[llvm] [LV] Use getSmallBestKnownTC in IV-overflow-check (PR #195226)
Luke Lau via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 24 00:35:06 PDT 2026
================
@@ -0,0 +1,101 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -passes=loop-vectorize -mtriple=riscv64 -S %s | FileCheck %s
+
+target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n64-S128"
+
+; VF is scalable, as is TC. We can prove that TC > VF.
+define void @iv_no_overflow(ptr noalias %a, ptr noalias %b) #0 {
+; CHECK-LABEL: define void @iv_no_overflow(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[VS:%.*]] = tail call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[N:%.*]] = mul nuw nsw i64 [[VS]], 100
+; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[CURRENT_ITERATION_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[AVL:%.*]] = phi i64 [ [[N]], %[[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 4, i1 true)
+; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX]]
+; CHECK-NEXT: [[VP_OP_LOAD:%.*]] = call <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr align 4 [[TMP1]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP0]])
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[B]], i64 [[INDEX]]
+; CHECK-NEXT: call void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float> [[VP_OP_LOAD]], ptr align 4 [[TMP2]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP0]])
+; CHECK-NEXT: [[TMP3:%.*]] = zext i32 [[TMP0]] to i64
+; CHECK-NEXT: [[CURRENT_ITERATION_NEXT]] = add nuw i64 [[TMP3]], [[INDEX]]
+; CHECK-NEXT: [[AVL_NEXT]] = sub nuw i64 [[AVL]], [[TMP3]]
+; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
+; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: br label %[[EXIT:.*]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %vs = tail call i64 @llvm.vscale.i64()
+ %n = mul nuw nsw i64 %vs, 100
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %gep.a = getelementptr inbounds nuw float, ptr %a, i64 %iv
+ %la = load float, ptr %gep.a, align 4
+ %gep.b = getelementptr inbounds nuw float, ptr %b, i64 %iv
+ store float %la, ptr %gep.b, align 4
+ %iv.next = add nuw nsw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, %n
+ br i1 %ec, label %exit, label %loop
+
+exit:
+ ret void
+}
+
+; VF is scalable, as is TC. TC * MaxVScale overflows,
+; make sure that the IV increment does not have nuw.
+define void @tc_maxvscale_overflow(ptr noalias %a, ptr noalias %b) #0 {
+; CHECK-LABEL: define void @tc_maxvscale_overflow(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[VS:%.*]] = tail call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[N:%.*]] = mul nuw nsw i64 [[VS]], 268435457
+; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[CURRENT_ITERATION_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[AVL:%.*]] = phi i64 [ [[N]], %[[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 4, i1 true)
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX]]
+; CHECK-NEXT: [[VP_OP_LOAD:%.*]] = call <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr align 4 [[TMP2]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP0]])
+; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw float, ptr [[B]], i64 [[INDEX]]
+; CHECK-NEXT: call void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float> [[VP_OP_LOAD]], ptr align 4 [[TMP3]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP0]])
+; CHECK-NEXT: [[TMP5:%.*]] = zext i32 [[TMP0]] to i64
+; CHECK-NEXT: [[CURRENT_ITERATION_NEXT]] = add i64 [[TMP5]], [[INDEX]]
+; CHECK-NEXT: [[AVL_NEXT]] = sub nuw i64 [[AVL]], [[TMP5]]
+; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[AVL_NEXT]], 0
+; CHECK-NEXT: br i1 [[TMP4]], label %[[SCALAR_PH:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %vs = tail call i64 @llvm.vscale.i64()
+ %n = mul nuw nsw i64 %vs, 268435457
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %gep.a = getelementptr inbounds nuw float, ptr %a, i64 %iv
+ %la = load float, ptr %gep.a, align 4
+ %gep.b = getelementptr inbounds nuw float, ptr %b, i64 %iv
+ store float %la, ptr %gep.b, align 4
+ %iv.next = add nuw nsw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, %n
+ br i1 %ec, label %exit, label %loop
+
+exit:
+ ret void
+}
+
+attributes #0 = { "target-features"="+v,+zvl512b" }
----------------
lukel97 wrote:
Better to just stick this in the RUN line as `-mattr=+v,+zvl512b`. Any reason why we need the extra zvl512b btw? I thought that just changes the minimum vscale value
https://github.com/llvm/llvm-project/pull/195226
More information about the llvm-commits
mailing list