RE: [AMDGPU] InstCombine: fold invalid calls to amdgcn intrinsics into poison values… by yoonseoch · Pull Request #196151 · llvm/llvm-project

yoonseoch · 2026-05-06T19:26:21Z

Replace a call to amdgpu intrinsic into a poison value when the call is invalid because of "amdgpu-no-" attribute in the caller function.

Upon #186925 (review)

Reintroducing the logic with a fix of special casing kernels, which was reverted by
#192514 (Original: #191904)

Assisted by claude-4.6-sonnet-medium through CURSOR.

…o poison values

llvmorg-github-actions · 2026-05-06T22:34:41Z

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-backend-amdgpu

Author: Yoonseo Choi (yoonseoch)

Changes

Replace a call to amdgpu intrinsic into a poison value when the call is invalid because of "amdgpu-no-" attribute in the caller function.

Upon #186925 (review)

Reintroducing the logic with a fix of special casing kernels, which was reverted by
#192514 (Original: #191904)

Assisted by claude-4.6-sonnet-medium through CURSOR.

Full diff: https://github.com/llvm/llvm-project/pull/196151.diff

8 Files Affected:

(modified) llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp (+80)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.cluster.id.ll (+59)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.id.ll (+32)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.ptr.ll (+48)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.lds.kernel.id.ll (+32)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.queue.ptr.ll (+48)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workgroup.id.ll (+57)
(added) llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workitem.id.ll (+57)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
index 463d63a88f690..d6d4b80c217af 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
@@ -749,6 +749,86 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
 
     return std::nullopt;
   }
+  case Intrinsic::amdgcn_dispatch_ptr: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-dispatch-ptr"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_queue_ptr: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-queue-ptr"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_dispatch_id: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-dispatch-id"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_workgroup_id_x:
+  case Intrinsic::amdgcn_workgroup_id_y:
+  case Intrinsic::amdgcn_workgroup_id_z: {
+    // These attributes are only for non-kernel functions and don't-cares for
+    // kernels. Thus, a kernel having one attr and calling the corresponding
+    // intrinsic is allowed.
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_workgroup_id_x   ? "amdgpu-no-workgroup-id-x"
+        : IID == Intrinsic::amdgcn_workgroup_id_y ? "amdgpu-no-workgroup-id-y"
+                                                  : "amdgpu-no-workgroup-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_workitem_id_x:
+  case Intrinsic::amdgcn_workitem_id_y:
+  case Intrinsic::amdgcn_workitem_id_z: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_workitem_id_x   ? "amdgpu-no-workitem-id-x"
+        : IID == Intrinsic::amdgcn_workitem_id_y ? "amdgpu-no-workitem-id-y"
+                                                 : "amdgpu-no-workitem-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_lds_kernel_id: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-lds-kernel-id"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_cluster_id_x:
+  case Intrinsic::amdgcn_cluster_id_y:
+  case Intrinsic::amdgcn_cluster_id_z: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_cluster_id_x   ? "amdgpu-no-cluster-id-x"
+        : IID == Intrinsic::amdgcn_cluster_id_y ? "amdgpu-no-cluster-id-y"
+                                                : "amdgpu-no-cluster-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
   case Intrinsic::amdgcn_rcp: {
     Value *Src = II.getArgOperand(0);
     if (isa<PoisonValue>(Src))
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.cluster.id.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.cluster.id.ll
new file mode 100644
index 0000000000000..3070498d12ca7
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.cluster.id.ll
@@ -0,0 +1,59 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck %s --check-prefix=GFX1250
+
+; When amdgpu-no-cluster-id-x is set, calling the intrinsic is UB and the
+; call should be folded to poison on any target.
+define i32 @no_cluster_id_x() #0 {
+; GFX1250-LABEL: define i32 @no_cluster_id_x(
+; GFX1250-SAME: ) #[[ATTR0:[0-9]+]] {
+; GFX1250-NEXT:  [[ENTRY:.*:]]
+; GFX1250-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.cluster.id.x()
+  ret i32 %tmp
+}
+
+define i32 @no_cluster_id_y() #1 {
+; GFX1250-LABEL: define i32 @no_cluster_id_y(
+; GFX1250-SAME: ) #[[ATTR1:[0-9]+]] {
+; GFX1250-NEXT:  [[ENTRY:.*:]]
+; GFX1250-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.cluster.id.y()
+  ret i32 %tmp
+}
+
+define i32 @no_cluster_id_z() #2 {
+; GFX1250-LABEL: define i32 @no_cluster_id_z(
+; GFX1250-SAME: ) #[[ATTR2:[0-9]+]] {
+; GFX1250-NEXT:  [[ENTRY:.*:]]
+; GFX1250-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.cluster.id.z()
+  ret i32 %tmp
+}
+
+; Without the attribute, the call folds to poison on targets without the
+; cluster feature (gfx1100), but is kept on cluster-capable targets (gfx1250).
+define i32 @cluster_id_x_no_attr() {
+; GFX1250-LABEL: define i32 @cluster_id_x_no_attr(
+; GFX1250-SAME: ) #[[ATTR3:[0-9]+]] {
+; GFX1250-NEXT:  [[ENTRY:.*:]]
+; GFX1250-NEXT:    [[TMP:%.*]] = call i32 @llvm.amdgcn.cluster.id.x()
+; GFX1250-NEXT:    ret i32 [[TMP]]
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.cluster.id.x()
+  ret i32 %tmp
+}
+
+declare i32 @llvm.amdgcn.cluster.id.x()
+declare i32 @llvm.amdgcn.cluster.id.y()
+declare i32 @llvm.amdgcn.cluster.id.z()
+
+attributes #0 = { "amdgpu-no-cluster-id-x" }
+attributes #1 = { "amdgpu-no-cluster-id-y" }
+attributes #2 = { "amdgpu-no-cluster-id-z" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.id.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.id.ll
new file mode 100644
index 0000000000000..d55beb67ca822
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.id.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-dispatch-id is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define i64 @no_dispatch_id() #0 {
+; CHECK-LABEL: define i64 @no_dispatch_id(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i64 poison
+;
+entry:
+  %tmp = call i64 @llvm.amdgcn.dispatch.id()
+  ret i64 %tmp
+}
+
+; Without the attribute the call is kept.
+define i64 @dispatch_id_no_attr() {
+; CHECK-LABEL: define i64 @dispatch_id_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call i64 @llvm.amdgcn.dispatch.id()
+; CHECK-NEXT:    ret i64 [[TMP]]
+;
+entry:
+  %tmp = call i64 @llvm.amdgcn.dispatch.id()
+  ret i64 %tmp
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare i64 @llvm.amdgcn.dispatch.id()
+
+attributes #0 = { "amdgpu-no-dispatch-id" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.ptr.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.ptr.ll
new file mode 100644
index 0000000000000..74cd63be6c706
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.ptr.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-dispatch-ptr is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define ptr addrspace(4) @no_dispatch_ptr() #0 {
+; CHECK-LABEL: define ptr addrspace(4) @no_dispatch_ptr(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret ptr addrspace(4) poison
+;
+entry:
+  %tmp = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+  ret ptr addrspace(4) %tmp
+}
+
+; When amdgpu-no-dispatch-ptr is set, a load from the intrinsic's returned
+; pointer is also UB. The call folds to poison and loading from poison yields
+; poison via existing InstCombine rules.
+define i32 @no_dispatch_ptr_load() #0 {
+; CHECK-LABEL: define i32 @no_dispatch_ptr_load(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    store i1 true, ptr poison, align 1
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %ptr = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+  %val = load i32, ptr addrspace(4) %ptr
+  ret i32 %val
+}
+
+; Without the attribute the call is kept.
+define ptr addrspace(4) @dispatch_ptr_no_attr() {
+; CHECK-LABEL: define ptr addrspace(4) @dispatch_ptr_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+; CHECK-NEXT:    ret ptr addrspace(4) [[TMP]]
+;
+entry:
+  %tmp = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+  ret ptr addrspace(4) %tmp
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+
+attributes #0 = { "amdgpu-no-dispatch-ptr" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.lds.kernel.id.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.lds.kernel.id.ll
new file mode 100644
index 0000000000000..c699b21c802f4
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.lds.kernel.id.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-lds-kernel-id is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define i32 @no_lds_kernel_id() #0 {
+; CHECK-LABEL: define i32 @no_lds_kernel_id(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.lds.kernel.id()
+  ret i32 %tmp
+}
+
+; Without the attribute the call is kept.
+define i32 @lds_kernel_id_no_attr() {
+; CHECK-LABEL: define i32 @lds_kernel_id_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call i32 @llvm.amdgcn.lds.kernel.id()
+; CHECK-NEXT:    ret i32 [[TMP]]
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.lds.kernel.id()
+  ret i32 %tmp
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare i32 @llvm.amdgcn.lds.kernel.id()
+
+attributes #0 = { "amdgpu-no-lds-kernel-id" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.queue.ptr.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.queue.ptr.ll
new file mode 100644
index 0000000000000..013e70e0b38d8
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.queue.ptr.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-queue-ptr is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define ptr addrspace(4) @no_queue_ptr() #0 {
+; CHECK-LABEL: define ptr addrspace(4) @no_queue_ptr(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret ptr addrspace(4) poison
+;
+entry:
+  %tmp = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+  ret ptr addrspace(4) %tmp
+}
+
+; When amdgpu-no-queue-ptr is set, a load from the intrinsic's returned
+; pointer is also UB. The call folds to poison and loading from poison yields
+; poison via existing InstCombine rules.
+define i32 @no_queue_ptr_load() #0 {
+; CHECK-LABEL: define i32 @no_queue_ptr_load(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    store i1 true, ptr poison, align 1
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %ptr = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+  %val = load i32, ptr addrspace(4) %ptr
+  ret i32 %val
+}
+
+; Without the attribute the call is kept.
+define ptr addrspace(4) @queue_ptr_no_attr() {
+; CHECK-LABEL: define ptr addrspace(4) @queue_ptr_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+; CHECK-NEXT:    ret ptr addrspace(4) [[TMP]]
+;
+entry:
+  %tmp = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+  ret ptr addrspace(4) %tmp
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+
+attributes #0 = { "amdgpu-no-queue-ptr" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workgroup.id.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workgroup.id.ll
new file mode 100644
index 0000000000000..34efd980c3e30
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workgroup.id.ll
@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-workgroup-id-x is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define i32 @no_workgroup_id_x() #0 {
+; CHECK-LABEL: define i32 @no_workgroup_id_x(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workgroup.id.x()
+  ret i32 %tmp
+}
+
+define i32 @no_workgroup_id_y() #1 {
+; CHECK-LABEL: define i32 @no_workgroup_id_y(
+; CHECK-SAME: ) #[[ATTR1:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workgroup.id.y()
+  ret i32 %tmp
+}
+
+define i32 @no_workgroup_id_z() #2 {
+; CHECK-LABEL: define i32 @no_workgroup_id_z(
+; CHECK-SAME: ) #[[ATTR2:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workgroup.id.z()
+  ret i32 %tmp
+}
+
+; Without the attribute the calls are kept.
+define i32 @workgroup_id_x_no_attr() {
+; CHECK-LABEL: define i32 @workgroup_id_x_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call i32 @llvm.amdgcn.workgroup.id.x()
+; CHECK-NEXT:    ret i32 [[TMP]]
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workgroup.id.x()
+  ret i32 %tmp
+}
+
+declare i32 @llvm.amdgcn.workgroup.id.x()
+declare i32 @llvm.amdgcn.workgroup.id.y()
+declare i32 @llvm.amdgcn.workgroup.id.z()
+
+attributes #0 = { "amdgpu-no-workgroup-id-x" }
+attributes #1 = { "amdgpu-no-workgroup-id-y" }
+attributes #2 = { "amdgpu-no-workgroup-id-z" }
diff --git a/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workitem.id.ll b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workitem.id.ll
new file mode 100644
index 0000000000000..3972b9ae8d168
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workitem.id.ll
@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='instcombine' -mtriple=amdgcn-amd-amdhsa < %s | FileCheck %s
+
+; When amdgpu-no-workitem-id-x is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define i32 @no_workitem_id_x() #0 {
+; CHECK-LABEL: define i32 @no_workitem_id_x(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workitem.id.x()
+  ret i32 %tmp
+}
+
+define i32 @no_workitem_id_y() #1 {
+; CHECK-LABEL: define i32 @no_workitem_id_y(
+; CHECK-SAME: ) #[[ATTR1:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workitem.id.y()
+  ret i32 %tmp
+}
+
+define i32 @no_workitem_id_z() #2 {
+; CHECK-LABEL: define i32 @no_workitem_id_z(
+; CHECK-SAME: ) #[[ATTR2:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 poison
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workitem.id.z()
+  ret i32 %tmp
+}
+
+; Without the attribute the calls are kept.
+define i32 @workitem_id_x_no_attr() {
+; CHECK-LABEL: define i32 @workitem_id_x_no_attr() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP:%.*]] = call i32 @llvm.amdgcn.workitem.id.x()
+; CHECK-NEXT:    ret i32 [[TMP]]
+;
+entry:
+  %tmp = call i32 @llvm.amdgcn.workitem.id.x()
+  ret i32 %tmp
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x()
+declare i32 @llvm.amdgcn.workitem.id.y()
+declare i32 @llvm.amdgcn.workitem.id.z()
+
+attributes #0 = { "amdgpu-no-workitem-id-x" }
+attributes #1 = { "amdgpu-no-workitem-id-y" }
+attributes #2 = { "amdgpu-no-workitem-id-z" }

Copilot

Pull request overview

This PR reintroduces/extends AMDGPU-specific InstCombine folding so calls to certain llvm.amdgcn.* intrinsics are replaced with poison when the caller function is marked with the corresponding amdgpu-no-* attribute (i.e., the intrinsic use is considered invalid/UB under that contract). It also adds focused InstCombine tests for each intrinsic/attribute pair.

Changes:

Teach AMDGPUInstCombineIntrinsic to fold invalid calls to amdgcn_dispatch_ptr, amdgcn_queue_ptr, amdgcn_dispatch_id, amdgcn_workgroup_id_{x,y,z}, amdgcn_workitem_id_{x,y,z}, amdgcn_lds_kernel_id, and amdgcn_cluster_id_{x,y,z} into poison when the matching amdgpu-no-* function attribute is present.
Add new InstCombine regression tests covering these fold-to-poison behaviors for non-entry functions.
Add a target-feature-specific test for cluster.id.* on a cluster-capable GPU.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp	Adds InstCombine folds that replace invalid AMDGPU intrinsic calls with `poison` based on `amdgpu-no-*` attributes, with special-casing for entry functions.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workitem.id.ll	New test coverage for folding `workitem.id.{x,y,z}` to `poison` under `amdgpu-no-workitem-id-*`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.workgroup.id.ll	New test coverage for folding `workgroup.id.{x,y,z}` to `poison` under `amdgpu-no-workgroup-id-*`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.queue.ptr.ll	New test coverage for folding `queue.ptr` to `poison` under `amdgpu-no-queue-ptr`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.lds.kernel.id.ll	New test coverage for folding `lds.kernel.id` to `poison` under `amdgpu-no-lds-kernel-id`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.ptr.ll	New test coverage for folding `dispatch.ptr` to `poison` under `amdgpu-no-dispatch-ptr`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.dispatch.id.ll	New test coverage for folding `dispatch.id` to `poison` under `amdgpu-no-dispatch-id`.
llvm/test/Transforms/InstCombine/AMDGPU/llvm.amdgcn.cluster.id.ll	New test coverage for folding `cluster.id.{x,y,z}` to `poison` under `amdgpu-no-cluster-id-*` (currently only on a cluster-capable target).

+  case Intrinsic::amdgcn_dispatch_ptr: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-dispatch-ptr"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_queue_ptr: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-queue-ptr"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_dispatch_id: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+    if (Fn->hasFnAttribute("amdgpu-no-dispatch-id"))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }


+  case Intrinsic::amdgcn_workgroup_id_x:
+  case Intrinsic::amdgcn_workgroup_id_y:
+  case Intrinsic::amdgcn_workgroup_id_z: {
+    // These attributes are only for non-kernel functions and don't-cares for
+    // kernels. Thus, a kernel having one attr and calling the corresponding
+    // intrinsic is allowed.
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_workgroup_id_x   ? "amdgpu-no-workgroup-id-x"
+        : IID == Intrinsic::amdgcn_workgroup_id_y ? "amdgpu-no-workgroup-id-y"
+                                                  : "amdgpu-no-workgroup-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }
+  case Intrinsic::amdgcn_workitem_id_x:
+  case Intrinsic::amdgcn_workitem_id_y:
+  case Intrinsic::amdgcn_workitem_id_z: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_workitem_id_x   ? "amdgpu-no-workitem-id-x"
+        : IID == Intrinsic::amdgcn_workitem_id_y ? "amdgpu-no-workitem-id-y"
+                                                 : "amdgpu-no-workitem-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;
+  }


+  case Intrinsic::amdgcn_cluster_id_x:
+  case Intrinsic::amdgcn_cluster_id_y:
+  case Intrinsic::amdgcn_cluster_id_z: {
+    const Function *Fn = II.getFunction();
+    if (AMDGPU::isEntryFunctionCC(Fn->getCallingConv()))
+      return std::nullopt;
+
+    StringRef Attr =
+        IID == Intrinsic::amdgcn_cluster_id_x   ? "amdgpu-no-cluster-id-x"
+        : IID == Intrinsic::amdgcn_cluster_id_y ? "amdgpu-no-cluster-id-y"
+                                                : "amdgpu-no-cluster-id-z";
+    if (Fn->hasFnAttribute(Attr))
+      return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType()));
+    return std::nullopt;


+; Without the attribute, the call folds to poison on targets without the
+; cluster feature (gfx1100), but is kept on cluster-capable targets (gfx1250).
+define i32 @cluster_id_x_no_attr() {
+; GFX1250-LABEL: define i32 @cluster_id_x_no_attr(
+; GFX1250-SAME: ) #[[ATTR3:[0-9]+]] {
+; GFX1250-NEXT:  [[ENTRY:.*:]]
+; GFX1250-NEXT:    [[TMP:%.*]] = call i32 @llvm.amdgcn.cluster.id.x()
+; GFX1250-NEXT:    ret i32 [[TMP]]
+;


+; When amdgpu-no-dispatch-ptr is set, calling the intrinsic is UB and the
+; call should be folded to poison.
+define ptr addrspace(4) @no_dispatch_ptr() #0 {
+; CHECK-LABEL: define ptr addrspace(4) @no_dispatch_ptr(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret ptr addrspace(4) poison
+;
+entry:
+  %tmp = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+  ret ptr addrspace(4) %tmp
+}


shiltian · 2026-05-06T23:05:12Z

I'm not really sure it's a good idea to special-case the semantics of those attributes. In other words, when the attribute appears on a kernel entry function, it means nothing, but when it appears on a regular device function, it suddenly has meaning.

IMO, we should have consistent behavior and avoid giving the same attribute different meanings depending on where it appears.

steffenlarsen · 2026-05-07T06:18:42Z

The pattern is a little repetitive. I'd suggest adding

static StringRef getAMDGPUConflictFuncAttrForIntrinsic(Intrinsic::ID IID) { switch (IID) { case Intrinsic::amdgcn_dispatch_ptr: return "amdgpu-no-dispatch-ptr"; case Intrinsic::amdgcn_queue_ptr: return "amdgpu-no-queue-ptr"; case Intrinsic::amdgcn_dispatch_id: return "amdgpu-no-dispatch-id"; case Intrinsic::amdgcn_workgroup_id_x: return "amdgpu-no-workgroup-id-x"; case Intrinsic::amdgcn_workgroup_id_y: return "amdgpu-no-workgroup-id-y"; case Intrinsic::amdgcn_workgroup_id_z: return "amdgpu-no-workgroup-id-z"; case Intrinsic::amdgcn_workitem_id_x: return "amdgpu-no-workitem-id-x"; case Intrinsic::amdgcn_workitem_id_y: return "amdgpu-no-workitem-id-y"; case Intrinsic::amdgcn_workitem_id_z: return "amdgpu-no-workitem-id-z"; case Intrinsic::amdgcn_lds_kernel_id: return "amdgpu-no-lds-kernel-id"; case Intrinsic::amdgcn_cluster_id_x: return "amdgpu-no-cluster-id-x"; case Intrinsic::amdgcn_cluster_id_y: return "amdgpu-no-cluster-id-y"; case Intrinsic::amdgcn_cluster_id_z: return "amdgpu-no-cluster-id-z"; default: llvm_unreachable("unexpected intrinsic"); return StringRef(); } } static std::optional<Instruction *> getPoisonedAMDGPUIntrinsicReplacement(Instruction &II, InstCombiner &IC) { Intrinsic::ID IID = II.getIntrinsicID(); StringRef ConflictAttr = getAMDGPUConflictFuncAttrForIntrinsic(IID); const Function *Fn = II.getFunction(); if (!AMDGPU::isEntryFunctionCC(Fn->getCallingConv()) && Fn->hasFnAttribute(ConflictAttr)) return IC.replaceInstUsesWith(II, PoisonValue::get(II.getType())); return std::nullopt; }

and then change this to

case Intrinsic::amdgcn_dispatch_ptr: case Intrinsic::amdgcn_queue_ptr: case Intrinsic::amdgcn_dispatch_id: case Intrinsic::amdgcn_workgroup_id_x: case Intrinsic::amdgcn_workgroup_id_y: case Intrinsic::amdgcn_workgroup_id_z: case Intrinsic::amdgcn_workitem_id_x: case Intrinsic::amdgcn_workitem_id_y: case Intrinsic::amdgcn_workitem_id_z: case Intrinsic::amdgcn_lds_kernel_id: case Intrinsic::amdgcn_cluster_id_x: case Intrinsic::amdgcn_cluster_id_y: case Intrinsic::amdgcn_cluster_id_z: return getPoisonedAMDGPUIntrinsicReplacement(II, IC);

RE: [AMDGPU] InstCombine: fold invalid calls to amdgcn intrinsics int…

a88cc33

…o poison values

yoonseoch marked this pull request as ready for review May 6, 2026 22:34

Copilot AI review requested due to automatic review settings May 6, 2026 22:34

llvmorg-github-actions Bot added backend:AMDGPU llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels May 6, 2026

yoonseoch requested review from arsenm and shiltian May 6, 2026 22:34

Copilot started reviewing on behalf of yoonseoch May 6, 2026 22:35 View session

Copilot AI reviewed May 6, 2026

View reviewed changes

steffenlarsen reviewed May 7, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

RE: [AMDGPU] InstCombine: fold invalid calls to amdgcn intrinsics into poison values…#196151

RE: [AMDGPU] InstCombine: fold invalid calls to amdgcn intrinsics into poison values…#196151
yoonseoch wants to merge 1 commit into
llvm:mainfrom
yoonseoch:re-amdgpu-no-poison

yoonseoch commented May 6, 2026 •

edited

Loading

Uh oh!

llvmorg-github-actions Bot commented May 6, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

shiltian commented May 6, 2026

Uh oh!

steffenlarsen May 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

yoonseoch commented May 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmorg-github-actions Bot commented May 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

shiltian commented May 6, 2026

Uh oh!

steffenlarsen May 7, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

yoonseoch commented May 6, 2026 •

edited

Loading

llvmorg-github-actions Bot commented May 6, 2026 •

edited

Loading