[llvm] [VPlan] Lower safe uniform load to unconditional scalar load (Alt. approach) (PR #217564)
Shih-Po Hung via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 20 02:56:41 PDT 2026
https://github.com/arcbbb created https://github.com/llvm/llvm-project/pull/217564
When a load's address is uniform across VFs/UFs and provably safe to
access unconditionally, lower it to an unconditional scalar load recipe.
This avoids generating unnecessary masked loads or gathers for uniform memory
accesses.
Alternative approach to #216767.
>From 7dbe19f0e547152e839f9fc19ed238f0be585ebe Mon Sep 17 00:00:00 2001
From: ShihPo Hung <shihpo.hung at sifive.com>
Date: Mon, 17 Aug 2026 01:58:51 -0700
Subject: [PATCH 1/2] [VPlan] precommit uniform gather test
---
.../LoopVectorize/uniform-gather.ll | 411 ++++++++++++++++++
1 file changed, 411 insertions(+)
create mode 100644 llvm/test/Transforms/LoopVectorize/uniform-gather.ll
diff --git a/llvm/test/Transforms/LoopVectorize/uniform-gather.ll b/llvm/test/Transforms/LoopVectorize/uniform-gather.ll
new file mode 100644
index 0000000000000..61af4c533d69a
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/uniform-gather.ll
@@ -0,0 +1,411 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -S -passes=loop-vectorize -force-vector-width=2 -force-target-supports-gather-scatter-ops=true -tail-folding-policy=must-fold-tail %s 2>&1 | FileCheck %s --check-prefix=FIXED
+; RUN: opt -S -passes=loop-vectorize -force-vector-width=2 -force-target-supports-gather-scatter-ops=true -tail-folding-policy=must-fold-tail -force-target-supports-scalable-vectors -scalable-vectorization=on %s 2>&1 | FileCheck %s --check-prefix=SCALABLE
+
+ at log_table = internal constant [4 x i32] [i32 10, i32 20, i32 30, i32 40]
+ at exp_table = internal constant [4 x i32] [i32 1, i32 2, i32 3, i32 4]
+
+; Uniform, dereferenceable global variable.
+define void @global_table_load(ptr noalias %out, ptr noalias %cond, i64 %n) {
+; FIXED-LABEL: define void @global_table_load(
+; FIXED-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], i64 [[N:%.*]]) {
+; FIXED-NEXT: [[ENTRY:.*:]]
+; FIXED-NEXT: br label %[[VECTOR_PH1:.*]]
+; FIXED: [[VECTOR_PH1]]:
+; FIXED-NEXT: [[N_RND_UP:%.*]] = add i64 [[N]], 1
+; FIXED-NEXT: [[TMP0:%.*]] = and i64 [[N_RND_UP]], 1
+; FIXED-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[TMP0]]
+; FIXED-NEXT: [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i64 0
+; FIXED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i64> [[BROADCAST_SPLATINSERT]], <2 x i64> poison, <2 x i32> zeroinitializer
+; FIXED-NEXT: br label %[[VECTOR_BODY1:.*]]
+; FIXED: [[VECTOR_BODY1]]:
+; FIXED-NEXT: [[INDEX1:%.*]] = phi i64 [ 0, %[[VECTOR_PH1]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY1]] ]
+; FIXED-NEXT: [[VEC_IND:%.*]] = phi <2 x i64> [ <i64 0, i64 1>, %[[VECTOR_PH1]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY1]] ]
+; FIXED-NEXT: [[TMP1:%.*]] = icmp ule <2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
+; FIXED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX1]]
+; FIXED-NEXT: [[WIDE_LOAD:%.*]] = call <2 x i32> @llvm.masked.load.v2i32.p0(ptr align 4 [[TMP40]], <2 x i1> [[TMP1]], <2 x i32> poison)
+; FIXED-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
+; FIXED-NEXT: [[TMP3:%.*]] = xor <2 x i1> [[TMP2]], splat (i1 true)
+; FIXED-NEXT: [[TMP7:%.*]] = select <2 x i1> [[TMP1]], <2 x i1> [[TMP3]], <2 x i1> zeroinitializer
+; FIXED-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 <ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4)>, <2 x i1> [[TMP7]], <2 x i32> poison)
+; FIXED-NEXT: [[TMP6:%.*]] = select <2 x i1> [[TMP1]], <2 x i1> [[TMP2]], <2 x i1> zeroinitializer
+; FIXED-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 <ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8)>, <2 x i1> [[TMP6]], <2 x i32> poison)
+; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP2]], <2 x i32> [[WIDE_MASKED_GATHER1]], <2 x i32> [[WIDE_MASKED_GATHER]]
+; FIXED-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX1]]
+; FIXED-NEXT: call void @llvm.masked.store.v2i32.p0(<2 x i32> [[PREDPHI]], ptr align 4 [[TMP4]], <2 x i1> [[TMP1]])
+; FIXED-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX1]], 2
+; FIXED-NEXT: [[VEC_IND_NEXT]] = add nuw <2 x i64> [[VEC_IND]], splat (i64 2)
+; FIXED-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; FIXED-NEXT: br i1 [[TMP5]], label %[[PRED_STORE_IF27:.*]], label %[[VECTOR_BODY1]], !llvm.loop [[LOOP0:![0-9]+]]
+; FIXED: [[PRED_STORE_IF27]]:
+; FIXED-NEXT: br label %[[PRED_STORE_CONTINUE28:.*]]
+; FIXED: [[PRED_STORE_CONTINUE28]]:
+; FIXED-NEXT: ret void
+;
+; SCALABLE-LABEL: define void @global_table_load(
+; SCALABLE-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], i64 [[N:%.*]]) {
+; SCALABLE-NEXT: [[SCALAR_PH:.*:]]
+; SCALABLE-NEXT: br label %[[VECTOR_PH:.*]]
+; SCALABLE: [[VECTOR_PH]]:
+; SCALABLE-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; SCALABLE-NEXT: [[TMP2:%.*]] = shl nuw i64 [[TMP0]], 1
+; SCALABLE-NEXT: [[TMP10:%.*]] = sub i64 [[TMP2]], 1
+; SCALABLE-NEXT: [[N_RND_UP:%.*]] = add i64 [[N]], [[TMP10]]
+; SCALABLE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP2]]
+; SCALABLE-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; SCALABLE-NEXT: [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], 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: [[TMP3:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP2]], 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: br label %[[FOR_BODY:.*]]
+; SCALABLE: [[FOR_BODY]]:
+; SCALABLE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[FOR_BODY]] ]
+; SCALABLE-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP3]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[FOR_BODY]] ]
+; SCALABLE-NEXT: [[TMP11:%.*]] = icmp ule <vscale x 2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
+; SCALABLE-NEXT: [[ARRAYIDX_COND:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[IV]]
+; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = call <vscale x 2 x i32> @llvm.masked.load.nxv2i32.p0(ptr align 4 [[ARRAYIDX_COND]], <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP4:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
+; SCALABLE-NEXT: [[TMP5:%.*]] = xor <vscale x 2 x i1> [[TMP4]], splat (i1 true)
+; SCALABLE-NEXT: [[TMP8:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i1> zeroinitializer
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 shufflevector (<vscale x 2 x ptr> insertelement (<vscale x 2 x ptr> poison, ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), i64 0), <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x i1> [[TMP8]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP9:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i1> zeroinitializer
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 shufflevector (<vscale x 2 x ptr> insertelement (<vscale x 2 x ptr> poison, ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), i64 0), <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x i1> [[TMP9]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER1]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]]
+; SCALABLE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[IV]]
+; SCALABLE-NEXT: call void @llvm.masked.store.nxv2i32.p0(<vscale x 2 x i32> [[PREDPHI]], ptr align 4 [[TMP6]], <vscale x 2 x i1> [[TMP11]])
+; SCALABLE-NEXT: [[INDEX_NEXT]] = add i64 [[IV]], [[TMP2]]
+; SCALABLE-NEXT: [[VEC_IND_NEXT]] = add nuw <vscale x 2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT2]]
+; SCALABLE-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SCALABLE-NEXT: br i1 [[TMP7]], label %[[IF_END:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SCALABLE: [[IF_END]]:
+; SCALABLE-NEXT: br label %[[FOR_INC:.*]]
+; SCALABLE: [[FOR_INC]]:
+; SCALABLE-NEXT: ret void
+;
+entry:
+ br label %for.body
+
+for.body:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.inc ]
+ %arrayidx.cond = getelementptr inbounds i32, ptr %cond, i64 %iv
+ %c = load i32, ptr %arrayidx.cond, align 4
+ %cmp = icmp ne i32 %c, 0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %gep.log = getelementptr inbounds [4 x i32], ptr @log_table, i64 0, i64 2
+ %v.then = load i32, ptr %gep.log, align 4
+ br label %if.end
+
+if.else:
+ %gep.exp = getelementptr inbounds [4 x i32], ptr @exp_table, i64 0, i64 1
+ %v.else = load i32, ptr %gep.exp, align 4
+ br label %if.end
+
+if.end:
+ %v = phi i32 [ %v.then, %if.then ], [ %v.else, %if.else ]
+ %arrayidx.out = getelementptr inbounds i32, ptr %out, i64 %iv
+ store i32 %v, ptr %arrayidx.out, align 4
+ br label %for.inc
+
+for.inc:
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exitcond = icmp eq i64 %iv.next, %n
+ br i1 %exitcond, label %exit, label %for.body
+
+exit:
+ ret void
+}
+
+; Uniform, dereferenceable(16) pointer parameter.
+; The access (bytes [12,16)) is inside the range.
+define void @deref_param_in_range(ptr noalias %out, ptr noalias %cond, ptr noalias nonnull align 4 dereferenceable(16) %table) nofree {
+; FIXED-LABEL: define void @deref_param_in_range(
+; FIXED-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(16) [[TABLE:%.*]]) #[[ATTR0:[0-9]+]] {
+; FIXED-NEXT: [[ENTRY:.*:]]
+; FIXED-NEXT: br label %[[VECTOR_PH:.*]]
+; FIXED: [[VECTOR_PH]]:
+; FIXED-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x ptr> poison, ptr [[TMP0]], i64 0
+; FIXED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x ptr> [[BROADCAST_SPLATINSERT]], <2 x ptr> poison, <2 x i32> zeroinitializer
+; FIXED-NEXT: br label %[[VECTOR_BODY:.*]]
+; FIXED: [[VECTOR_BODY]]:
+; FIXED-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; FIXED-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
+; FIXED-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP1]], align 4
+; FIXED-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
+; FIXED-NEXT: [[TMP6:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 [[BROADCAST_SPLAT]], <2 x i1> [[TMP2]], <2 x i32> poison)
+; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP2]], <2 x i32> [[TMP6]], <2 x i32> zeroinitializer
+; FIXED-NEXT: [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
+; FIXED-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP11]], align 4
+; FIXED-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; FIXED-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 16
+; FIXED-NEXT: br i1 [[TMP12]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; FIXED: [[MIDDLE_BLOCK]]:
+; FIXED-NEXT: br label %[[EXIT:.*]]
+; FIXED: [[EXIT]]:
+; FIXED-NEXT: ret void
+;
+; SCALABLE-LABEL: define void @deref_param_in_range(
+; SCALABLE-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(16) [[TABLE:%.*]]) #[[ATTR0:[0-9]+]] {
+; SCALABLE-NEXT: [[VECTOR_PH:.*:]]
+; SCALABLE-NEXT: br label %[[VECTOR_PH1:.*]]
+; SCALABLE: [[VECTOR_PH1]]:
+; SCALABLE-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
+; SCALABLE-NEXT: [[TMP2:%.*]] = shl nuw i64 [[TMP4]], 1
+; SCALABLE-NEXT: [[TMP9:%.*]] = sub i64 [[TMP2]], 1
+; SCALABLE-NEXT: [[N_RND_UP:%.*]] = add i64 16, [[TMP9]]
+; SCALABLE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP2]]
+; SCALABLE-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; SCALABLE-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[TMP3]], i64 0
+; SCALABLE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[TMP10:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP2]], i64 0
+; SCALABLE-NEXT: [[BROADCAST_SPLAT1:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: br label %[[VECTOR_BODY:.*]]
+; SCALABLE: [[VECTOR_BODY]]:
+; SCALABLE-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH1]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SCALABLE-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP10]], %[[VECTOR_PH1]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SCALABLE-NEXT: [[TMP11:%.*]] = icmp ule <vscale x 2 x i64> [[VEC_IND]], splat (i64 15)
+; SCALABLE-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
+; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = call <vscale x 2 x i32> @llvm.masked.load.nxv2i32.p0(ptr align 4 [[TMP1]], <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP5:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
+; SCALABLE-NEXT: [[TMP8:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i1> zeroinitializer
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 [[BROADCAST_SPLAT]], <vscale x 2 x i1> [[TMP8]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]], <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
+; SCALABLE-NEXT: call void @llvm.masked.store.nxv2i32.p0(<vscale x 2 x i32> [[PREDPHI]], ptr align 4 [[TMP6]], <vscale x 2 x i1> [[TMP11]])
+; SCALABLE-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP2]]
+; SCALABLE-NEXT: [[VEC_IND_NEXT]] = add nuw <vscale x 2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT1]]
+; SCALABLE-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SCALABLE-NEXT: br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; SCALABLE: [[MIDDLE_BLOCK]]:
+; SCALABLE-NEXT: br label %[[EXIT:.*]]
+; SCALABLE: [[EXIT]]:
+; SCALABLE-NEXT: ret void
+;
+entry:
+ br label %for.body
+
+for.body:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.inc ]
+ %arrayidx.cond = getelementptr inbounds i32, ptr %cond, i64 %iv
+ %c = load i32, ptr %arrayidx.cond, align 4
+ %cmp = icmp ne i32 %c, 0
+ br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+ %gep.table = getelementptr inbounds i32, ptr %table, i64 3
+ %v.then = load i32, ptr %gep.table, align 4
+ br label %if.end
+
+if.end:
+ %v = phi i32 [ %v.then, %if.then ], [ 0, %for.body ]
+ %arrayidx.out = getelementptr inbounds i32, ptr %out, i64 %iv
+ store i32 %v, ptr %arrayidx.out, align 4
+ br label %for.inc
+
+for.inc:
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exitcond = icmp eq i64 %iv.next, 16
+ br i1 %exitcond, label %exit, label %for.body
+
+exit:
+ ret void
+}
+
+; Uniform, dereferenceable(15) pointer parameter.
+; The last byte of the accessed i32 falls outside the range.
+define void @deref_param_out_of_range_last_byte(ptr noalias %out, ptr noalias %cond, ptr noalias nonnull align 4 dereferenceable(15) %table) nofree {
+; FIXED-LABEL: define void @deref_param_out_of_range_last_byte(
+; FIXED-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(15) [[TABLE:%.*]]) #[[ATTR0]] {
+; FIXED-NEXT: [[ENTRY:.*:]]
+; FIXED-NEXT: br label %[[VECTOR_PH:.*]]
+; FIXED: [[VECTOR_PH]]:
+; FIXED-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x ptr> poison, ptr [[TMP0]], i64 0
+; FIXED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x ptr> [[BROADCAST_SPLATINSERT]], <2 x ptr> poison, <2 x i32> zeroinitializer
+; FIXED-NEXT: br label %[[VECTOR_BODY:.*]]
+; FIXED: [[VECTOR_BODY]]:
+; FIXED-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; FIXED-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
+; FIXED-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP1]], align 4
+; FIXED-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
+; FIXED-NEXT: [[TMP10:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 [[BROADCAST_SPLAT]], <2 x i1> [[TMP2]], <2 x i32> poison)
+; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP2]], <2 x i32> [[TMP10]], <2 x i32> zeroinitializer
+; FIXED-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
+; FIXED-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP19]], align 4
+; FIXED-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; FIXED-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 16
+; FIXED-NEXT: br i1 [[TMP20]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; FIXED: [[MIDDLE_BLOCK]]:
+; FIXED-NEXT: br label %[[EXIT:.*]]
+; FIXED: [[EXIT]]:
+; FIXED-NEXT: ret void
+;
+; SCALABLE-LABEL: define void @deref_param_out_of_range_last_byte(
+; SCALABLE-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(15) [[TABLE:%.*]]) #[[ATTR0]] {
+; SCALABLE-NEXT: [[ENTRY:.*:]]
+; SCALABLE-NEXT: br label %[[VECTOR_PH:.*]]
+; SCALABLE: [[VECTOR_PH]]:
+; SCALABLE-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; SCALABLE-NEXT: [[TMP2:%.*]] = shl nuw i64 [[TMP0]], 1
+; SCALABLE-NEXT: [[TMP9:%.*]] = sub i64 [[TMP2]], 1
+; SCALABLE-NEXT: [[N_RND_UP:%.*]] = add i64 16, [[TMP9]]
+; SCALABLE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP2]]
+; SCALABLE-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; SCALABLE-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[TMP3]], i64 0
+; SCALABLE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[TMP4:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP2]], i64 0
+; SCALABLE-NEXT: [[BROADCAST_SPLAT1:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: br label %[[FOR_BODY:.*]]
+; SCALABLE: [[FOR_BODY]]:
+; SCALABLE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[FOR_BODY]] ]
+; SCALABLE-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP4]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[FOR_BODY]] ]
+; SCALABLE-NEXT: [[TMP10:%.*]] = icmp ule <vscale x 2 x i64> [[VEC_IND]], splat (i64 15)
+; SCALABLE-NEXT: [[ARRAYIDX_COND:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[IV]]
+; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = call <vscale x 2 x i32> @llvm.masked.load.nxv2i32.p0(ptr align 4 [[ARRAYIDX_COND]], <vscale x 2 x i1> [[TMP10]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP5:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
+; SCALABLE-NEXT: [[TMP8:%.*]] = select <vscale x 2 x i1> [[TMP10]], <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i1> zeroinitializer
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 [[BROADCAST_SPLAT]], <vscale x 2 x i1> [[TMP8]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]], <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[IV]]
+; SCALABLE-NEXT: call void @llvm.masked.store.nxv2i32.p0(<vscale x 2 x i32> [[PREDPHI]], ptr align 4 [[TMP6]], <vscale x 2 x i1> [[TMP10]])
+; SCALABLE-NEXT: [[INDEX_NEXT]] = add i64 [[IV]], [[TMP2]]
+; SCALABLE-NEXT: [[VEC_IND_NEXT]] = add nuw <vscale x 2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT1]]
+; SCALABLE-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SCALABLE-NEXT: br i1 [[TMP7]], label %[[IF_END:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; SCALABLE: [[IF_END]]:
+; SCALABLE-NEXT: br label %[[FOR_INC:.*]]
+; SCALABLE: [[FOR_INC]]:
+; SCALABLE-NEXT: ret void
+;
+entry:
+ br label %for.body
+
+for.body:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.inc ]
+ %arrayidx.cond = getelementptr inbounds i32, ptr %cond, i64 %iv
+ %c = load i32, ptr %arrayidx.cond, align 4
+ %cmp = icmp ne i32 %c, 0
+ br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+ %gep.table = getelementptr inbounds i32, ptr %table, i64 3
+ %v.then = load i32, ptr %gep.table, align 4
+ br label %if.end
+
+if.end:
+ %v = phi i32 [ %v.then, %if.then ], [ 0, %for.body ]
+ %arrayidx.out = getelementptr inbounds i32, ptr %out, i64 %iv
+ store i32 %v, ptr %arrayidx.out, align 4
+ br label %for.inc
+
+for.inc:
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exitcond = icmp eq i64 %iv.next, 16
+ br i1 %exitcond, label %exit, label %for.body
+
+exit:
+ ret void
+}
+
+; Uniform, dereferenceable(16) pointer parameter.
+; Since %iv < 1024, The offset (bytes [0,4)) is inside the range.
+define void @deref_param_in_range_uniform_but_not_loop_invariant(ptr noalias %out, ptr noalias %cond, ptr noalias nonnull align 4 dereferenceable(16) %table) nofree vscale_range(4, 4) {
+; FIXED-LABEL: define void @deref_param_in_range_uniform_but_not_loop_invariant(
+; FIXED-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(16) [[TABLE:%.*]]) #[[ATTR1:[0-9]+]] {
+; FIXED-NEXT: [[ENTRY:.*:]]
+; FIXED-NEXT: br label %[[VECTOR_PH:.*]]
+; FIXED: [[VECTOR_PH]]:
+; FIXED-NEXT: br label %[[VECTOR_BODY:.*]]
+; FIXED: [[VECTOR_BODY]]:
+; FIXED-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; FIXED-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
+; FIXED-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP0]], align 4
+; FIXED-NEXT: [[TMP1:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
+; FIXED-NEXT: [[TMP2:%.*]] = lshr i64 [[INDEX]], 10
+; FIXED-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 [[TMP2]]
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x ptr> poison, ptr [[TMP3]], i64 0
+; FIXED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x ptr> [[BROADCAST_SPLATINSERT]], <2 x ptr> poison, <2 x i32> zeroinitializer
+; FIXED-NEXT: [[TMP11:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 [[BROADCAST_SPLAT]], <2 x i1> [[TMP1]], <2 x i32> poison)
+; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP1]], <2 x i32> [[TMP11]], <2 x i32> zeroinitializer
+; FIXED-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
+; FIXED-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP20]], align 4
+; FIXED-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; FIXED-NEXT: [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], 32
+; FIXED-NEXT: br i1 [[TMP21]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
+; FIXED: [[MIDDLE_BLOCK]]:
+; FIXED-NEXT: br label %[[EXIT:.*]]
+; FIXED: [[EXIT]]:
+; FIXED-NEXT: ret void
+;
+; SCALABLE-LABEL: define void @deref_param_in_range_uniform_but_not_loop_invariant(
+; SCALABLE-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(16) [[TABLE:%.*]]) #[[ATTR1:[0-9]+]] {
+; SCALABLE-NEXT: [[ENTRY:.*:]]
+; SCALABLE-NEXT: br label %[[FOR_BODY:.*]]
+; SCALABLE: [[FOR_BODY]]:
+; SCALABLE-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; SCALABLE-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 1
+; SCALABLE-NEXT: [[TMP2:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP1]], 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: br label %[[IF_END:.*]]
+; SCALABLE: [[IF_END]]:
+; SCALABLE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[FOR_BODY]] ], [ [[IV_NEXT:%.*]], %[[IF_END]] ]
+; SCALABLE-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP2]], %[[FOR_BODY]] ], [ [[VEC_IND_NEXT:%.*]], %[[IF_END]] ]
+; SCALABLE-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[IV]]
+; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i32>, ptr [[TMP3]], align 4
+; SCALABLE-NEXT: [[TMP4:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
+; SCALABLE-NEXT: [[TMP5:%.*]] = lshr <vscale x 2 x i64> [[VEC_IND]], splat (i64 10)
+; SCALABLE-NEXT: [[WIDE_GEP:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], <vscale x 2 x i64> [[TMP5]]
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 [[WIDE_GEP]], <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]], <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[ARRAYIDX_OUT:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[IV]]
+; SCALABLE-NEXT: store <vscale x 2 x i32> [[PREDPHI]], ptr [[ARRAYIDX_OUT]], align 4
+; SCALABLE-NEXT: [[IV_NEXT]] = add nuw i64 [[IV]], [[TMP1]]
+; SCALABLE-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 2 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
+; SCALABLE-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 32
+; SCALABLE-NEXT: br i1 [[EXITCOND]], label %[[EXIT:.*]], label %[[IF_END]], !llvm.loop [[LOOP5:![0-9]+]]
+; SCALABLE: [[EXIT]]:
+; SCALABLE-NEXT: br label %[[EXIT1:.*]]
+; SCALABLE: [[EXIT1]]:
+; SCALABLE-NEXT: ret void
+;
+entry:
+ br label %for.body
+
+for.body:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.inc ]
+ %arrayidx.cond = getelementptr inbounds i32, ptr %cond, i64 %iv
+ %c = load i32, ptr %arrayidx.cond, align 4
+ %cmp = icmp ne i32 %c, 0
+ br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+ %offset = udiv i64 %iv, 1024
+ %gep.table = getelementptr inbounds i32, ptr %table, i64 %offset
+ %v.then = load i32, ptr %gep.table, align 4
+ br label %if.end
+
+if.end:
+ %v = phi i32 [ %v.then, %if.then ], [ 0, %for.body ]
+ %arrayidx.out = getelementptr inbounds i32, ptr %out, i64 %iv
+ store i32 %v, ptr %arrayidx.out, align 4
+ br label %for.inc
+
+for.inc:
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exitcond = icmp eq i64 %iv.next, 32
+ br i1 %exitcond, label %exit, label %for.body
+
+exit:
+ ret void
+}
+
>From 45b346d3e7ea86443194eabe35eec2c9cd6dab72 Mon Sep 17 00:00:00 2001
From: ShihPo Hung <shihpo.hung at sifive.com>
Date: Mon, 17 Aug 2026 09:27:32 -0700
Subject: [PATCH 2/2] [VPlan] Lower safe uniform load to unconditional scalar
load
When a load's address is uniform across VFs/UFs and provably safe to
access unconditionally, lower it to an unconditional scalar load recipe.
This avoids generating unnecessary masked loads or gathers for uniform memory
accesses.
---
.../Transforms/Vectorize/LoopVectorize.cpp | 1 +
.../Transforms/Vectorize/VPlanTransforms.cpp | 34 +++++++++++++++++-
.../Transforms/Vectorize/VPlanTransforms.h | 3 ++
.../VPlan/vplan-print-before-after-all.ll | 1 +
.../LoopVectorize/uniform-gather.ll | 36 ++++++++++---------
5 files changed, 57 insertions(+), 18 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index c12a4f562f600..dc2388aa38668 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -6871,6 +6871,7 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlan(VPlanPtr Plan,
// legal and profitable.
RUN_VPLAN_PASS(VPlanTransforms::convertToStridedAccesses, *Plan, PSE,
*OrigLoop, CostCtx, Range);
+ RUN_VPLAN_PASS(VPlanTransforms::lowerSafeUniformLoads, *Plan, *TLI);
// Ensure scalar VF plans only contain VF=1, as required by hasScalarVFOnly.
if (Range.Start.isScalar())
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index e75e9cf4e9103..24d061ef2d287 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -5716,7 +5716,6 @@ void VPlanTransforms::convertToStridedAccesses(VPlan &Plan,
for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
auto *MemR = dyn_cast<VPWidenMemoryRecipe>(&R);
// TODO: Transform reverse access into strided access with -1 stride.
- // TODO: Transform gather/scatter with uniform address into strided access
// with 0 stride.
// TODO: Transform interleave access into multiple strided accesses.
if (!MemR || MemR->isConsecutive())
@@ -5823,3 +5822,36 @@ void VPlanTransforms::convertToStridedAccesses(VPlan &Plan,
}
}
}
+
+void VPlanTransforms::lowerSafeUniformLoads(VPlan &Plan,
+ const TargetLibraryInfo &TLI) {
+ for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+ vp_depth_first_deep(Plan.getEntry()))) {
+ for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+
+ auto *UniformGather = dyn_cast<VPWidenLoadRecipe>(&R);
+ if (!UniformGather ||
+ !vputils::isUniformAcrossVFsAndUFs(UniformGather->getAddr()))
+ continue;
+
+ auto &LI = cast<LoadInst>(UniformGather->getIngredient());
+ VPValue *Addr = UniformGather->getAddr();
+ Value *Underlying = Addr->getUnderlyingValue();
+ if (!Underlying ||
+ !isSafeToLoadUnconditionally(Underlying, LI.getType(), LI.getAlign(),
+ Plan.getDataLayout(), &LI,
+ /*AC=*/nullptr, /*DT=*/nullptr, &TLI))
+ continue;
+ DebugLoc DbgLoc = UniformGather->getDebugLoc();
+ auto *ScalarLoad =
+ new VPReplicateRecipe(&LI, {Addr}, /*IsSingleScalar=*/true,
+ /*Mask=*/nullptr, {}, *UniformGather, DbgLoc);
+ ScalarLoad->insertBefore(UniformGather);
+ auto *Broadcast =
+ new VPInstruction(VPInstruction::Broadcast, {ScalarLoad});
+ Broadcast->insertBefore(UniformGather);
+ UniformGather->replaceAllUsesWith(Broadcast);
+ UniformGather->eraseFromParent();
+ }
+ }
+}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index e1832cd2de27a..15f9e0988f69e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -352,6 +352,9 @@ struct VPlanTransforms {
static void convertToStridedAccesses(VPlan &Plan,
PredicatedScalarEvolution &PSE, Loop &L,
VPCostContext &Ctx, VFRange &Range);
+ // Lower safe uniform (potentially masked) loads to unconditional scalar
+ // loads.
+ static void lowerSafeUniformLoads(VPlan &Plan, const TargetLibraryInfo &TLI);
/// Remove dead recipes from \p Plan.
static void removeDeadRecipes(VPlan &Plan);
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-before-after-all.ll b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-before-after-all.ll
index a7271524f0191..ea6162d77f250 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-before-after-all.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-before-after-all.ll
@@ -37,6 +37,7 @@
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::convertToAbstractRecipes
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::createInterleaveGroups
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::convertToStridedAccesses
+; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::lowerSafeUniformLoads
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::dropPoisonGeneratingRecipes
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::hoistPredicatedLoads
; CHECK: VPlan for loop in 'foo' [[BEFORE_OR_AFTER]] VPlanTransforms::sinkPredicatedStores
diff --git a/llvm/test/Transforms/LoopVectorize/uniform-gather.ll b/llvm/test/Transforms/LoopVectorize/uniform-gather.ll
index 61af4c533d69a..8ee6b2ea33c0f 100644
--- a/llvm/test/Transforms/LoopVectorize/uniform-gather.ll
+++ b/llvm/test/Transforms/LoopVectorize/uniform-gather.ll
@@ -26,11 +26,12 @@ define void @global_table_load(ptr noalias %out, ptr noalias %cond, i64 %n) {
; FIXED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX1]]
; FIXED-NEXT: [[WIDE_LOAD:%.*]] = call <2 x i32> @llvm.masked.load.v2i32.p0(ptr align 4 [[TMP40]], <2 x i1> [[TMP1]], <2 x i32> poison)
; FIXED-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
-; FIXED-NEXT: [[TMP3:%.*]] = xor <2 x i1> [[TMP2]], splat (i1 true)
-; FIXED-NEXT: [[TMP7:%.*]] = select <2 x i1> [[TMP1]], <2 x i1> [[TMP3]], <2 x i1> zeroinitializer
-; FIXED-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 <ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4)>, <2 x i1> [[TMP7]], <2 x i32> poison)
-; FIXED-NEXT: [[TMP6:%.*]] = select <2 x i1> [[TMP1]], <2 x i1> [[TMP2]], <2 x i1> zeroinitializer
-; FIXED-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 <ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8)>, <2 x i1> [[TMP6]], <2 x i32> poison)
+; FIXED-NEXT: [[TMP6:%.*]] = load i32, ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), align 4
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <2 x i32> poison, i32 [[TMP6]], i64 0
+; FIXED-NEXT: [[WIDE_MASKED_GATHER:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT3]], <2 x i32> poison, <2 x i32> zeroinitializer
+; FIXED-NEXT: [[TMP7:%.*]] = load i32, ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), align 4
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <2 x i32> poison, i32 [[TMP7]], i64 0
+; FIXED-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT1]], <2 x i32> poison, <2 x i32> zeroinitializer
; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP2]], <2 x i32> [[WIDE_MASKED_GATHER1]], <2 x i32> [[WIDE_MASKED_GATHER]]
; FIXED-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX1]]
; FIXED-NEXT: call void @llvm.masked.store.v2i32.p0(<2 x i32> [[PREDPHI]], ptr align 4 [[TMP4]], <2 x i1> [[TMP1]])
@@ -68,11 +69,12 @@ define void @global_table_load(ptr noalias %out, ptr noalias %cond, i64 %n) {
; SCALABLE-NEXT: [[ARRAYIDX_COND:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[IV]]
; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = call <vscale x 2 x i32> @llvm.masked.load.nxv2i32.p0(ptr align 4 [[ARRAYIDX_COND]], <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i32> poison)
; SCALABLE-NEXT: [[TMP4:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
-; SCALABLE-NEXT: [[TMP5:%.*]] = xor <vscale x 2 x i1> [[TMP4]], splat (i1 true)
-; SCALABLE-NEXT: [[TMP8:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i1> zeroinitializer
-; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 shufflevector (<vscale x 2 x ptr> insertelement (<vscale x 2 x ptr> poison, ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), i64 0), <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x i1> [[TMP8]], <vscale x 2 x i32> poison)
-; SCALABLE-NEXT: [[TMP9:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i1> zeroinitializer
-; SCALABLE-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 shufflevector (<vscale x 2 x ptr> insertelement (<vscale x 2 x ptr> poison, ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), i64 0), <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x i1> [[TMP9]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP9:%.*]] = load i32, ptr getelementptr inbounds nuw (i8, ptr @exp_table, i64 4), align 4
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <vscale x 2 x i32> poison, i32 [[TMP9]], i64 0
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = shufflevector <vscale x 2 x i32> [[BROADCAST_SPLATINSERT5]], <vscale x 2 x i32> poison, <vscale x 2 x i32> zeroinitializer
+; SCALABLE-NEXT: [[TMP8:%.*]] = load i32, ptr getelementptr inbounds nuw (i8, ptr @log_table, i64 8), align 4
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 2 x i32> poison, i32 [[TMP8]], i64 0
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER1:%.*]] = shufflevector <vscale x 2 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 2 x i32> poison, <vscale x 2 x i32> zeroinitializer
; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP4]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER1]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]]
; SCALABLE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[IV]]
; SCALABLE-NEXT: call void @llvm.masked.store.nxv2i32.p0(<vscale x 2 x i32> [[PREDPHI]], ptr align 4 [[TMP6]], <vscale x 2 x i1> [[TMP11]])
@@ -129,15 +131,15 @@ define void @deref_param_in_range(ptr noalias %out, ptr noalias %cond, ptr noali
; FIXED-NEXT: br label %[[VECTOR_PH:.*]]
; FIXED: [[VECTOR_PH]]:
; FIXED-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
-; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x ptr> poison, ptr [[TMP0]], i64 0
-; FIXED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x ptr> [[BROADCAST_SPLATINSERT]], <2 x ptr> poison, <2 x i32> zeroinitializer
; FIXED-NEXT: br label %[[VECTOR_BODY:.*]]
; FIXED: [[VECTOR_BODY]]:
; FIXED-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; FIXED-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
; FIXED-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP1]], align 4
; FIXED-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[WIDE_LOAD]], zeroinitializer
-; FIXED-NEXT: [[TMP6:%.*]] = call <2 x i32> @llvm.masked.gather.v2i32.v2p0(<2 x ptr> align 4 [[BROADCAST_SPLAT]], <2 x i1> [[TMP2]], <2 x i32> poison)
+; FIXED-NEXT: [[TMP3:%.*]] = load i32, ptr [[TMP0]], align 4
+; FIXED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[TMP3]], i64 0
+; FIXED-NEXT: [[TMP6:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer
; FIXED-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP2]], <2 x i32> [[TMP6]], <2 x i32> zeroinitializer
; FIXED-NEXT: [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
; FIXED-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP11]], align 4
@@ -161,8 +163,6 @@ define void @deref_param_in_range(ptr noalias %out, ptr noalias %cond, ptr noali
; SCALABLE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP2]]
; SCALABLE-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
; SCALABLE-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TABLE]], i64 3
-; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x ptr> poison, ptr [[TMP3]], i64 0
-; SCALABLE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 2 x ptr> [[BROADCAST_SPLATINSERT]], <vscale x 2 x ptr> poison, <vscale x 2 x i32> zeroinitializer
; SCALABLE-NEXT: [[TMP10:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()
; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP2]], i64 0
; SCALABLE-NEXT: [[BROADCAST_SPLAT1:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer
@@ -174,8 +174,9 @@ define void @deref_param_in_range(ptr noalias %out, ptr noalias %cond, ptr noali
; SCALABLE-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[COND]], i64 [[INDEX]]
; SCALABLE-NEXT: [[WIDE_LOAD:%.*]] = call <vscale x 2 x i32> @llvm.masked.load.nxv2i32.p0(ptr align 4 [[TMP1]], <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i32> poison)
; SCALABLE-NEXT: [[TMP5:%.*]] = icmp ne <vscale x 2 x i32> [[WIDE_LOAD]], zeroinitializer
-; SCALABLE-NEXT: [[TMP8:%.*]] = select <vscale x 2 x i1> [[TMP11]], <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i1> zeroinitializer
-; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 2 x i32> @llvm.masked.gather.nxv2i32.nxv2p0(<vscale x 2 x ptr> align 4 [[BROADCAST_SPLAT]], <vscale x 2 x i1> [[TMP8]], <vscale x 2 x i32> poison)
+; SCALABLE-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP3]], align 4
+; SCALABLE-NEXT: [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <vscale x 2 x i32> poison, i32 [[TMP8]], i64 0
+; SCALABLE-NEXT: [[WIDE_MASKED_GATHER:%.*]] = shufflevector <vscale x 2 x i32> [[BROADCAST_SPLATINSERT2]], <vscale x 2 x i32> poison, <vscale x 2 x i32> zeroinitializer
; SCALABLE-NEXT: [[PREDPHI:%.*]] = select <vscale x 2 x i1> [[TMP5]], <vscale x 2 x i32> [[WIDE_MASKED_GATHER]], <vscale x 2 x i32> zeroinitializer
; SCALABLE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[OUT]], i64 [[INDEX]]
; SCALABLE-NEXT: call void @llvm.masked.store.nxv2i32.p0(<vscale x 2 x i32> [[PREDPHI]], ptr align 4 [[TMP6]], <vscale x 2 x i1> [[TMP11]])
@@ -318,6 +319,7 @@ exit:
; Uniform, dereferenceable(16) pointer parameter.
; Since %iv < 1024, The offset (bytes [0,4)) is inside the range.
+; TODO: Fold GEP offset to 0 using IV upper bound (< 1024).
define void @deref_param_in_range_uniform_but_not_loop_invariant(ptr noalias %out, ptr noalias %cond, ptr noalias nonnull align 4 dereferenceable(16) %table) nofree vscale_range(4, 4) {
; FIXED-LABEL: define void @deref_param_in_range_uniform_but_not_loop_invariant(
; FIXED-SAME: ptr noalias [[OUT:%.*]], ptr noalias [[COND:%.*]], ptr noalias nonnull align 4 dereferenceable(16) [[TABLE:%.*]]) #[[ATTR1:[0-9]+]] {
More information about the llvm-commits
mailing list