[SYCL][Driver] Warn when total Device code-size causes relocation errors by bviyer · Pull Request #23523 · intel/llvm

bviyer · 2026-10-16T20:04:17Z

Large device code can push the 32-bit PC relative relocations past their 2GB reach. this causes host linker to fail with errors such as "relocation truncated to fit." In this PR, such cases are dignosed at linked time via -Wlarge-device-kernel flag and providing user suggestion to use the flink-huge-device-code as a fix.

elizabethandrews

This is primarily a Driver Change with just a warning group declared in FE. I defer to @mdtoguchi for review. I agree that the current proposed warning is too verbose.

… bviyer-make-better-warning

mdtoguchi · 2026-10-26T21:11:39Z

+
+  for (const InputInfo &II : Inputs)
+    if (II.isFilename() &&
+        isa_and_nonnull<OffloadWrapperJobAction>(II.getAction())) {


This is good for the old offload model. We should also look into a solution for the new model - I would expect that to be more in the clang-linker-wrapper space.

Can we do that in a separate PR?

Yup, this can be done in a separate PR.

mdtoguchi · 2026-10-26T21:12:37Z

+static constexpr uint64_t HugeDeviceCodeWarnThreshold = 1ull << 30; // 1GiB
+
+// Sums the size of only those link inputs that carry wrapped device code,
+// i.e. skips host objects, static archives, and other linker inputs that


the comment on static archives isn't true - as static archives should be queried as well, as they can be 'fat archives'. But that query happens in a different area so the device objects from any static archives should already be processed along with the other device binaries.

mdtoguchi · 2026-10-27T00:00:46Z

+                   "clang LLVM compiler", InMemoryFileSystem);
+  TheDriver.setCheckInputsExist(false);
+  std::unique_ptr<Compilation> C(TheDriver.BuildCompilation(
+      {"/usr/bin/clang++", "-fsycl", "-nostdlib", "foo.cpp", "-o", "foo.out"}));


We probably don't want the link step to occur when the warning is triggered. Adding -Werror=large-device-kernel will prevent it from running the link.

mdtoguchi · 2026-10-27T00:05:12Z

+
+// RUN: rm -f %t-big.o
+
+// NO-WARN-NOT: total link input size


Missing tests that check that -flink-huge-device-code and -Wno-large-device-kernel prevent the warning from being emitted.

I may have commented here to add these checks - but given that the originating objects don't have any device code in the first place - the additional tests aren't going to be doing anything here and will just be spinning compilation time on large inputs.

mdtoguchi · 2026-10-27T00:09:01Z

+// RUN: %python -c "open(r'%t-big.o', 'wb').truncate(1100000000)"
+// RUN: %python -c "open(r'%t-small.o', 'wb').truncate(1000)"
+
+// RUN: %clangxx -fsycl %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1 || true


Suggested change

// RUN: %clangxx -fsycl %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1 || true

// RUN: %clangxx -fsycl %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1

No expectation for this command to fail is there? I don't think the || true is necessary.

mdtoguchi · 2026-10-27T22:14:04Z

+// RUN: %clangxx -fsycl -fuse-ld=/bin/true %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1
+// RUN: FileCheck --check-prefix=NO-WARN --allow-empty %s < %t-warn.log
+
+// RUN: %clangxx -fsycl -fuse-ld=/bin/true -flink-huge-device-code %t-big.o %t-small.o -o %t.out > %t-flink.log 2>&1
+// RUN: FileCheck --check-prefix=NO-WARN --allow-empty %s < %t-flink.log
+
+// RUN: %clangxx -fsycl -fuse-ld=/bin/true -Wno-large-device-kernel %t-big.o %t-small.o -o %t.out > %t-wno.log 2>&1
+// RUN: FileCheck --check-prefix=NO-WARN --allow-empty %s < %t-wno.log


I would expect these tests to fail in the new offload model setting. Probably should add --no-offload-new-driver to these command lines.

mdtoguchi · 2026-10-27T22:37:48Z

+  // short-circuit instead of actually invoking the (nonexistent, in this
+  // test) linker -- that's what we want, since the point here is just to


Comment says that the linker is non-existent. Probably not a true statement. Should just say we don't want to invoke the linker.

mdtoguchi · 2026-10-27T22:39:28Z

+  EXPECT_TRUE(DiagConsumer->Msgs.empty());
+  EXPECT_TRUE(DiagConsumer->Errors.empty());


These will fail if any diagnostic is emitted. Probably should tighten this up to check for more of the specific diagnostic messaging (not) expected.

mdtoguchi · 2026-10-27T22:40:07Z

+  EXPECT_TRUE(DiagConsumer->Msgs.empty());
+  EXPECT_TRUE(DiagConsumer->Errors.empty());


Same comment as below in regards to tightening up the diagnostic to not be emitted.

bviyer added 3 commits October 16, 2026 12:59

Init commit

c4c65c5

Added test case.

6cd8d03

Did clang format

a70693b

bviyer requested review from a team as code owners October 16, 2026 20:04

mdtoguchi reviewed Oct 19, 2026

View reviewed changes

elizabethandrews reviewed Oct 20, 2026

View reviewed changes

Fixed all issues mentioned by Mike

2cf1108

bviyer requested review from elizabethandrews and mdtoguchi October 20, 2026 19:22

elizabethandrews approved these changes Oct 21, 2026

View reviewed changes

mdtoguchi changed the title ~~[SYCL] Warn when total Device code-size causes relocation errors~~ [SYCL][Driver] Warn when total Device code-size causes relocation errors Oct 22, 2026

mdtoguchi reviewed Oct 22, 2026

View reviewed changes

Comment thread clang/test/Driver/sycl-huge-device-code-warning.cpp

bviyer requested a review from mdtoguchi October 22, 2026 17:23

Merge branch 'sycl' into bviyer-make-better-warning

d7ba756

mdtoguchi reviewed Oct 23, 2026

View reviewed changes

Comment thread clang/lib/Driver/ToolChains/Gnu.cpp Outdated

bviyer added 2 commits October 23, 2026 16:47

Added changes requested by Mike

00362e2

Merge remote-tracking branch 'origin/bviyer-make-better-warning' into…

31b1ee1

… bviyer-make-better-warning

bviyer requested a review from mdtoguchi October 26, 2026 20:36

mdtoguchi reviewed Oct 26, 2026

View reviewed changes

mdtoguchi reviewed Oct 27, 2026

View reviewed changes

Fixed issues mentioned by Mike

89c5b11

bviyer requested a review from mdtoguchi October 27, 2026 19:52

mdtoguchi reviewed Oct 27, 2026

View reviewed changes

	// RUN: %clangxx -fsycl %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1 \|\| true
	// RUN: %clangxx -fsycl %t-big.o %t-small.o -o %t.out > %t-warn.log 2>&1

		// short-circuit instead of actually invoking the (nonexistent, in this
		// test) linker -- that's what we want, since the point here is just to

		EXPECT_TRUE(DiagConsumer->Msgs.empty());
		EXPECT_TRUE(DiagConsumer->Errors.empty());

Conversation

bviyer commented Oct 16, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

elizabethandrews left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants