP3]], <32 x i32> poison, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 1, i32 9, i32 17, i32 25, i32 2, i32 10, i32 18, i32 26, i32 3, i32 11, i32 19, i32 27, i32 4, i32 12, i32 20, i32 28, i32 5, i32 13, i32 21, i32 29, i32 6, i32 14, i32 22, i32 30, i32 7, i32 15, i32 23, i32 31>
+; NO-VP-NEXT: call void @llvm.masked.store.v32i32.p0(<32 x i32> [[INTERLEAVED_VEC]], ptr align 4 [[TMP0]], <32 x i1> <i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 false>)
+; NO-VP-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
+; NO-VP-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <8 x i32> [[VEC_IND]], splat (i32 8)
; NO-VP-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; NO-VP-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
; NO-VP: middle.block:
More information about the llvm-commits
mailing list