[libspirv] Replace libclc-remangler with SYCLRemangleLibspirvPass by wenju-he · Pull Request #21588 · intel/llvm

wenju-he · 2026-03-23T08:48:56Z

Remove libclc-remangler tool and replace it with SYCLRemangleLibspirvPass
that remangles OpenCL C mangling in libspirv to match with SYCL mangling.
The pass ports the functionalities in the tool into compile pipeline.

Motivation:

libclc-remangler executable doesn't fit in runtime build.
libclc-remangler can't justify as a standalone tool in llvm/tools.

Functional changes:

Add SYCLRemangleLibspirvPass. It uses Itanium demangler and adapts SPIR type system and SPIR mangler from SPIRV-LLVM-Translator.
The pass runs at the end of compile pipeline for each .cl file.
libspirv variants are created without separate remangling step.
Drop libclc-remangler tool.

llvm-diff changes to remangled libspirv:

_clc functions from CLC library are now not remangled because we use
a single clc library that defaults to OpenCL C mangling.
Some _clc functions from CLC library are no longer inlined for NVPTX.
Previously, these were called only once and inlined before remangling.
Now, because remangling can generate two caller variants, the _clc
function is called twice, causing the inliner to skip them.
Changes in _CLC_DEFINE_MIPMAP_BINDLESS_READS_BUILTIN also exposed a bug in libclc-remangler which would remangle Z30__spirv_ImageSampleExplicitLodImDv4_cDv3_fET0_T_T1_iS4_S4 (correct) to Z30__spirv_ImageSampleExplicitLodImDv4_aDv3_fET0_T_T1_iS1_S1
This bug was hidden since correct symbol was hardcoded in the source.
Fixed 4 __spirv_ImageArrayWrite builtins in nvptx l64.signed_char, e.g.
_Z23__spirv_ImageArrayWriteIyiiEvT_T0_T1_i -> Z23__spirv_ImageArrayWriteIyiiEvT_T0_iT1;
_Z23__spirv_ImageArrayWriteImiiEvT_T0_T1_i -> Z23__spirv_ImageArrayWriteImiiEvT_T0_iT1.

Remove libclc-remangler tool and replace it with SYCLBuiltinRemanglePass that remangles SPIR-V built-ins in SYCL user device code to match OpenCL C mangling in libspirv. The pass is inverse of what libclc-remangler did. Motivation: - libclc-remangler executable doesn't fit in runtime build. - libclc-remangler can't justify as a standalone tool in llvm/tools. Changes: - Add SYCLBuiltinRemanglePass. It uses Itanium demangler and adapts SPIR type system and SPIR mangler from SPIRV-LLVM-Translator. - Type transformations: - long long -> long (OpenCL has no long long) - long -> int (Windows or 32-bit only) - signed char -> char - char -> unsigned char (only if char is signed on the host) - _Float16 -> half (with a few w/a for native-cpu libdevice) - Address space adjustments if target's default addrspace is private. - libspirv is now linked into user code after the optimization pipeline by -mlink-builtin-bitcode-postopt flag since the linking must be after remangling. In the future, the linking will be moved to LTO. - Changes in _CLC_DEFINE_MIPMAP_BINDLESS_READS_BUILTIN also exposed a bug in libclc-remangler which would remangle _Z30__spirv_ImageSampleExplicitLodImDv4_cDv3_fET0_T_T1_iS4_S4_ (correct) to _Z30__spirv_ImageSampleExplicitLodImDv4_aDv3_fET0_T_T1_iS1_S1_ This bug was hidden since correct symbol was hardcoded in the source. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Maetveis

libclc changes LGTM. I didn't check the new pass or any of the other changes.

Maetveis · 2026-03-23T10:50:24Z

sycl/doc/design/CompilerAndRuntimeDesign.md

-When the SYCL compiler is in device mode and targeting the NVPTX backend, the
-compiler exposes NVPTX builtins supported by clang.
+incompatible libclc built-ins. When building a SYCL application targeting the
+CUDA backend, the SYCLBuiltinRemangle pass remangles SPIR-V builtins in device


I'm wondering could we make clang emit calls to these functions using the correct name mangling from the start instead of re-mangling as a pass?
I was thinking a function attribute that allows to switch the name mangling to be OpenCL compatible could theoretically work for this. We would need something similar if we wanted to author libclc sources in C++.

I'm wondering could we make clang emit calls to these functions using the correct name mangling from the start instead of re-mangling as a pass?

This is interesting idea, but I don't know about how to implement the automatic mapping.

We would need something similar if we wanted to author libclc sources in C++.

Implement libspirv in C++ and compile in sycl device mode like libdevice does can indeed avoid the remangling issue. However, one problem is that it will be very difficult to share implementation between libclc OpenCL and libspirv builtin wrappers, since eventually the sharing requires a stable interface which SPV-IR intends to be.

clang/test/Driver/sycl-libspirv-toolchain.cpp

llvm/include/llvm/SYCLLowerIR/SYCLBuiltinRemangle.h

llvm/lib/SYCLLowerIR/SYCLBuiltinRemangle.cpp

llvm/test/SYCLLowerIR/SYCLBuiltinRemangle/char-signedness.ll

llvm/test/SYCLLowerIR/SYCLBuiltinRemangle/char-types.ll

llvm/test/SYCLLowerIR/SYCLBuiltinRemangle/float16-types.ll

elizabethandrews

FE changes just add a pass. FE changes LGTM. I have not reviewed the functionality of the pass.

…RemangleLibspirvPass Not inlined __clc_ functions are not remangled because we only have a single clc library that has OpenCL C mangling.

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

YuriPlyakhin · 2026-03-30T22:51:03Z

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

+          if (AS == 1)
+            Attr = SPIR::ATTR_GLOBAL;
+          else if (AS == 2)
+            Attr = SPIR::ATTR_CONSTANT;
+          else if (AS == 3)
+            Attr = SPIR::ATTR_LOCAL;
+          else if (AS == 4)
+            Attr = SPIR::ATTR_GENERIC_EXPLICIT;


what is AS == 5 or something? should we add assert or llvm_unreachable?

what is AS == 5 or something? should we add assert or llvm_unreachable?

This will be handled in follow-up PR since it will cause llvm-diff change to remangled libspirv. Current address space mapping is incorrect for non-spirv targets. The issues exists before this PR.

could you please add TODO comment and maybe point to github issue?

could you please add TODO comment and maybe point to github issue?

added comment linking to #21664

could you please add TODO comment and maybe point to github issue?

added comment linking to #21664

sorry, the above issue is a false alarm. It is that this pass should not use SPIR address space values. Fixed in b0fba79

YuriPlyakhin · 2026-03-30T22:55:09Z

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

+    // Skip __clc_ functions. If we remangle them, it should be done in clc
+    // library and it requires creating multiple variants of clc libaries,
+    // with mangling scheme aligning with the corresponding libspirv remangling.
+    if (F.isDeclaration() && F.getName().contains("__clc_"))


this condition skips only __clc_ declarations. Should we also skip definitions?

this condition is correct. Only __clc_ declarations from CLC library, which is built before libspirv build, should be ignored.
__clc_ definitions in libspirv should be remangled because they are not from CLC library. Actually libspirv shouldn't provide any __clc_ definitions, but this is just current status.
I think it is feasible to skip __clc_ definitions here since they are not public APIs, but I'd not do so in this PR to avoid llvm-diff change from old remangled libspirv prior to this PR.

I see. could you please clarify the comment then. currently it is saying to skip __clc_ functions, but without details you provided in the reply, it is not clear, why we are skipping only declarations, but not definitions.

libspirv should not contain _clc definitions. Adding a comment to justify a mistake, only to remove it later, don't have much value.

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

YuriPlyakhin · 2026-03-30T23:43:53Z

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

+      MangleVisitor Visitor(MangledName, Stream, 1);
+      for (const auto &TemplateArg : TemplateArgs) {
+        MangleError Err = TemplateArg->accept(&Visitor);
+        if (Err != MANGLE_SUCCESS)


The only value of MangleError is MANGLE_SUCCESS. What is the scenario, where Err would be different? Should more MangleError cases be added to enum, like MANGLE_FAILURE and then we should use it somewhere?

Current error macros in llvm-spirv mangler don't suite here. I kept MANGLE_SUCCESS here for API consistency. Should we just delete MANGLE_SUCCESS?

yes, I think it would be fine to delete, if it is not used meaningfully

yes, I think it would be fine to delete, if it is not used meaningfully

deleted MANGLE_SUCCESS

clang/lib/CodeGen/BackendUtil.cpp

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

YuriPlyakhin · 2026-03-31T21:54:38Z

llvm/lib/SYCLLowerIR/SYCLRemangleLibspirv.cpp

+//     CloneOriginal recreates it so callers using OpenCL type names still link.
+//     Clone source is found by applying CloneOriginal's map to original name:
+//       char -> signed char (signed) / unsigned char (unsigned)
+//       _Float16 -> half, long -> int (L32) or long long (L64),


based on further comments, there is no CloneOriginal entry for 'half'

Suggested change

// _Float16 -> half, long -> int (L32) or long long (L64),

// long -> int (L32) or long long (L64),

based on further comments, there is no CloneOriginal entry for 'half'

done, thanks

YuriPlyakhin

4 comments left to address for files owned by me.

wenju-he requested review from a team, Maetveis and cperkinsintel as code owners March 23, 2026 08:48

wenju-he requested a review from bratpiorka March 23, 2026 08:49

Maetveis approved these changes Mar 23, 2026

View reviewed changes

mdtoguchi reviewed Mar 23, 2026

View reviewed changes

clang/test/Driver/sycl-libspirv-toolchain.cpp Outdated Show resolved Hide resolved