[llvm] [VPlan] Lower safe uniform load to unconditional scalar load (PR #216767)
Shih-Po Hung via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 27 00:07:56 PDT 2026
================
@@ -0,0 +1,413 @@
+; 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: [[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]])
+; 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: [[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]])
+; 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 {
----------------
arcbbb wrote:
@global_table_load was actually extracted from a real benchmark. @deref_param_in_range is just a simplified version of it to ensure test coverage, so I don't have any other test cases besides these.
https://github.com/llvm/llvm-project/pull/216767
More information about the llvm-commits
mailing list