[DirectX] Denote dx.resource.getpointer with IntrInaccessibleMemOnly and IntrReadMem#193593
Merged
Conversation
…ly` and `IntrReadMem` `IntrConvergent` was originally added to `dx.resource.getpointer` to prevent optimization passes (SimplifyCFG, GVN) from sinking the intrinsic out of control flow branches, which would create phi nodes on the returned pointer. Using `IntrInaccessibleMemOnly` and `IntrReadMem` semantics still prevent passes from merging or sinking identical calls across branches. However, this allows the call to be moved within a single control flow path and allows for further legal optimizations. Updates relevant tests and adds a new test to demonstrate a potential optimization
Member
|
@llvm/pr-subscribers-llvm-ir @llvm/pr-subscribers-backend-directx Author: Finn Plummer (inbelic) Changes
Using Updates relevant tests and adds a new test to demonstrate a now legal potential optimization. Assisted by: Claude Opus 4.6 Full diff: https://github.com/llvm/llvm-project/pull/193593.diff 4 Files Affected:
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 13fd0e2dc6255..54e783b0ee337 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -38,7 +38,7 @@ def int_dx_resource_handlefromimplicitbinding
def int_dx_resource_getpointer
: DefaultAttrsIntrinsic<[llvm_anyptr_ty], [llvm_any_ty, llvm_any_ty],
- [IntrConvergent, IntrNoMem]>;
+ [IntrReadMem, IntrInaccessibleMemOnly]>;
def int_dx_resource_nonuniformindex
: DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
diff --git a/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll b/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll
new file mode 100644
index 0000000000000..d560a2af4e3c5
--- /dev/null
+++ b/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll
@@ -0,0 +1,31 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=sink -S %s | FileCheck %s
+
+; Verify that dx.resource.getpointer can be sunk into a branch where it is
+; only used, now that it is no longer marked convergent.
+
+define void @can_sink_into_branch(i1 %cond) {
+; CHECK-LABEL: define void @can_sink_into_branch(
+; CHECK-SAME: i1 [[COND:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: br i1 [[COND]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
+; CHECK: [[IF_THEN]]:
+; CHECK-NEXT: [[BUF:%.*]] = call target("dx.RawBuffer", i32, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_i32_1_0t(i32 0, i32 0, i32 1, i32 0, ptr null)
+; CHECK-NEXT: [[PTR:%.*]] = call noundef nonnull align 4 dereferenceable(4) ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i32_1_0t.i32(target("dx.RawBuffer", i32, 1, 0) [[BUF]], i32 0)
+; CHECK-NEXT: store i32 42, ptr [[PTR]], align 4
+; CHECK-NEXT: br label %[[IF_END]]
+; CHECK: [[IF_END]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %buf = call target("dx.RawBuffer", i32, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_i32_1_0t(i32 0, i32 0, i32 1, i32 0, ptr null)
+ %ptr = call noundef nonnull align 4 dereferenceable(4) ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i32_1_0t.i32(target("dx.RawBuffer", i32, 1, 0) %buf, i32 0)
+ br i1 %cond, label %if.then, label %if.end
+
+if.then:
+ store i32 42, ptr %ptr, align 4
+ br label %if.end
+
+if.end:
+ ret void
+}
diff --git a/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll b/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
index 2d5a07562f4a4..eeaead53f893a 100644
--- a/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
+++ b/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
@@ -1,8 +1,9 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -passes=gvn -passes=instnamer -S %s | FileCheck %s
-; This test ensures that given dx.resource.getpointer is marked convergent, the
-; GVN pass is prevented from sinking these intrinsics.
+; This test ensures that given dx.resource.getpointer reads inaccessible memory,
+; the GVN pass is prevented from sinking these intrinsics out of branches which
+; would create phi nodes on the returned ptr.
;
; NOTE: The following ir represents case F and G from:
; https://godbolt.org/z/cK4xh1P49.
diff --git a/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll b/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
index 038b25d765d6f..576acce2b8dd2 100644
--- a/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
+++ b/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
@@ -1,9 +1,9 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -passes=simplifycfg -passes=instnamer -S %s | FileCheck %s
-; This test ensures that given dx.resource.getpointer is marked convergent, the
-; SimplifyCFG pass will be prevented from moving these intrinsics into the
-; branches required for sinking handle retrieve before resource access.
+; This test ensures that given dx.resource.getpointer reads inaccessible memory,
+; the SimplifyCFG pass will be prevented from sinking these intrinsics out of
+; branches which would create phi nodes on the returned ptr.
;
; NOTE: The following test ir is generated from:
; https://godbolt.org/z/1EdGTbscE.
|
Member
|
@llvm/pr-subscribers-llvm-transforms Author: Finn Plummer (inbelic) Changes
Using Updates relevant tests and adds a new test to demonstrate a now legal potential optimization. Assisted by: Claude Opus 4.6 Full diff: https://github.com/llvm/llvm-project/pull/193593.diff 4 Files Affected:
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 13fd0e2dc6255..54e783b0ee337 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -38,7 +38,7 @@ def int_dx_resource_handlefromimplicitbinding
def int_dx_resource_getpointer
: DefaultAttrsIntrinsic<[llvm_anyptr_ty], [llvm_any_ty, llvm_any_ty],
- [IntrConvergent, IntrNoMem]>;
+ [IntrReadMem, IntrInaccessibleMemOnly]>;
def int_dx_resource_nonuniformindex
: DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
diff --git a/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll b/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll
new file mode 100644
index 0000000000000..d560a2af4e3c5
--- /dev/null
+++ b/llvm/test/Transforms/DirectX/getpointer-sink-behavior.ll
@@ -0,0 +1,31 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=sink -S %s | FileCheck %s
+
+; Verify that dx.resource.getpointer can be sunk into a branch where it is
+; only used, now that it is no longer marked convergent.
+
+define void @can_sink_into_branch(i1 %cond) {
+; CHECK-LABEL: define void @can_sink_into_branch(
+; CHECK-SAME: i1 [[COND:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: br i1 [[COND]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
+; CHECK: [[IF_THEN]]:
+; CHECK-NEXT: [[BUF:%.*]] = call target("dx.RawBuffer", i32, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_i32_1_0t(i32 0, i32 0, i32 1, i32 0, ptr null)
+; CHECK-NEXT: [[PTR:%.*]] = call noundef nonnull align 4 dereferenceable(4) ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i32_1_0t.i32(target("dx.RawBuffer", i32, 1, 0) [[BUF]], i32 0)
+; CHECK-NEXT: store i32 42, ptr [[PTR]], align 4
+; CHECK-NEXT: br label %[[IF_END]]
+; CHECK: [[IF_END]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %buf = call target("dx.RawBuffer", i32, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_i32_1_0t(i32 0, i32 0, i32 1, i32 0, ptr null)
+ %ptr = call noundef nonnull align 4 dereferenceable(4) ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i32_1_0t.i32(target("dx.RawBuffer", i32, 1, 0) %buf, i32 0)
+ br i1 %cond, label %if.then, label %if.end
+
+if.then:
+ store i32 42, ptr %ptr, align 4
+ br label %if.end
+
+if.end:
+ ret void
+}
diff --git a/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll b/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
index 2d5a07562f4a4..eeaead53f893a 100644
--- a/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
+++ b/llvm/test/Transforms/GVN/no-sink-dxgetpointer.ll
@@ -1,8 +1,9 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -passes=gvn -passes=instnamer -S %s | FileCheck %s
-; This test ensures that given dx.resource.getpointer is marked convergent, the
-; GVN pass is prevented from sinking these intrinsics.
+; This test ensures that given dx.resource.getpointer reads inaccessible memory,
+; the GVN pass is prevented from sinking these intrinsics out of branches which
+; would create phi nodes on the returned ptr.
;
; NOTE: The following ir represents case F and G from:
; https://godbolt.org/z/cK4xh1P49.
diff --git a/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll b/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
index 038b25d765d6f..576acce2b8dd2 100644
--- a/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
+++ b/llvm/test/Transforms/SimplifyCFG/DirectX/no-sink-dxgetpointer.ll
@@ -1,9 +1,9 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -passes=simplifycfg -passes=instnamer -S %s | FileCheck %s
-; This test ensures that given dx.resource.getpointer is marked convergent, the
-; SimplifyCFG pass will be prevented from moving these intrinsics into the
-; branches required for sinking handle retrieve before resource access.
+; This test ensures that given dx.resource.getpointer reads inaccessible memory,
+; the SimplifyCFG pass will be prevented from sinking these intrinsics out of
+; branches which would create phi nodes on the returned ptr.
;
; NOTE: The following test ir is generated from:
; https://godbolt.org/z/1EdGTbscE.
|
bogner
approved these changes
Apr 23, 2026
Co-authored-by: Justin Bogner <mail@justinbogner.com>
yingopq
pushed a commit
to yingopq/llvm-project
that referenced
this pull request
Apr 29, 2026
…ly` and `IntrReadMem` (llvm#193593) `IntrConvergent` was originally added to `dx.resource.getpointer` to prevent optimization passes (`SimplifyCFG`, `GVN`) from sinking the intrinsic out of control flow branches, which would create phi nodes on the returned pointer. Using `IntrInaccessibleMemOnly` and `IntrReadMem` semantics still prevent passes from merging or sinking identical calls across branches. However, this allows the call to be moved within a single control flow path. Updates relevant tests and adds a new test to demonstrate a now legal potential optimization. This was discovered when llvm#188792 caused the following failure: https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618. When emitting convergence control tokens, each resource access is then a user of the convergence control tokens, which makes it's use more unnecessarily restrictive for optimizations and in this case would prevent a loop unroll from taking place. Assisted by: Claude Opus 4.6
KHicketts
pushed a commit
to KHicketts/llvm-project
that referenced
this pull request
Apr 30, 2026
…ly` and `IntrReadMem` (llvm#193593) `IntrConvergent` was originally added to `dx.resource.getpointer` to prevent optimization passes (`SimplifyCFG`, `GVN`) from sinking the intrinsic out of control flow branches, which would create phi nodes on the returned pointer. Using `IntrInaccessibleMemOnly` and `IntrReadMem` semantics still prevent passes from merging or sinking identical calls across branches. However, this allows the call to be moved within a single control flow path. Updates relevant tests and adds a new test to demonstrate a now legal potential optimization. This was discovered when llvm#188792 caused the following failure: https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618. When emitting convergence control tokens, each resource access is then a user of the convergence control tokens, which makes it's use more unnecessarily restrictive for optimizations and in this case would prevent a loop unroll from taking place. Assisted by: Claude Opus 4.6
inbelic
added a commit
that referenced
this pull request
May 26, 2026
…g DirectX" (#194452) The initial landing surfaced 3 somewhat orthogonal issues related to loop unrolling. These are addressed: [here](#193592), [here](#193593) and [here](#193590). These caused these [tests](https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618#step:8:87913) to fail in the offload test suite. We can verify that these are now passing as expected (fixing any of the 3 issues would resolve this and allow us to reland) Some additional tests were added since the revert that are now accounted for and updated in the reland fixes commit. This relands #188792
llvm-upstreamsync Bot
pushed a commit
to qualcomm/cpullvm-toolchain
that referenced
this pull request
May 26, 2026
…en targeting DirectX" (#194452) The initial landing surfaced 3 somewhat orthogonal issues related to loop unrolling. These are addressed: [here](llvm/llvm-project#193592), [here](llvm/llvm-project#193593) and [here](llvm/llvm-project#193590). These caused these [tests](https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618#step:8:87913) to fail in the offload test suite. We can verify that these are now passing as expected (fixing any of the 3 issues would resolve this and allow us to reland) Some additional tests were added since the revert that are now accounted for and updated in the reland fixes commit. This relands llvm/llvm-project#188792
llvm-sync Bot
pushed a commit
to arm/arm-toolchain
that referenced
this pull request
May 26, 2026
…en targeting DirectX" (#194452) The initial landing surfaced 3 somewhat orthogonal issues related to loop unrolling. These are addressed: [here](llvm/llvm-project#193592), [here](llvm/llvm-project#193593) and [here](llvm/llvm-project#193590). These caused these [tests](https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618#step:8:87913) to fail in the offload test suite. We can verify that these are now passing as expected (fixing any of the 3 issues would resolve this and allow us to reland) Some additional tests were added since the revert that are now accounted for and updated in the reland fixes commit. This relands llvm/llvm-project#188792
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
IntrConvergentwas originally added todx.resource.getpointerto prevent optimization passes (SimplifyCFG,GVN) from sinking the intrinsic out of control flow branches, which would create phi nodes on the returned pointer.Using
IntrInaccessibleMemOnlyandIntrReadMemsemantics still prevent passes from merging or sinking identical calls across branches. However, this allows the call to be moved within a single control flow path.Updates relevant tests and adds a new test to demonstrate a now legal potential optimization.
This was discovered when #188792 caused the following failure: https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618. When emitting convergence control tokens, each resource access is then a user of the convergence control tokens, which makes it's use more unnecessarily restrictive for optimizations and in this case would prevent a loop unroll from taking place.
Assisted by: Claude Opus 4.6