[llvm] [AMDGPU] A SCHED_BARRIER in a bundle should not prevent other SCHED_BARRIERs to be considered (PR #152627)
Yoonseo Choi via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 8 15:33:35 PDT 2025
https://github.com/yoonseoch updated https://github.com/llvm/llvm-project/pull/152627
>From f06e57fe0e160e31db35e68d7ea320e5bc68e206 Mon Sep 17 00:00:00 2001
From: Yoonseo Choi <yoonchoi at amd.com>
Date: Thu, 7 Aug 2025 19:08:57 -0500
Subject: [PATCH 1/2] A SCHED_BARRIER in a bundle should not prevent other
SCHED_BARRIER to be considered
---
llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp | 4 +-
.../CodeGen/AMDGPU/sched-barrier-post-RA.mir | 286 ++++++++++++++++++
2 files changed, 289 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
index dbe74b1b08f8c..8c92060128a12 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
@@ -2508,7 +2508,9 @@ bool SchedGroup::canAddSU(SUnit &SU) const {
++E;
// Return true if all of the bundled MIs can be added to this group.
- return std::all_of(B, E, [this](MachineInstr &MI) { return canAddMI(MI); });
+ return std::all_of(B, E, [this](MachineInstr &MI) {
+ return (MI.isMetaInstruction()) || canAddMI(MI);
+ });
}
void SchedGroup::initSchedGroup() {
diff --git a/llvm/test/CodeGen/AMDGPU/sched-barrier-post-RA.mir b/llvm/test/CodeGen/AMDGPU/sched-barrier-post-RA.mir
index 7bdb8f5b35ec5..a49d2188ce704 100644
--- a/llvm/test/CodeGen/AMDGPU/sched-barrier-post-RA.mir
+++ b/llvm/test/CodeGen/AMDGPU/sched-barrier-post-RA.mir
@@ -5,6 +5,10 @@
define amdgpu_kernel void @no_sched_barrier(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
define amdgpu_kernel void @sched_barrier_mask_0(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
define amdgpu_kernel void @sched_barrier_mask_1(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
+ define amdgpu_kernel void @no_sched_barrier_many_loads(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
+ define amdgpu_kernel void @sched_barrier_mask_1924(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
+ define amdgpu_kernel void @extra_sched_barrier_in_bundle(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
+ define amdgpu_kernel void @extra_sched_barrier_in_bundle_2(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) { ret void }
!0 = distinct !{!0}
!1 = !{!1, !0}
@@ -120,3 +124,285 @@ body: |
}
S_ENDPGM 0
...
+
+
+# No SCHED_BARRIER - First two bundles global_loads get clustered and VALU instrucions are scheduled later.
+
+---
+name: no_sched_barrier_many_loads
+tracksRegLiveness: true
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: no_sched_barrier_many_loads
+ ; CHECK: renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit killed $sgpr0_sgpr1, implicit killed $vgpr0, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR killed renamable $sgpr0_sgpr1, killed renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr10 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: S_ENDPGM 0
+ renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit $sgpr0_sgpr1, implicit $vgpr0, implicit $exec {
+ renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ renamable $vgpr10 = IMPLICIT_DEF
+ renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+ BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ S_ENDPGM 0
+...
+
+# MASK = 0b 0111 1000 0100 SALU, MFMA/WMMA, All DS, All DS Read, All DS Write, All Trans may be
+# scheduled across SCHED_BARRIER. VALU and all VMEM, VMEM Read/Write cannot be
+# scheduled across SCHED_BARRIER. V_MULs is not shceduled after the second bundle of loads.
+
+---
+name: sched_barrier_mask_1924
+tracksRegLiveness: true
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: sched_barrier_mask_1924
+ ; CHECK: renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit killed $sgpr0_sgpr1, implicit killed $vgpr0, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR killed renamable $sgpr0_sgpr1, killed renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr10 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+ ; CHECK-NEXT: SCHED_BARRIER 1924
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: S_ENDPGM 0
+ renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit $sgpr0_sgpr1, implicit $vgpr0, implicit $exec {
+ renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ renamable $vgpr10 = IMPLICIT_DEF
+ renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+
+ SCHED_BARRIER 1924
+
+ BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ S_ENDPGM 0
+...
+
+# MASK = 0b 0111 1000 0100 SALU, MFMA/WMMA, All DS, All DS Read, All DS Write, All Trans may be
+# scheduled across SCHED_BARRIER. VALU and all VMEM, VMEM Read/Write cannot be
+# scheduled across SCHED_BARRIER. V_MULs is not shceduled after the second bundle of loads.
+#
+# A SCHED_BARRIER before the seconde bundle is honored.
+
+
+---
+name: extra_sched_barrier_in_bundle
+tracksRegLiveness: true
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: extra_sched_barrier_in_bundle
+ ; CHECK: renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit killed $sgpr0_sgpr1, implicit killed $vgpr0, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR killed renamable $sgpr0_sgpr1, killed renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr10 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+ ; CHECK-NEXT: SCHED_BARRIER 1924
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: SCHED_BARRIER 1924
+ ; CHECK-NEXT: renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: S_ENDPGM 0
+ renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit $sgpr0_sgpr1, implicit $vgpr0, implicit $exec {
+ renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ renamable $vgpr10 = IMPLICIT_DEF
+ renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+
+ SCHED_BARRIER 1924
+
+ BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ SCHED_BARRIER 1924
+ renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ S_ENDPGM 0
+...
+
+# MASK = 0b 0111 1000 0100 SALU, MFMA/WMMA, All DS, All DS Read, All DS Write, All Trans may be
+# scheduled across SCHED_BARRIER. VALU and all VMEM, VMEM Read/Write cannot be
+# scheduled across SCHED_BARRIER. V_MULs is not shceduled after the second bundle of loads.
+#
+# A SCHED_BARRIER in the seconde bundle is ignored.
+# TODO: The SCHED_BARRIER may be honored.
+
+---
+name: extra_sched_barrier_in_bundle_2
+tracksRegLiveness: true
+body: |
+ bb.0:
+ ; CHECK-LABEL: name: extra_sched_barrier_in_bundle_2
+ ; CHECK: renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit killed $sgpr0_sgpr1, implicit killed $vgpr0, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR killed renamable $sgpr0_sgpr1, killed renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr10 = IMPLICIT_DEF
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ ; CHECK-NEXT: renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: SCHED_BARRIER 1924
+ ; CHECK-NEXT: renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK-NEXT: renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+ ; CHECK-NEXT: renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: S_ENDPGM 0
+ renamable $sgpr0_sgpr1 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ BUNDLE implicit-def $vgpr1, implicit-def $vgpr1_lo16, implicit-def $vgpr1_hi16, implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit $sgpr0_sgpr1, implicit $vgpr0, implicit $exec {
+ renamable $vgpr1 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ renamable $vgpr2 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr0_sgpr1, renamable $vgpr0, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr2_sgpr3 = IMPLICIT_DEF
+ renamable $vgpr10 = IMPLICIT_DEF
+ renamable $vgpr1 = nsw V_MUL_LO_U32_e64 killed $vgpr1, $vgpr1, implicit $exec
+ renamable $vgpr2 = nsw V_MUL_LO_U32_e64 killed $vgpr2, $vgpr2, implicit $exec
+
+ BUNDLE implicit-def $vgpr11, implicit-def $vgpr11_lo16, implicit-def $vgpr11_hi16, implicit-def $vgpr12, implicit-def $vgpr12_lo16, implicit-def $vgpr12_hi16, implicit $sgpr2_sgpr3, implicit $vgpr10, implicit $exec {
+ renamable $vgpr11 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 0, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ SCHED_BARRIER 1924
+ renamable $vgpr12 = GLOBAL_LOAD_DWORD_SADDR renamable $sgpr2_sgpr3, renamable $vgpr10, 512, 0, implicit $exec :: (load (s32) from %ir.in, !alias.scope !0, addrspace 1)
+ }
+ renamable $sgpr4_sgpr5 = IMPLICIT_DEF
+ renamable $vgpr0 = IMPLICIT_DEF
+ renamable $vgpr11 = nsw V_MUL_LO_U32_e64 killed $vgpr11, $vgpr11, implicit $exec
+ renamable $vgpr12 = nsw V_MUL_LO_U32_e64 killed $vgpr12, $vgpr12, implicit $exec
+ BUNDLE implicit killed $vgpr10, implicit killed $vgpr11, implicit killed $sgpr2_sgpr3, implicit $exec, implicit killed $vgpr12 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr10, killed renamable $vgpr11, renamable $sgpr2_sgpr3, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr10, killed renamable $vgpr12, killed renamable $sgpr2_sgpr3, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ BUNDLE implicit killed $vgpr0, implicit killed $vgpr1, implicit killed $sgpr4_sgpr5, implicit $exec, implicit killed $vgpr2 {
+ GLOBAL_STORE_DWORD_SADDR renamable $vgpr0, killed renamable $vgpr1, renamable $sgpr4_sgpr5, 0, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ GLOBAL_STORE_DWORD_SADDR killed renamable $vgpr0, killed renamable $vgpr2, killed renamable $sgpr4_sgpr5, 512, 0, implicit $exec :: (store (s32) into %ir.out, !noalias !0, addrspace 1)
+ }
+ S_ENDPGM 0
+...
>From c1b4e379d6d9f6485da506a0199a2eb4a7ec7e41 Mon Sep 17 00:00:00 2001
From: Yoonseo Choi <Yoonseo.Choi at amd.com>
Date: Fri, 8 Aug 2025 17:29:24 -0500
Subject: [PATCH 2/2] Simplifying scanning inside a bundle
---
llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp | 17 ++++++++++-------
1 file changed, 10 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
index 8c92060128a12..c47ca42007f82 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
@@ -2502,15 +2502,18 @@ bool SchedGroup::canAddSU(SUnit &SU) const {
return canAddMI(MI);
// Special case for bundled MIs.
+ // Return true if all of the bundled MIs can be added to this group.
+ // A meta instruction in a bundle is an exception.
const MachineBasicBlock *MBB = MI.getParent();
- MachineBasicBlock::instr_iterator B = MI.getIterator(), E = ++B;
- while (E != MBB->end() && E->isBundledWithPred())
- ++E;
+ // Initially, iterator is on a bundler header.
+ MachineBasicBlock::instr_iterator B = std::next(MI.getIterator());
+ while (B != MBB->end() && B->isBundledWithPred()) {
+ if (!B->isMetaInstruction() && !canAddMI(*B))
+ return false;
+ ++B;
+ }
- // Return true if all of the bundled MIs can be added to this group.
- return std::all_of(B, E, [this](MachineInstr &MI) {
- return (MI.isMetaInstruction()) || canAddMI(MI);
- });
+ return true;
}
void SchedGroup::initSchedGroup() {
More information about the llvm-commits
mailing list