[SPIRV] Support -fembed-bitcode=marker for non-shader modules #162082

jmmartinez · 2025-10-06T12:54:43Z

-fembed-bitcode=marker gets lowered as a [0 x i8] zeroinitialized global
variable. This is interpreted as a runtime-array and is not supported in
non-shaders.

To work around this, we replace the [0 x i8] by a zeroinitialized
single-element array.

I'm not sure this is the way we want to support this, neither for shaders (using runtime arrays) nor for kernels (using the 1-sized zeroinitialized array).
A part of me would like to just delete the variable since we do not really need the delimiter.

Depends on #162081

llvmbot · 2025-10-06T12:55:17Z

@llvm/pr-subscribers-backend-spir-v

Author: Juan Manuel Martinez Caamaño (jmmartinez)

Changes

-fembed-bitcode=marker gets lowered as a [0 x i8] zeroinitialized global
variable. This is interpreted as a runtime-array and is not supported in
non-shaders.

To work around this, we replace the [0 x i8] by a zeroinitialized
single-element array.

I'm not sure this is the way we want to support this, neither for shaders (using runtime arrays) nor for kernels (using the 1-sized zeroinitialized array).
A part of me would like to just delete the variable since we do not really need the delimiter.

Depends on #162081

Full diff: https://github.com/llvm/llvm-project/pull/162082.diff

4 Files Affected:

(modified) llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp (+36)
(added) llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll (+21)
(added) llvm/test/CodeGen/SPIRV/fembed-bitcode-marker.ll (+21)
(added) llvm/test/CodeGen/SPIRV/fembed-bitcode.ll (+25)

diff --git a/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp b/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
index 9f2e07508a36a..447fbb9d1ac0a 100644
--- a/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
@@ -236,6 +236,8 @@ class SPIRVEmitIntrinsics
 
   Instruction *buildLogicalAccessChainFromGEP(GetElementPtrInst &GEP);
 
+  bool promoteEmbddedBitcodeMarker(Module &M) const;
+
 public:
   static char ID;
   SPIRVEmitIntrinsics(SPIRVTargetMachine *TM = nullptr)
@@ -3007,9 +3009,43 @@ void SPIRVEmitIntrinsics::parseFunDeclarations(Module &M) {
   }
 }
 
+bool SPIRVEmitIntrinsics::promoteEmbddedBitcodeMarker(Module &M) const {
+  const SPIRVSubtarget *STI = TM->getSubtargetImpl();
+  if (STI->isShader())
+    return false;
+
+  GlobalVariable *EmbeddedBitcode = M.getNamedGlobal("llvm.embedded.module");
+  if (!EmbeddedBitcode)
+    return false;
+
+  ArrayType *AT = cast<ArrayType>(EmbeddedBitcode->getValueType());
+  if (AT->getNumElements() != 0)
+    return false;
+
+  // When compiling with -fembed-bitcode=marker, LLVM generates a [0 x i8]
+  // zeroinitialized global variable containing the bitcode. This results in an
+  // assert outside of shaders. As a workaround, we replace this global with a
+  // zeroinitialized [1 x i8].
+  ArrayType *AT1 = ArrayType::get(AT->getElementType(), 1);
+  Constant *ZeroInit = Constant::getNullValue(AT1);
+  GlobalVariable *NewEmbeddedBitcode = new GlobalVariable(
+      AT1, EmbeddedBitcode->isConstant(), EmbeddedBitcode->getLinkage(),
+      ZeroInit, "", EmbeddedBitcode->getThreadLocalMode(),
+      EmbeddedBitcode->getAddressSpace(),
+      EmbeddedBitcode->isExternallyInitialized());
+  NewEmbeddedBitcode->setSection(NewEmbeddedBitcode->getSection());
+  NewEmbeddedBitcode->takeName(EmbeddedBitcode);
+
+  M.insertGlobalVariable(NewEmbeddedBitcode);
+  EmbeddedBitcode->replaceAllUsesWith(NewEmbeddedBitcode);
+  EmbeddedBitcode->eraseFromParent();
+  return true;
+}
+
 bool SPIRVEmitIntrinsics::runOnModule(Module &M) {
   bool Changed = false;
 
+  Changed |= promoteEmbddedBitcodeMarker(M);
   parseFunDeclarations(M);
   insertConstantsForFPFastMathDefault(M);
 
diff --git a/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll b/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll
new file mode 100644
index 0000000000000..20ac543486766
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll
@@ -0,0 +1,21 @@
+; RUN: llc -verify-machineinstrs -mtriple=spirv-vulkan-unknown %s -o - | FileCheck %s
+
+@llvm.embedded.module = private constant [0 x i8] zeroinitializer, section ".llvmbc", align 1
+@llvm.cmdline = private constant [5 x i8] c"-cc1\00", section ".llvmcmd", align 1
+@llvm.compiler.used = appending global [2 x ptr] [ptr @llvm.embedded.module, ptr @llvm.cmdline], section "llvm.metadata"
+
+; CHECK-DAG: OpName [[FOO:%[0-9]+]] "foo"
+; CHECK-DAG: OpName [[MODULE:%[0-9]+]] "llvm.embedded.module"
+; CHECK-DAG: [[INT8:%[0-9]+]] = OpTypeInt 8 0
+; CHECK-DAG: [[RUNTIME_ARRAY_INT8x1:%[0-9]+]] = OpTypeRuntimeArray [[INT8]]
+; CHECK-DAG: [[POINTER:%[0-9]+]] = OpTypePointer Function [[RUNTIME_ARRAY_INT8x1]]
+; CHECK-DAG: [[EMBEDDED_MODULE_INIT:%[0-9]+]] = OpConstantNull [[RUNTIME_ARRAY_INT8x1]]
+; CHECK: [[FOO]] = OpFunction {{.*}} None {{.*}} 
+; CHECK-DAG: {{%[0-9]+}} = OpVariable [[POINTER]] Function [[EMBEDDED_MODULE_INIT]]
+
+define void @foo() #1 {
+entry:
+  ret void
+}
+
+attributes #1 = { "hlsl.numthreads"="4,8,16" "hlsl.shader"="compute" }
diff --git a/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker.ll b/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker.ll
new file mode 100644
index 0000000000000..e3d21fcdc8084
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/fembed-bitcode-marker.ll
@@ -0,0 +1,21 @@
+; RUN: llc -verify-machineinstrs -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s
+
+@llvm.embedded.module = private constant [0 x i8] zeroinitializer, section ".llvmbc", align 1
+@llvm.cmdline = private constant [5 x i8] c"-cc1\00", section ".llvmcmd", align 1
+@llvm.compiler.used = appending global [2 x ptr] [ptr @llvm.embedded.module, ptr @llvm.cmdline], section "llvm.metadata"
+
+; CHECK-DAG: OpName [[FOO:%[0-9]+]] "foo"
+; CHECK-DAG: OpName [[MODULE:%[0-9]+]] "llvm.embedded.module"
+; CHECK-DAG: [[INT8:%[0-9]+]] = OpTypeInt 8 0
+; CHECK-DAG: [[INT32:%[0-9]+]] = OpTypeInt 32 0
+; CHECK-DAG: [[CONST_1_32:%[0-9]+]] = OpConstant [[INT32]] 1
+; CHECK-DAG: [[ARRAY_INT8x1:%[0-9]+]] = OpTypeArray [[INT8]] [[CONST_1_32]]
+; CHECK-DAG: [[POINTER:%[0-9]+]] = OpTypePointer Function [[ARRAY_INT8x1]]
+; CHECK-DAG: [[EMBEDDED_MODULE_INIT:%[0-9]+]] = OpConstantNull [[ARRAY_INT8x1]]
+; CHECK: [[FOO]] = OpFunction {{.*}} None {{.*}}
+; CHECK-DAG: {{%[0-9]+}} = OpVariable [[POINTER]] Function [[EMBEDDED_MODULE_INIT]]
+
+define spir_kernel void @foo() {
+entry:
+  ret void
+}
diff --git a/llvm/test/CodeGen/SPIRV/fembed-bitcode.ll b/llvm/test/CodeGen/SPIRV/fembed-bitcode.ll
new file mode 100644
index 0000000000000..0044e09aceaab
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/fembed-bitcode.ll
@@ -0,0 +1,25 @@
+; RUN: llc -verify-machineinstrs -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s
+
+@llvm.embedded.module = private constant [4 x i8] c"BC\C0\DE", section ".llvmbc", align 1
+@llvm.cmdline = private constant [5 x i8] c"-cc1\00", section ".llvmcmd", align 1
+@llvm.compiler.used = appending global [2 x ptr] [ptr @llvm.embedded.module, ptr @llvm.cmdline], section "llvm.metadata"
+
+; CHECK-DAG: OpName [[FOO:%[0-9]+]] "foo"
+; CHECK-DAG: OpName [[MODULE:%[0-9]+]] "llvm.embedded.module"
+; CHECK-DAG: [[INT8:%[0-9]+]] = OpTypeInt 8 0
+; CHECK-DAG: [[INT32:%[0-9]+]] = OpTypeInt 32 0
+; CHECK-DAG: [[CONST_4_32:%[0-9]+]] = OpConstant [[INT32]] 4
+; CHECK-DAG: [[ARRAY_INT8x4:%[0-9]+]] = OpTypeArray [[INT8]] [[CONST_4_32]]
+; CHECK-DAG: [[POINTER:%[0-9]+]] = OpTypePointer Function [[ARRAY_INT8x4]]
+; CHECK-DAG: [[CONST_B_8:%[0-9]+]] = OpConstant [[INT8]] 66
+; CHECK-DAG: [[CONST_C_8:%[0-9]+]] = OpConstant [[INT8]] 67
+; CHECK-DAG: [[CONST_0xC0_8:%[0-9]+]] = OpConstant [[INT8]] 192
+; CHECK-DAG: [[CONST_0xDE_8:%[0-9]+]] = OpConstant [[INT8]] 222
+; CHECK-DAG: [[EMBEDDED_MODULE_INIT:%[0-9]+]] = OpConstantComposite [[ARRAY_INT8x4]] [[CONST_B_8]] [[CONST_C_8]] [[CONST_0xC0_8]] [[CONST_0xDE_8]]
+; CHECK: [[FOO]] = OpFunction {{.*}} None {{.*}} 
+; CHECK-DAG: {{%[0-9]+}} = OpVariable [[POINTER]] Function [[EMBEDDED_MODULE_INIT]]
+
+define spir_kernel void @foo() {
+entry:
+  ret void
+}

llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp

maarquitos14 · 2025-10-06T16:12:31Z

llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll

@@ -0,0 +1,21 @@
+; RUN: llc -verify-machineinstrs -mtriple=spirv-vulkan-unknown %s -o - | FileCheck %s


Please, add also a check with spirv-val. Applies to all tests.

Done in the latest update. I'll also update the parent PR #162081

The test started failing once I added the spirv-val line. I'm checking this

I think that I have now more context about all of this.

The @llvm.embedded.module stuff only make sense when the spirv flavour is AMD's, and not for the others. In this case, the variables are in the constant addrspace(1) and are addrspace-casted to the generic addrspace(4) when added to llvm.compiler.used.

This is allowed only in amd's flavour but is outside of spirv standard.

I think I'm going to change my strategy:

I'll add an AMD specific pass that is only added to the pipeline if the vendor is AMD to put this kind of stuff over there.

Maybe put this kind of test not in test/CodeGen/SPIRV but in a subfolder test/CodeGen/SPIRV/amd or something like that.

What do you think ?

llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp

maarquitos14 · 2025-10-06T16:16:29Z

There was this thread in a related PR that might be relevant here: #149522 (comment). I'm not against this, but there were concerns in the past with a similar solution, although in this case, the variable might not actually be used. Did you try simply deleting it and seeing what happens? What would be the cons of deleting it?

jmmartinez · 2025-10-07T08:28:35Z

Did you try simply deleting it and seeing what happens? What would be the cons of deleting it?

When-fembed-bitcode=marker is specified, I'm pretty sure that deleting the global could work; but I decided to preserve the global to respect what is expected from the option.

In the shader case, I'm leaning towards handling the bitcode-marker in the same way as for the kernel case.

I would have liked to have an -fembed-bitcode=cmdline-only option; but I prefer to not add yet-another-obscure-option that is exposed to the user. That's why I've settled to modify the backend instead.

…er modules

-fembed-bitcode=marker gets lowered as a [0 x i8] zeroinitialized global variable. This is interpreted as a runtime-array and is not supported in non-shaders. To work around this, we replace the [0 x i8] by a zeroinitialized single-element array.

mgcarrasco · 2025-10-07T15:51:45Z

llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp


  Instruction *buildLogicalAccessChainFromGEP(GetElementPtrInst &GEP);

+  bool promoteEmbeddedBitcodeMarkear(Module &M) const;


"Markear" (possible typo)

maarquitos14 · 2025-10-07T19:01:47Z

llvm/test/CodeGen/SPIRV/fembed-bitcode-marker-shader.ll

@@ -0,0 +1,22 @@
+; RUN: llc -verify-machineinstrs -mtriple=spirv-vulkan-unknown %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -mtriple=spirv-unknown-unknown %s -o - -filetype=obj | spirv-val %}


Suggested change

; RUN: %if spirv-tools %{ llc -mtriple=spirv-unknown-unknown %s -o - -filetype=obj | spirv-val %}

; RUN: %if spirv-tools %{ llc -mtriple=spirv-vulkan-unknown %s -o - -filetype=obj | spirv-val %}

jmmartinez · 2025-10-08T11:36:27Z

I'm closing this PR (and its parent #162081). I think I got this wrong.

The AS of the variables in the tests should not be the default one, it should be 1. But that introduces an addrspace cast from 1 to 0 to add this to llvm.used; which is illegal, even if they all are pseudo-variables.

I'm going to rethink about this and tackle this issue later.

Thanks a lot for the comments still !

jmmartinez requested review from AlexVlx, Keenuts, VyacheslavLevytskyy and maarquitos14 October 6, 2025 12:54

jmmartinez self-assigned this Oct 6, 2025

llvmbot added the backend:SPIR-V label Oct 6, 2025

jmmartinez mentioned this pull request Oct 6, 2025

[amd-spirv-backend][Backend] Identify and address translator/backend delta ROCm/llvm-project#306

Open

jmmartinez removed the request for review from VyacheslavLevytskyy October 6, 2025 16:14

maarquitos14 reviewed Oct 6, 2025

View reviewed changes

jmmartinez force-pushed the spirv/fembed branch from 39d7fbf to 8367a5c Compare October 7, 2025 08:16

jmmartinez added 4 commits October 7, 2025 10:32

Pre-Commit tests: [SPIRV] Support -fembed-bitcode=marker for non-shad…

2cb721d

…er modules

[Review] Add spirv-val call to tests

806bb39

[Review] Fix names

5c2271c

jmmartinez force-pushed the spirv/fembed branch from 8367a5c to 5c2271c Compare October 7, 2025 08:32

mgcarrasco mentioned this pull request Oct 7, 2025

[HIP][SPIRV] Enable the SPIRV backend instead of the translator through an experimental flag. #162282

Open

mgcarrasco reviewed Oct 7, 2025

View reviewed changes

maarquitos14 reviewed Oct 7, 2025

View reviewed changes

maarquitos14 mentioned this pull request Oct 8, 2025

Pre-Commit tests: [SPIRV] Support -fembed-bitcode=marker #162081

Closed

jmmartinez closed this Oct 8, 2025

jmmartinez mentioned this pull request Oct 10, 2025

[amd-spirv-backend][Driver] Develop opt-in pipeline for AMDGPU ROCm/llvm-project#305

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[SPIRV] Support -fembed-bitcode=marker for non-shader modules #162082

[SPIRV] Support -fembed-bitcode=marker for non-shader modules #162082

Uh oh!

jmmartinez commented Oct 6, 2025

Uh oh!

llvmbot commented Oct 6, 2025

Uh oh!

Uh oh!

maarquitos14 Oct 6, 2025

Uh oh!

jmmartinez Oct 7, 2025

Uh oh!

jmmartinez Oct 7, 2025

Uh oh!

jmmartinez Oct 8, 2025

Uh oh!

Uh oh!

maarquitos14 commented Oct 6, 2025

Uh oh!

jmmartinez commented Oct 7, 2025

Uh oh!

mgcarrasco Oct 7, 2025

Uh oh!

maarquitos14 Oct 7, 2025

Uh oh!

jmmartinez commented Oct 8, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		@@ -0,0 +1,21 @@
		; RUN: llc -verify-machineinstrs -mtriple=spirv-vulkan-unknown %s -o - \| FileCheck %s


		Instruction *buildLogicalAccessChainFromGEP(GetElementPtrInst &GEP);

		bool promoteEmbeddedBitcodeMarkear(Module &M) const;

		@@ -0,0 +1,22 @@
		; RUN: llc -verify-machineinstrs -mtriple=spirv-vulkan-unknown %s -o - \| FileCheck %s
		; RUN: %if spirv-tools %{ llc -mtriple=spirv-unknown-unknown %s -o - -filetype=obj \| spirv-val %}

[SPIRV] Support -fembed-bitcode=marker for non-shader modules #162082

[SPIRV] Support -fembed-bitcode=marker for non-shader modules #162082

Uh oh!

Conversation

jmmartinez commented Oct 6, 2025

Uh oh!

llvmbot commented Oct 6, 2025

Uh oh!

Uh oh!

maarquitos14 Oct 6, 2025

Choose a reason for hiding this comment

Uh oh!

jmmartinez Oct 7, 2025

Choose a reason for hiding this comment

Uh oh!

jmmartinez Oct 7, 2025

Choose a reason for hiding this comment

Uh oh!

jmmartinez Oct 8, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

maarquitos14 commented Oct 6, 2025

Uh oh!

jmmartinez commented Oct 7, 2025

Uh oh!

mgcarrasco Oct 7, 2025

Choose a reason for hiding this comment

Uh oh!

maarquitos14 Oct 7, 2025

Choose a reason for hiding this comment

Uh oh!

jmmartinez commented Oct 8, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants