[llvm] [NVPTX] Disable relative lookup tables (PR #159748)

[Nikita Popov via llvm-commits]

https://github.com/nikic created https://github.com/llvm/llvm-project/pull/159748

Relative lookup tables result in "LLVM ERROR: Circular dependency found in global variable set", so disable them for this target.

>From 4651eb7df79ec80a46662398924c2fe31ce469d1 Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Fri, 19 Sep 2025 12:22:09 +0200
Subject: [PATCH 1/2] Add test

---
 .../RelLookupTableConverter/nvptx.ll          | 33 +++++++++++++++++++
 1 file changed, 33 insertions(+)
 create mode 100644 llvm/test/Transforms/RelLookupTableConverter/nvptx.ll

diff --git a/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll b/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll
new file mode 100644
index 0000000000000..ac4cde795f7e2
--- /dev/null
+++ b/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll
@@ -0,0 +1,33 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 5
+; RUN: opt < %s -passes=rel-lookup-table-converter -relocation-model=pic -S | FileCheck %s
+; REQUIRES: nvptx-registered-target
+target triple = "nvptx64-nvidia-cuda"
+
+ at a1 = internal constant i32 0, align 4
+ at b1 = internal constant i32 0, align 4
+ at c1 = internal constant i32 0, align 4
+ at d1 = internal constant i32 0, align 4
+
+ at switch.table = private unnamed_addr constant [3 x ptr] [ptr @a1, ptr @b1, ptr @c1], align 8
+
+;.
+; CHECK: @a1 = internal constant i32 0, align 4
+; CHECK: @b1 = internal constant i32 0, align 4
+; CHECK: @c1 = internal constant i32 0, align 4
+; CHECK: @d1 = internal constant i32 0, align 4
+; CHECK: @switch.table.rel = private unnamed_addr constant [3 x i32] [i32 trunc (i64 sub (i64 ptrtoint (ptr @a1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @b1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @c1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32)], align 4
+;.
+define ptr @internal_linkage(i32 %cond) {
+; CHECK-LABEL: define ptr @internal_linkage(
+; CHECK-SAME: i32 [[COND:%.*]]) {
+; CHECK-NEXT:    [[RELTABLE_SHIFT:%.*]] = shl i32 [[COND]], 2
+; CHECK-NEXT:    [[RELTABLE_INTRINSIC:%.*]] = call ptr @llvm.load.relative.i32(ptr @switch.table.rel, i32 [[RELTABLE_SHIFT]])
+; CHECK-NEXT:    ret ptr [[RELTABLE_INTRINSIC]]
+;
+  %switch.gep = getelementptr inbounds [3 x ptr], ptr @switch.table, i32 0, i32 %cond
+  %switch.load = load ptr, ptr %switch.gep, align 8
+  ret ptr %switch.load
+}
+;.
+; CHECK: attributes #[[ATTR0:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(argmem: read) }
+;.

>From 005378330fe1328ce9e7139cd1c9564b7a8bdf7f Mon Sep 17 00:00:00 2001
From: Nikita Popov <npopov at redhat.com>
Date: Fri, 19 Sep 2025 12:24:29 +0200
Subject: [PATCH 2/2] Disable relative lookup tables on nvptx

---
 llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.h      |  5 +++++
 llvm/test/Transforms/RelLookupTableConverter/nvptx.ll | 11 +++++------
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.h b/llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.h
index b32d931bd3074..78eb751cf3c2e 100644
--- a/llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.h
+++ b/llvm/lib/Target/NVPTX/NVPTXTargetTransformInfo.h
@@ -190,6 +190,11 @@ class NVPTXTTIImpl final : public BasicTTIImplBase<NVPTXTTIImpl> {
   void collectKernelLaunchBounds(
       const Function &F,
       SmallVectorImpl<std::pair<StringRef, int64_t>> &LB) const override;
+
+  bool shouldBuildRelLookupTables() const override {
+    // Self-referential globals are not supported.
+    return false;
+  }
 };
 
 } // end namespace llvm
diff --git a/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll b/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll
index ac4cde795f7e2..70ebf220c369c 100644
--- a/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll
+++ b/llvm/test/Transforms/RelLookupTableConverter/nvptx.ll
@@ -3,6 +3,8 @@
 ; REQUIRES: nvptx-registered-target
 target triple = "nvptx64-nvidia-cuda"
 
+; Do not produce relative lookup table for nvptx target.
+
 @a1 = internal constant i32 0, align 4
 @b1 = internal constant i32 0, align 4
 @c1 = internal constant i32 0, align 4
@@ -15,19 +17,16 @@ target triple = "nvptx64-nvidia-cuda"
 ; CHECK: @b1 = internal constant i32 0, align 4
 ; CHECK: @c1 = internal constant i32 0, align 4
 ; CHECK: @d1 = internal constant i32 0, align 4
-; CHECK: @switch.table.rel = private unnamed_addr constant [3 x i32] [i32 trunc (i64 sub (i64 ptrtoint (ptr @a1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @b1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (ptr @c1 to i64), i64 ptrtoint (ptr @switch.table.rel to i64)) to i32)], align 4
+; CHECK: @switch.table = private unnamed_addr constant [3 x ptr] [ptr @a1, ptr @b1, ptr @c1], align 8
 ;.
 define ptr @internal_linkage(i32 %cond) {
 ; CHECK-LABEL: define ptr @internal_linkage(
 ; CHECK-SAME: i32 [[COND:%.*]]) {
-; CHECK-NEXT:    [[RELTABLE_SHIFT:%.*]] = shl i32 [[COND]], 2
-; CHECK-NEXT:    [[RELTABLE_INTRINSIC:%.*]] = call ptr @llvm.load.relative.i32(ptr @switch.table.rel, i32 [[RELTABLE_SHIFT]])
+; CHECK-NEXT:    [[SWITCH_GEP:%.*]] = getelementptr inbounds [3 x ptr], ptr @switch.table, i32 0, i32 [[COND]]
+; CHECK-NEXT:    [[RELTABLE_INTRINSIC:%.*]] = load ptr, ptr [[SWITCH_GEP]], align 8
 ; CHECK-NEXT:    ret ptr [[RELTABLE_INTRINSIC]]
 ;
   %switch.gep = getelementptr inbounds [3 x ptr], ptr @switch.table, i32 0, i32 %cond
   %switch.load = load ptr, ptr %switch.gep, align 8
   ret ptr %switch.load
 }
-;.
-; CHECK: attributes #[[ATTR0:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(argmem: read) }
-;.