[llvm] [AMDGPU] Expand constant named barrier matching in SDag (PR #191506)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Apr 10 12:58:54 PDT 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Krzysztof Drewniak (krzysz00)
<details>
<summary>Changes</summary>
This commit factors out some utilities used to enable M0-less s.signal.barrier.var on gfx1250 to use them for other intrinsics that take named barrier objects, such as s.barrier.joint, eliminating a bunch of needless m0 writes when the barrier doesn't change.
In addition, this commit specifies that the "NULL barrier object" is represented by `ptr addrspace(3) nullptr`, enabling `s_barrier_join 0` to be emitted from user code, and also adds tests for other uses of the null barrier (which are generally no-ops).
---
Full diff: https://github.com/llvm/llvm-project/pull/191506.diff
4 Files Affected:
- (modified) llvm/docs/AMDGPUUsage.rst (+3)
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+29-23)
- (modified) llvm/test/CodeGen/AMDGPU/s-barrier.ll (+87-58)
- (modified) llvm/test/CodeGen/AMDGPU/s-wakeup-barrier.ll (+19-2)
``````````diff
diff --git a/llvm/docs/AMDGPUUsage.rst b/llvm/docs/AMDGPUUsage.rst
index fb440bde68b8f..d9079f21fe544 100644
--- a/llvm/docs/AMDGPUUsage.rst
+++ b/llvm/docs/AMDGPUUsage.rst
@@ -1307,6 +1307,9 @@ as the pointer operand of a regular memory access instruction or intrinsic.
Pointers to named barrier objects are intended to be used with dedicated
intrinsics. Reading from or writing to such pointers is undefined behavior.
+The NULL named barrier (needed for un-joining a barrier) is represented in IR by
+``ptr addrspace(3) nullptr``.
+
LLVM IR Intrinsics
------------------
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 6750dfcbaac62..32398c7dac7ce 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -10909,6 +10909,21 @@ SITargetLowering::lowerStructBufferAtomicIntrin(SDValue Op, SelectionDAG &DAG,
M->getMemOperand());
}
+// Try to resolve a named barrier object to a constant barrier ID.
+// Named barrier IDs are encoded at bits [9:4] of a "LDS address".
+static std::optional<uint32_t> getConstantNamedBarrierNumber(SDValue BarOp) {
+ std::optional<uint64_t> BarVal;
+ if (auto *C = dyn_cast<ConstantSDNode>(BarOp))
+ BarVal = C->getZExtValue();
+ else if (auto *GA = dyn_cast<GlobalAddressSDNode>(BarOp))
+ if (auto Addr =
+ AMDGPUMachineFunctionInfo::getLDSAbsoluteAddress(*GA->getGlobal()))
+ BarVal = *Addr + GA->getOffset();
+ if (BarVal)
+ return (*BarVal >> 4) & 0x3F;
+ return std::nullopt;
+}
+
SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
SelectionDAG &DAG) const {
unsigned IntrID = Op.getConstantOperandVal(1);
@@ -11466,12 +11481,15 @@ SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
SmallVector<SDValue, 2> Ops;
unsigned Opc;
- if (isa<ConstantSDNode>(Op->getOperand(2))) {
- uint64_t BarID = cast<ConstantSDNode>(Op->getOperand(2))->getZExtValue();
- if (IntrID == Intrinsic::amdgcn_s_get_named_barrier_state)
- BarID = (BarID >> 4) & 0x3F;
+ std::optional<uint32_t> BarID;
+ if (IntrID == Intrinsic::amdgcn_s_get_named_barrier_state)
+ BarID = getConstantNamedBarrierNumber(Op->getOperand(2));
+ else if (auto *C = dyn_cast<ConstantSDNode>(Op->getOperand(2)))
+ BarID = C->getZExtValue();
+
+ if (BarID) {
Opc = AMDGPU::S_GET_BARRIER_STATE_IMM;
- SDValue K = DAG.getTargetConstant(BarID, DL, MVT::i32);
+ SDValue K = DAG.getTargetConstant(*BarID, DL, MVT::i32);
Ops.push_back(K);
Ops.push_back(Chain);
} else {
@@ -12066,19 +12084,10 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
if (CntC && CntC->isZero()) {
SDValue Chain = Op->getOperand(0);
SDValue BarOp = Op->getOperand(2);
- SmallVector<SDValue, 2> Ops;
-
- std::optional<uint64_t> BarVal;
- if (auto *C = dyn_cast<ConstantSDNode>(BarOp))
- BarVal = C->getZExtValue();
- else if (auto *GA = dyn_cast<GlobalAddressSDNode>(BarOp))
- if (auto Addr = AMDGPUMachineFunctionInfo::getLDSAbsoluteAddress(
- *GA->getGlobal()))
- BarVal = *Addr + GA->getOffset();
-
- if (BarVal) {
- unsigned BarID = (*BarVal >> 4) & 0x3F;
- Ops.push_back(DAG.getTargetConstant(BarID, DL, MVT::i32));
+
+ if (auto BarID = getConstantNamedBarrierNumber(BarOp)) {
+ SmallVector<SDValue, 2> Ops;
+ Ops.push_back(DAG.getTargetConstant(*BarID, DL, MVT::i32));
Ops.push_back(Chain);
auto *NewMI = DAG.getMachineNode(AMDGPU::S_BARRIER_SIGNAL_IMM, DL,
Op->getVTList(), Ops);
@@ -12135,8 +12144,7 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
SDValue BarOp = Op->getOperand(2);
unsigned Opc;
- if (isa<ConstantSDNode>(BarOp)) {
- uint64_t BarVal = cast<ConstantSDNode>(BarOp)->getZExtValue();
+ if (auto BarID = getConstantNamedBarrierNumber(BarOp)) {
switch (IntrinsicID) {
default:
return SDValue();
@@ -12147,9 +12155,7 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
Opc = AMDGPU::S_WAKEUP_BARRIER_IMM;
break;
}
- // extract the BarrierID from bits 4-9 of the immediate
- unsigned BarID = (BarVal >> 4) & 0x3F;
- SDValue K = DAG.getTargetConstant(BarID, DL, MVT::i32);
+ SDValue K = DAG.getTargetConstant(*BarID, DL, MVT::i32);
Ops.push_back(K);
Ops.push_back(Chain);
} else {
diff --git a/llvm/test/CodeGen/AMDGPU/s-barrier.ll b/llvm/test/CodeGen/AMDGPU/s-barrier.ll
index 6c438ed94c863..a0df95715f782 100644
--- a/llvm/test/CodeGen/AMDGPU/s-barrier.ll
+++ b/llvm/test/CodeGen/AMDGPU/s-barrier.ll
@@ -7,32 +7,18 @@
@bar3 = internal addrspace(3) global target("amdgcn.named.barrier", 0) poison
define void @func1() {
-; GFX12-SDAG-LABEL: func1:
-; GFX12-SDAG: ; %bb.0:
-; GFX12-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-SDAG-NEXT: s_wait_expcnt 0x0
-; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
-; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
-; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 3
-; GFX12-SDAG-NEXT: s_barrier_join m0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 0x70003
-; GFX12-SDAG-NEXT: s_barrier_signal m0
-; GFX12-SDAG-NEXT: s_barrier_wait 1
-; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-GISEL-LABEL: func1:
-; GFX12-GISEL: ; %bb.0:
-; GFX12-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-GISEL-NEXT: s_wait_expcnt 0x0
-; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0
-; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0
-; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0
-; GFX12-GISEL-NEXT: s_mov_b32 m0, 0x70003
-; GFX12-GISEL-NEXT: s_barrier_join 3
-; GFX12-GISEL-NEXT: s_barrier_signal m0
-; GFX12-GISEL-NEXT: s_barrier_wait 1
-; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31]
+; GFX12-LABEL: func1:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_mov_b32 m0, 0x70003
+; GFX12-NEXT: s_barrier_join 3
+; GFX12-NEXT: s_barrier_signal m0
+; GFX12-NEXT: s_barrier_wait 1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
call void @llvm.amdgcn.s.barrier.join(ptr addrspace(3) @bar3)
call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(3) @bar3, i32 7)
call void @llvm.amdgcn.s.barrier.wait(i16 1)
@@ -40,32 +26,18 @@ define void @func1() {
}
define void @func2() {
-; GFX12-SDAG-LABEL: func2:
-; GFX12-SDAG: ; %bb.0:
-; GFX12-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-SDAG-NEXT: s_wait_expcnt 0x0
-; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
-; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
-; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 1
-; GFX12-SDAG-NEXT: s_barrier_join m0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 0x70001
-; GFX12-SDAG-NEXT: s_barrier_signal m0
-; GFX12-SDAG-NEXT: s_barrier_wait 1
-; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-GISEL-LABEL: func2:
-; GFX12-GISEL: ; %bb.0:
-; GFX12-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-GISEL-NEXT: s_wait_expcnt 0x0
-; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0
-; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0
-; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0
-; GFX12-GISEL-NEXT: s_mov_b32 m0, 0x70001
-; GFX12-GISEL-NEXT: s_barrier_join 1
-; GFX12-GISEL-NEXT: s_barrier_signal m0
-; GFX12-GISEL-NEXT: s_barrier_wait 1
-; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31]
+; GFX12-LABEL: func2:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_mov_b32 m0, 0x70001
+; GFX12-NEXT: s_barrier_join 1
+; GFX12-NEXT: s_barrier_signal m0
+; GFX12-NEXT: s_barrier_wait 1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
call void @llvm.amdgcn.s.barrier.join(ptr addrspace(3) @bar2)
call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(3) @bar2, i32 7)
call void @llvm.amdgcn.s.barrier.wait(i16 1)
@@ -100,12 +72,11 @@ define amdgpu_kernel void @kernel1(ptr addrspace(1) %out, ptr addrspace(3) %in)
; GFX12-SDAG-NEXT: s_barrier_signal -1
; GFX12-SDAG-NEXT: s_barrier_join m0
; GFX12-SDAG-NEXT: s_barrier_signal_isfirst -1
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 2
; GFX12-SDAG-NEXT: s_barrier_wait 1
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT: s_barrier_leave
-; GFX12-SDAG-NEXT: s_get_barrier_state s3, m0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, s2
+; GFX12-SDAG-NEXT: s_get_barrier_state s2, 2
+; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT: s_get_barrier_state s2, m0
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT: s_getpc_b64 s[2:3]
@@ -218,12 +189,11 @@ define amdgpu_kernel void @kernel2(ptr addrspace(1) %out, ptr addrspace(3) %in)
; GFX12-SDAG-NEXT: s_load_b64 s[12:13], s[6:7], 0x0
; GFX12-SDAG-NEXT: s_mov_b32 m0, 0x70002
; GFX12-SDAG-NEXT: s_add_nc_u64 s[8:9], s[4:5], 48
-; GFX12-SDAG-NEXT: s_barrier_signal m0
-; GFX12-SDAG-NEXT: s_mov_b32 m0, 2
; GFX12-SDAG-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX12-SDAG-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX12-SDAG-NEXT: s_mov_b32 s32, 0
-; GFX12-SDAG-NEXT: s_barrier_join m0
+; GFX12-SDAG-NEXT: s_barrier_signal m0
+; GFX12-SDAG-NEXT: s_barrier_join 2
; GFX12-SDAG-NEXT: s_barrier_wait 1
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT: s_swappc_b64 s[30:31], s[12:13]
@@ -308,6 +278,65 @@ define amdgpu_ps void @test_barrier_leave_write_to_scc(i32 inreg %val, ptr addrs
ret void
}
+define void @signal_var_cnt0_null_bar() {
+; GFX12-LABEL: signal_var_cnt0_null_bar:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_barrier_signal 0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+ call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(3) null, i32 0)
+ ret void
+}
+
+define void @barrier_join_null_bar() {
+; GFX12-LABEL: barrier_join_null_bar:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_barrier_join 0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+ call void @llvm.amdgcn.s.barrier.join(ptr addrspace(3) null)
+ ret void
+}
+
+define void @barrier_init_null_bar() {
+; GFX12-LABEL: barrier_init_null_bar:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_mov_b32 m0, 0x30000
+; GFX12-NEXT: s_barrier_init m0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+ call void @llvm.amdgcn.s.barrier.init(ptr addrspace(3) null, i32 3)
+ ret void
+}
+
+define void @get_named_barrier_state_null_bar() {
+; GFX12-LABEL: get_named_barrier_state_null_bar:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_get_barrier_state s0, 0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_alu depctr_sa_sdst(0)
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+ %state = call i32 @llvm.amdgcn.s.get.named.barrier.state(ptr addrspace(3) null)
+ ret void
+}
+
declare void @llvm.amdgcn.s.barrier() #1
declare void @llvm.amdgcn.s.barrier.wait(i16) #1
declare void @llvm.amdgcn.s.barrier.signal(i32) #1
diff --git a/llvm/test/CodeGen/AMDGPU/s-wakeup-barrier.ll b/llvm/test/CodeGen/AMDGPU/s-wakeup-barrier.ll
index bc995f4738af8..996ef0691e143 100644
--- a/llvm/test/CodeGen/AMDGPU/s-wakeup-barrier.ll
+++ b/llvm/test/CodeGen/AMDGPU/s-wakeup-barrier.ll
@@ -9,8 +9,7 @@ define amdgpu_kernel void @kernel1(ptr addrspace(1) %out, ptr addrspace(3) %in)
; GFX1250-SDAG: ; %bb.0:
; GFX1250-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x2c nv
-; GFX1250-SDAG-NEXT: s_mov_b32 m0, 1
-; GFX1250-SDAG-NEXT: s_wakeup_barrier m0
+; GFX1250-SDAG-NEXT: s_wakeup_barrier 1
; GFX1250-SDAG-NEXT: s_wait_kmcnt 0x0
; GFX1250-SDAG-NEXT: s_lshr_b32 s0, s0, 4
; GFX1250-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -35,6 +34,24 @@ define amdgpu_kernel void @kernel1(ptr addrspace(1) %out, ptr addrspace(3) %in)
}
+define void @wakeup_barrier_null_bar() {
+; GFX1250-SDAG-LABEL: wakeup_barrier_null_bar:
+; GFX1250-SDAG: ; %bb.0:
+; GFX1250-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-SDAG-NEXT: s_wait_kmcnt 0x0
+; GFX1250-SDAG-NEXT: s_wakeup_barrier 0
+; GFX1250-SDAG-NEXT: s_set_pc_i64 s[30:31]
+;
+; GFX1250-GISEL-LABEL: wakeup_barrier_null_bar:
+; GFX1250-GISEL: ; %bb.0:
+; GFX1250-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-GISEL-NEXT: s_wait_kmcnt 0x0
+; GFX1250-GISEL-NEXT: s_wakeup_barrier 0
+; GFX1250-GISEL-NEXT: s_set_pc_i64 s[30:31]
+ call void @llvm.amdgcn.s.wakeup.barrier(ptr addrspace(3) null)
+ ret void
+}
+
declare void @llvm.amdgcn.s.wakeup.barrier(ptr addrspace(3)) #1
attributes #0 = { nounwind }
``````````
</details>
https://github.com/llvm/llvm-project/pull/191506
More information about the llvm-commits
mailing list