[llvm] [AMDGPU] Use 8-bit barrier member count on GFX13 (PR #230884)

via llvm-commits llvm-commits at lists.llvm.org
Sat Oct 10 21:03:07 PDT 2026


llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Chinmay Deshpande (chinmaydd)

<details>
<summary>Changes</summary>

GFX13 widens the barrier member count in M0 to 8 bits for cluster named barriers. `s_barrier_init` and `s_barrier_signal_var` lowering masked it to 6 bits, truncating counts above 63.

---
Full diff: https://github.com/llvm/llvm-project/pull/230884.diff


5 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp (+4-2) 
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+4-2) 
- (modified) llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp (+9) 
- (modified) llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h (+3) 
- (added) llvm/test/CodeGen/AMDGPU/s-barrier-member-count.ll (+137) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 39c8f65f4f9c7..f05c726b71504 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -7445,11 +7445,13 @@ bool AMDGPUInstructionSelector::selectNamedBarrierInit(
       .addImm(0x3F)
       .setOperandDead(3); // Dead scc
 
-  // MO = ((CntOp & 0x3F) << shAmt) | BarID
+  // MO = ((CntOp & CntMask) << shAmt) | BarID
+  unsigned CntMask =
+      AMDGPU::getBarrierMemberCountBitMask(AMDGPU::getIsaVersion(STI.getCPU()));
   Register TmpReg2 = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
   BuildMI(*MBB, &I, DL, TII.get(AMDGPU::S_AND_B32), TmpReg2)
       .add(CntOp)
-      .addImm(0x3F)
+      .addImm(CntMask)
       .setOperandDead(3); // Dead scc
 
   Register TmpReg3 = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 7be36bf8f9174..fa7c49528a820 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -13103,10 +13103,12 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
     // extract the BarrierID from bits 0-5 of BarOp
     SDValue BarID = DAG.getNode(ISD::AND, DL, MVT::i32, BarOp,
                                 DAG.getConstant(0x3F, DL, MVT::i32));
-    // Member count should be put into M0[ShAmt:+6]
+    // Member count should be put into M0[ShAmt:+CntMask width]
     // Barrier ID should be put into M0[5:0]
+    unsigned CntMask = AMDGPU::getBarrierMemberCountBitMask(
+        AMDGPU::getIsaVersion(Subtarget->getCPU()));
     SDValue MemberCnt = DAG.getNode(ISD::AND, DL, MVT::i32, CntOp,
-                                    DAG.getConstant(0x3F, DL, MVT::i32));
+                                    DAG.getConstant(CntMask, DL, MVT::i32));
     constexpr unsigned ShAmt = 16;
     M0Val = DAG.getNode(ISD::SHL, DL, MVT::i32, MemberCnt,
                         DAG.getShiftAmountConstant(ShAmt, MVT::i32, DL));
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index bca4a408f5ceb..4a464dc3c3666 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -145,6 +145,11 @@ unsigned getLoadcntStorecntBitShift(unsigned VersionMajor) {
   return VersionMajor >= 12 ? 8 : 0;
 }
 
+/// \returns barrier member count bit width in M0.
+unsigned getBarrierMemberCountBitWidth(unsigned VersionMajor) {
+  return VersionMajor >= 13 ? 8 : 6;
+}
+
 /// \returns VaSdst bit width
 inline unsigned getVaSdstBitWidth() { return 3; }
 
@@ -1688,6 +1693,10 @@ unsigned getStorecntBitMask(const IsaVersion &Version) {
   return (1 << getStorecntBitWidth(Version.Major)) - 1;
 }
 
+unsigned getBarrierMemberCountBitMask(const IsaVersion &Version) {
+  return (1 << getBarrierMemberCountBitWidth(Version.Major)) - 1;
+}
+
 unsigned getWaitcntBitMask(const IsaVersion &Version) {
   unsigned VmcntLo = getBitMask(getVmcntBitShiftLo(Version.Major),
                                 getVmcntBitWidthLo(Version.Major));
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 50ee28931eda4..dfee7bbdbf251 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1178,6 +1178,9 @@ unsigned getXcntBitMask(const IsaVersion &Version);
 /// depends on the ISA version.
 unsigned getStorecntBitMask(const IsaVersion &Version);
 
+/// \returns barrier member count bit mask in M0 for given isa \p Version.
+unsigned getBarrierMemberCountBitMask(const IsaVersion &Version);
+
 namespace Hwreg {
 
 using HwregId = EncodingField<5, 0>;
diff --git a/llvm/test/CodeGen/AMDGPU/s-barrier-member-count.ll b/llvm/test/CodeGen/AMDGPU/s-barrier-member-count.ll
new file mode 100644
index 0000000000000..bee55193d21d8
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/s-barrier-member-count.ll
@@ -0,0 +1,137 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -global-isel=0 -mtriple=amdgpu12.00 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG %s
+; RUN: llc -global-isel=1 -mtriple=amdgpu12.00 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL %s
+; RUN: llc -global-isel=0 -mtriple=amdgpu13.10 < %s | FileCheck -check-prefixes=GFX13,GFX13-SDAG %s
+; RUN: llc -global-isel=1 -mtriple=amdgpu13.10 < %s | FileCheck -check-prefixes=GFX13,GFX13-GISEL %s
+
+; Member count is 6 bits on GFX12 and 8 bits on GFX13.
+
+ at bar = internal addrspace(15) global target("amdgcn.named.barrier", 0) poison
+
+define amdgpu_kernel void @barrier_init_var_cnt(i32 inreg %cnt) {
+; GFX12-SDAG-LABEL: barrier_init_var_cnt:
+; GFX12-SDAG:       ; %bb.0:
+; GFX12-SDAG-NEXT:    s_load_b32 s0, s[4:5], 0x24
+; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
+; GFX12-SDAG-NEXT:    s_and_b32 s0, s0, 63
+; GFX12-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-SDAG-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX12-SDAG-NEXT:    s_or_b32 m0, s0, 1
+; GFX12-SDAG-NEXT:    s_barrier_init m0
+; GFX12-SDAG-NEXT:    s_endpgm
+;
+; GFX12-GISEL-LABEL: barrier_init_var_cnt:
+; GFX12-GISEL:       ; %bb.0:
+; GFX12-GISEL-NEXT:    s_load_b32 s0, s[4:5], 0x24
+; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
+; GFX12-GISEL-NEXT:    s_and_b32 s0, s0, 63
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-GISEL-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX12-GISEL-NEXT:    s_or_b32 m0, 1, s0
+; GFX12-GISEL-NEXT:    s_barrier_init m0
+; GFX12-GISEL-NEXT:    s_endpgm
+;
+; GFX13-SDAG-LABEL: barrier_init_var_cnt:
+; GFX13-SDAG:       ; %bb.0:
+; GFX13-SDAG-NEXT:    s_load_b32 s0, s[4:5], 0x24 nv
+; GFX13-SDAG-NEXT:    s_wait_kmcnt 0x0
+; GFX13-SDAG-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX13-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX13-SDAG-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX13-SDAG-NEXT:    s_or_b32 m0, s0, 1
+; GFX13-SDAG-NEXT:    s_barrier_init m0
+; GFX13-SDAG-NEXT:    s_endpgm
+;
+; GFX13-GISEL-LABEL: barrier_init_var_cnt:
+; GFX13-GISEL:       ; %bb.0:
+; GFX13-GISEL-NEXT:    s_load_b32 s0, s[4:5], 0x24 nv
+; GFX13-GISEL-NEXT:    s_wait_kmcnt 0x0
+; GFX13-GISEL-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX13-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX13-GISEL-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX13-GISEL-NEXT:    s_or_b32 m0, 1, s0
+; GFX13-GISEL-NEXT:    s_barrier_init m0
+; GFX13-GISEL-NEXT:    s_endpgm
+  call void @llvm.amdgcn.s.barrier.init(ptr addrspace(15) @bar, i32 %cnt)
+  ret void
+}
+
+define amdgpu_kernel void @barrier_init_cnt_255() {
+; GFX12-LABEL: barrier_init_cnt_255:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_mov_b32 m0, 0x3f0001
+; GFX12-NEXT:    s_barrier_init m0
+; GFX12-NEXT:    s_endpgm
+;
+; GFX13-LABEL: barrier_init_cnt_255:
+; GFX13:       ; %bb.0:
+; GFX13-NEXT:    s_mov_b32 m0, 0xff0001
+; GFX13-NEXT:    s_barrier_init m0
+; GFX13-NEXT:    s_endpgm
+  call void @llvm.amdgcn.s.barrier.init(ptr addrspace(15) @bar, i32 255)
+  ret void
+}
+
+define amdgpu_kernel void @barrier_signal_var_var_cnt(i32 inreg %cnt) {
+; GFX12-SDAG-LABEL: barrier_signal_var_var_cnt:
+; GFX12-SDAG:       ; %bb.0:
+; GFX12-SDAG-NEXT:    s_load_b32 s0, s[4:5], 0x24
+; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
+; GFX12-SDAG-NEXT:    s_and_b32 s0, s0, 63
+; GFX12-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-SDAG-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX12-SDAG-NEXT:    s_or_b32 m0, s0, 1
+; GFX12-SDAG-NEXT:    s_barrier_signal m0
+; GFX12-SDAG-NEXT:    s_endpgm
+;
+; GFX12-GISEL-LABEL: barrier_signal_var_var_cnt:
+; GFX12-GISEL:       ; %bb.0:
+; GFX12-GISEL-NEXT:    s_load_b32 s0, s[4:5], 0x24
+; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
+; GFX12-GISEL-NEXT:    s_and_b32 s0, s0, 63
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-GISEL-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX12-GISEL-NEXT:    s_or_b32 m0, 1, s0
+; GFX12-GISEL-NEXT:    s_barrier_signal m0
+; GFX12-GISEL-NEXT:    s_endpgm
+;
+; GFX13-SDAG-LABEL: barrier_signal_var_var_cnt:
+; GFX13-SDAG:       ; %bb.0:
+; GFX13-SDAG-NEXT:    s_load_b32 s0, s[4:5], 0x24 nv
+; GFX13-SDAG-NEXT:    s_wait_kmcnt 0x0
+; GFX13-SDAG-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX13-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX13-SDAG-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX13-SDAG-NEXT:    s_or_b32 m0, s0, 1
+; GFX13-SDAG-NEXT:    s_barrier_signal m0
+; GFX13-SDAG-NEXT:    s_endpgm
+;
+; GFX13-GISEL-LABEL: barrier_signal_var_var_cnt:
+; GFX13-GISEL:       ; %bb.0:
+; GFX13-GISEL-NEXT:    s_load_b32 s0, s[4:5], 0x24 nv
+; GFX13-GISEL-NEXT:    s_wait_kmcnt 0x0
+; GFX13-GISEL-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX13-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX13-GISEL-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX13-GISEL-NEXT:    s_or_b32 m0, 1, s0
+; GFX13-GISEL-NEXT:    s_barrier_signal m0
+; GFX13-GISEL-NEXT:    s_endpgm
+  call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(15) @bar, i32 %cnt)
+  ret void
+}
+
+define amdgpu_kernel void @barrier_signal_var_cnt_255() {
+; GFX12-LABEL: barrier_signal_var_cnt_255:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_mov_b32 m0, 0x3f0001
+; GFX12-NEXT:    s_barrier_signal m0
+; GFX12-NEXT:    s_endpgm
+;
+; GFX13-LABEL: barrier_signal_var_cnt_255:
+; GFX13:       ; %bb.0:
+; GFX13-NEXT:    s_mov_b32 m0, 0xff0001
+; GFX13-NEXT:    s_barrier_signal m0
+; GFX13-NEXT:    s_endpgm
+  call void @llvm.amdgcn.s.barrier.signal.var(ptr addrspace(15) @bar, i32 255)
+  ret void
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/230884


More information about the llvm-commits mailing list