[ROCDL] Added missing named barrier ops (gfx1250) #162488

ravil-mobile · 2025-10-08T14:20:42Z

This patch introduces some missing s.barrier instructions in the ROCDL dialect handling named barriers

Specifically:

@llvm.amdgcn.s.barrier.init - s_barrier_init
@llvm.amdgcn.s.barrier.join - s_barrier_join
@llvm.amdgcn.s.barrier.leave - s_barrier_leave
@llvm.amdgcn.s.barrier.signal.isfirst - s_barrier_signal_isfirst
@llvm.amdgcn.s.get.barrier.state - s_get_barrier_state

Tests:

Added lit-tests to check MLIR -> LLVM lowering

cc @krzysz00, @amd-eochoalo

llvmbot · 2025-10-08T14:21:21Z

@llvm/pr-subscribers-mlir-llvm

@llvm/pr-subscribers-mlir

Author: None (ravil-mobile)

Changes

This patch introduces some missing s.barrier instructions in the ROCDL dialect handling named barriers

Specifically:

@<!-- -->llvm.amdgcn.s.barrier.init - s_barrier_init
@<!-- -->llvm.amdgcn.s.barrier.join - s_barrier_join
@<!-- -->llvm.amdgcn.s.barrier.leave - s_barrier_leave
@<!-- -->llvm.amdgcn.s.barrier.signal.isfirst - s_barrier_signal_isfirst

Tests:

Added lit-tests to check MLIR -> LLVM lowering

cc @krzysz00, @amd-eochoalo

Full diff: https://github.com/llvm/llvm-project/pull/162488.diff

3 Files Affected:

(modified) mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td (+32-1)
(modified) mlir/test/Dialect/LLVMIR/rocdl.mlir (+35)
(modified) mlir/test/Target/LLVMIR/rocdl.mlir (+35)

diff --git a/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td b/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
index db1b7e3af62fd..7560b61137afc 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
@@ -292,18 +292,50 @@ def ROCDL_BarrierOp : ROCDL_Op<"barrier"> {
   let assemblyFormat = "attr-dict";
 }
 
+def ROCDLBufferLDS : LLVM_PointerInAddressSpace<3>;
+
+def ROCDL_BarrierInitOp : ROCDL_IntrOp<"s.barrier.init", [], [], [Pure], 0, 0, 0, 0, [1], ["id"]>,
+  Arguments<(ins Arg<ROCDLBufferLDS, "", [MemRead]>:$ptr, I32Attr:$id)> {
+  let results = (outs);
+  let assemblyFormat = "$ptr `,` $id attr-dict";
+}
+
 def ROCDL_BarrierSignalOp : ROCDL_ConcreteNonMemIntrOp<"s.barrier.signal", [], 0, [0], ["id"]>,
   Arguments<(ins I32Attr:$id)> {
   let results = (outs);
   let assemblyFormat = "$id attr-dict";
 }
 
+def ROCDL_BarrierSignalVarOp : ROCDL_IntrOp<"s.barrier.signal.var", [], [], [Pure], 0, 0, 0, 0, [1], ["id"]>,
+  Arguments<(ins Arg<ROCDLBufferLDS, "", [MemRead]>:$ptr, I32Attr:$id)> {
+  let results = (outs);
+  let assemblyFormat = "$ptr `,` $id attr-dict";
+}
+
+def ROCDL_BarrierJoinOp : ROCDL_IntrOp<"s.barrier.join", [], [], [Pure], 0>,
+  Arguments<(ins Arg<ROCDLBufferLDS, "", [MemRead]>:$ptr)> {
+  let results = (outs);
+  let assemblyFormat = "$ptr attr-dict";
+}
+
+def ROCDL_BarrierLeaveOp : ROCDL_ConcreteNonMemIntrOp<"s.barrier.leave", [], 0, [0], ["id"]>,
+  Arguments<(ins I16Attr:$id)> {
+  let results = (outs);
+  let assemblyFormat = "$id attr-dict";
+}
+
 def ROCDL_BarrierWaitOp : ROCDL_ConcreteNonMemIntrOp<"s.barrier.wait", [], 0, [0], ["id"]>,
   Arguments<(ins I16Attr:$id)> {
   let results = (outs);
   let assemblyFormat = "$id attr-dict";
 }
 
+def ROCDL_BarrierSignalIsfirstOp : ROCDL_ConcreteNonMemIntrOp<"s.barrier.signal.isfirst", [], 1, [0], ["id"]>,
+  Arguments<(ins I32Attr:$id)> {
+  let results = (outs I1:$res);
+  let assemblyFormat = "$id attr-dict `:` type($res)";
+}
+
 def ROCDL_WaitDscntOp: ROCDL_ConcreteNonMemIntrOp<"s.wait.dscnt", [], 0, [0], ["count"]>,
   Arguments<(ins I16Attr:$count)> {
   let summary = "Wait until DSCNT is less than or equal to `count`";
@@ -497,7 +529,6 @@ def ROCDL_wmma_i32_16x16x32_iu4 : ROCDL_Wmma_IntrOp<"wmma.i32.16x16x32.iu4", [1]
 // LDS transpose intrinsics (available in GFX950)
 
 def ROCDLGlobalBuffer : LLVM_PointerInAddressSpace<1>;
-def ROCDLBufferLDS : LLVM_PointerInAddressSpace<3>;
 
 class ROCDL_LDS_Read_Tr_IntrOp<string mnemonic> :
   ROCDL_IntrOp<mnemonic, [1], [], [], 1, 0, 1> {
diff --git a/mlir/test/Dialect/LLVMIR/rocdl.mlir b/mlir/test/Dialect/LLVMIR/rocdl.mlir
index a88b59aeb61b2..9e13baf77a689 100644
--- a/mlir/test/Dialect/LLVMIR/rocdl.mlir
+++ b/mlir/test/Dialect/LLVMIR/rocdl.mlir
@@ -951,6 +951,13 @@ llvm.func @rocdl.s.barrier() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.init(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.init
+  // CHECK: rocdl.s.barrier.init %[[PTR:.+]], 1
+  rocdl.s.barrier.init %ptr, 1
+  llvm.return
+}
+
 llvm.func @rocdl.s.barrier.signal() {
   // CHECK-LABEL: rocdl.s.barrier.signal
   // CHECK: rocdl.s.barrier.signal -1
@@ -958,6 +965,27 @@ llvm.func @rocdl.s.barrier.signal() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.signal.var(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.signal.var
+  // CHECK: rocdl.s.barrier.signal.var %[[PTR:.+]], 1
+  rocdl.s.barrier.signal.var %ptr, 1
+  llvm.return
+}
+
+llvm.func @rocdl.s.barrier.join(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.join
+  // CHECK: rocdl.s.barrier.join %[[PTR:.+]]
+  rocdl.s.barrier.join %ptr
+  llvm.return
+}
+
+llvm.func @rocdl.s.barrier.leave() {
+  // CHECK-LABEL: rocdl.s.barrier.leave
+  // CHECK: rocdl.s.barrier.leave 1
+  rocdl.s.barrier.leave 1
+  llvm.return
+}
+
 llvm.func @rocdl.s.barrier.wait() {
   // CHECK-LABEL: rocdl.s.barrier.wait
   // CHECK: rocdl.s.barrier.wait -1
@@ -965,6 +993,13 @@ llvm.func @rocdl.s.barrier.wait() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.signal.isfirst() {
+  // CHECK-LABEL: rocdl.s.barrier.signal.isfirst
+  // CHECK: rocdl.s.barrier.signal.isfirst 1
+  %0 = rocdl.s.barrier.signal.isfirst 1 : i1
+  llvm.return
+}
+
 llvm.func @rocdl.s.wait.dscnt() {
   // CHECK-LABEL: rocdl.s.wait.dscnt
   // CHECK: rocdl.s.wait.dscnt 0
diff --git a/mlir/test/Target/LLVMIR/rocdl.mlir b/mlir/test/Target/LLVMIR/rocdl.mlir
index 1c0c2eba002aa..3eb404bcb8a40 100644
--- a/mlir/test/Target/LLVMIR/rocdl.mlir
+++ b/mlir/test/Target/LLVMIR/rocdl.mlir
@@ -192,6 +192,13 @@ llvm.func @rocdl.barrier() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.init(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.init
+  // CHECK: call void @llvm.amdgcn.s.barrier.init(ptr addrspace(3) %[[PTR:.+]], i32 1)
+  rocdl.s.barrier.init %ptr, 1
+  llvm.return
+}
+
 llvm.func @rocdl.s.barrier.signal() {
   // CHECK-LABEL: rocdl.s.barrier.signal
   // CHECK-NEXT: call void @llvm.amdgcn.s.barrier.signal(i32 -1)
@@ -199,6 +206,27 @@ llvm.func @rocdl.s.barrier.signal() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.signal.var(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.signal.var
+  // CHECK: call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(3) %[[PTR:.+]], i32 1)
+  rocdl.s.barrier.signal.var %ptr, 1
+  llvm.return
+}
+
+llvm.func @rocdl.s.barrier.join(%ptr : !llvm.ptr<3>) {
+  // CHECK-LABEL: rocdl.s.barrier.join
+  // CHECK: call void @llvm.amdgcn.s.barrier.join(ptr addrspace(3) %[[PTR:.+]])
+  rocdl.s.barrier.join %ptr
+  llvm.return
+}
+
+llvm.func @rocdl.s.barrier.leave() {
+  // CHECK-LABEL: rocdl.s.barrier.leave
+  // CHECK: call void @llvm.amdgcn.s.barrier.leave(i16 1)
+  rocdl.s.barrier.leave 1
+  llvm.return
+}
+
 llvm.func @rocdl.s.barrier.wait() {
   // CHECK-LABEL: rocdl.s.barrier.wait
   // CHECK-NEXT: call void @llvm.amdgcn.s.barrier.wait(i16 -1)
@@ -206,6 +234,13 @@ llvm.func @rocdl.s.barrier.wait() {
   llvm.return
 }
 
+llvm.func @rocdl.s.barrier.signal.isfirst() {
+  // CHECK-LABEL: rocdl.s.barrier.signal.isfirst
+  // CHECK:  %[[OUT:.+]] = call i1 @llvm.amdgcn.s.barrier.signal.isfirst(i32 1)
+  %0 = rocdl.s.barrier.signal.isfirst 1 : i1
+  llvm.return
+}
+
 llvm.func @rocdl.s.wait.dscnt() {
   // CHECK-LABEL: rocdl.s.wait.dscnt
   // CHECK-NEXT: call void @llvm.amdgcn.s.wait.dscnt(i16 0)

kuhar · 2025-10-08T17:39:40Z

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

@@ -292,18 +292,56 @@ def ROCDL_BarrierOp : ROCDL_Op<"barrier"> {
  let assemblyFormat = "attr-dict";
 }

+def ROCDLBufferLDS : LLVM_PointerInAddressSpace<3>;
+
+def ROCDL_BarrierInitOp : ROCDL_IntrOp<"s.barrier.init", [], [], [Pure], 0, 0, 0, 0, [1], ["id"]>,


Is this pure? I don't see any results returned by this op

Yeah, I don't think this is pure

... Heck, the LLVM has IntrHasSideEffects

yeah should not be pure. this op does not have a return value so it will get eliminated if it is pure.

Agree. Removed the trait. As I can see there is not HasSideEffect trait in MLIR. So, I left the list of traits empty

kuhar

Can you add descriptions saying which gfxip version introduced these? Example: https://github.com/ravil-mobile/llvm-project/blob/5135c4fdae71ef27192ec190298f3a00ba85ae64/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td#L1043-L1045

krzysz00 · 2025-10-08T18:23:38Z

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

+}
+
+def ROCDL_BarrierJoinOp : ROCDL_IntrOp<"s.barrier.join", [], [], [Pure], 0>,
+  Arguments<(ins Arg<ROCDLBufferLDS, "", [MemRead]>:$ptr)> {


I don't think it's necessary to mark this as MemRed - the LLVM intrinsics are marked NoMem.

Yep. You are right. All this functions are marked with IntrNoMem trait in LLVM. Removed it

lialan · 2025-10-09T00:53:42Z

All LLVM intrinsics added in this op have property IntrHasSideEffects.

lialan · 2025-10-09T00:59:45Z

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

+def ROCDLBufferLDS : LLVM_PointerInAddressSpace<3>;
+
+def ROCDL_BarrierInitOp : ROCDL_IntrOp<"s.barrier.init", [], [], [Pure], 0, 0, 0, 0, [1], ["id"]>,
+  Arguments<(ins Arg<ROCDLBufferLDS, "", [MemRead]>:$ptr, I32Attr:$id)> {


See: https://github.com/llvm/llvm-project/blob/main/llvm/test/CodeGen/AMDGPU/s-barrier.ll#L83
so the s_barrier_init takes a single m0 so it should not read mem.

Same is signal.var, join.

lialan · 2025-10-09T01:04:53Z

The discrepancy here being: barriers need to be defined as a global variable at the LLVM IR level.
However those intrinsics do not rely on actually reading a global address in the emitted assembly (can see in the lit tests).

ravil-mobile · 2025-10-09T09:55:16Z

Can you add descriptions saying which gfxip version introduced these? Example: https://github.com/ravil-mobile/llvm-project/blob/5135c4fdae71ef27192ec190298f3a00ba85ae64/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td#L1043-L1045

Done

lialan · 2025-10-09T17:46:21Z

just want to make sure: do you know if on the rocdl level, is there any way we can declare a amdgcn.named.barrier?

krzysz00

Overall lgtm

krzysz00 · 2025-10-09T21:09:26Z

@lialan To answer your question, the declarations I see in tests for these named barriers are

@bar3 = internal addrspace(3) global target("amdgcn.named.barrier", 0) poison`

So, at the MLIR level, you just need a poison global value, in LDS, of type
!llvm.target<"amdgcn.named.barrier", 0>

(No, I don't know what the 0 is for)

So future work will be adding amdgpu.named_barrier @symbol

lialan · 2025-10-09T21:35:25Z

So, at the MLIR level, you just need a poison global value, in LDS, of type !llvm.target<"amdgcn.named.barrier", 0>

@krzysz00 Yeah I do not see the wrapper for type amdgcn.named.barrier in rocdl dialect. Or actually I do not need this?

krzysz00 · 2025-10-09T21:36:17Z

@lialan You don't need such a wrapper, just use llvm.target types.

(We can add a wrapper to AMDGPU, but we don't need one in rocdl)

llvmbot added mlir:llvm mlir labels Oct 8, 2025

ravil-mobile force-pushed the ravil/rocdl-barriers branch from 5ea786d to 5135c4f Compare October 8, 2025 15:05

searlmc1 requested review from jayfoad, krzysz00 and kuhar October 8, 2025 16:23

kuhar reviewed Oct 8, 2025

View reviewed changes

krzysz00 reviewed Oct 8, 2025

View reviewed changes

lialan reviewed Oct 9, 2025

View reviewed changes

[ROCDL] Added missing named barrier ops (gfx1250)

41c84bc

ravil-mobile force-pushed the ravil/rocdl-barriers branch from 5135c4f to 41c84bc Compare October 9, 2025 09:12

ravil-mobile requested review from kuhar, krzysz00 and lialan October 9, 2025 09:57

krzysz00 approved these changes Oct 9, 2025

View reviewed changes

kuhar approved these changes Oct 10, 2025

View reviewed changes

kuhar merged commit 3c080ed into llvm:main Oct 10, 2025
9 checks passed

[ROCDL] Added missing named barrier ops (gfx1250) #162488

[ROCDL] Added missing named barrier ops (gfx1250) #162488

Conversation

ravil-mobile commented Oct 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Oct 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

kuhar left a comment

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!

lialan commented Oct 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lialan commented Oct 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ravil-mobile commented Oct 9, 2025

Uh oh!

lialan commented Oct 9, 2025

Uh oh!

krzysz00 left a comment

Choose a reason for hiding this comment

Uh oh!

krzysz00 commented Oct 9, 2025

Uh oh!

lialan commented Oct 9, 2025

Uh oh!

krzysz00 commented Oct 9, 2025

Uh oh!

Uh oh!

Uh oh!

ravil-mobile commented Oct 8, 2025 •

edited

Loading

llvmbot commented Oct 8, 2025 •

edited

Loading

lialan commented Oct 9, 2025 •

edited

Loading

lialan commented Oct 9, 2025 •

edited

Loading