[llvm-branch-commits] [llvm] [AMDGPU] System scope atomics are emulated over PCIe in gfx1250 (PR #152369)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Aug 6 12:21:45 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Stanislav Mekhanoshin (rampitec)
<details>
<summary>Changes</summary>
HW will emulate unsupported PCIe atomics via CAS loop, we do not need to
expand these anymore.
---
Patch is 73.26 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/152369.diff
5 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPU.td (+9)
- (modified) llvm/lib/Target/AMDGPU/GCNSubtarget.h (+7)
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+3-2)
- (added) llvm/test/CodeGen/AMDGPU/atomics-system-scope.ll (+1486)
- (modified) llvm/test/CodeGen/AMDGPU/literal64.ll (+2-18)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index d84f512f4976d..ddeca07e51103 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1013,6 +1013,14 @@ def FeatureAgentScopeFineGrainedRemoteMemoryAtomics
"device memory."
>;
+def FeatureEmulatedSystemScopeAtomics
+ : SubtargetFeature<"emulated-system-scope-atomics",
+ "HasEmulatedSystemScopeAtomics",
+ "true",
+ "System scope atomics unsupported by the PCI-e are emulated in HW via CAS "
+ "loop and functional."
+>;
+
def FeatureDefaultComponentZero : SubtargetFeature<"default-component-zero",
"HasDefaultComponentZero",
"true",
@@ -2062,6 +2070,7 @@ def FeatureISAVersion12_50 : FeatureSet<
FeatureAtomicFMinFMaxF64FlatInsts,
FeatureFlatBufferGlobalAtomicFaddF64Inst,
FeatureMemoryAtomicFAddF32DenormalSupport,
+ FeatureEmulatedSystemScopeAtomics,
FeatureGloballyAddressableScratch,
FeatureKernargPreload,
FeatureVmemPrefInsts,
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 9114f249c92a7..1c3749d81eec8 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -187,6 +187,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
bool HasFlatBufferGlobalAtomicFaddF64Inst = false;
bool HasDefaultComponentZero = false;
bool HasAgentScopeFineGrainedRemoteMemoryAtomics = false;
+ bool HasEmulatedSystemScopeAtomics = false;
bool HasDefaultComponentBroadcast = false;
bool HasXF32Insts = false;
/// The maximum number of instructions that may be placed within an S_CLAUSE,
@@ -950,6 +951,12 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
return HasAgentScopeFineGrainedRemoteMemoryAtomics;
}
+ /// \return true is HW emulates system scope atomics unsupported by the PCI-e
+ /// via CAS loop.
+ bool hasEmulatedSystemScopeAtomics() const {
+ return HasEmulatedSystemScopeAtomics;
+ }
+
bool hasDefaultComponentZero() const { return HasDefaultComponentZero; }
bool hasDefaultComponentBroadcast() const {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 63826b782a377..8f44c03d95b43 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -17695,6 +17695,8 @@ static bool globalMemoryFPAtomicIsLegal(const GCNSubtarget &Subtarget,
if (Subtarget.supportsAgentScopeFineGrainedRemoteMemoryAtomics() &&
RMW->hasMetadata("amdgpu.no.remote.memory"))
return true;
+ if (Subtarget.hasEmulatedSystemScopeAtomics())
+ return true;
} else if (Subtarget.supportsAgentScopeFineGrainedRemoteMemoryAtomics())
return true;
@@ -17942,8 +17944,7 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
case AtomicRMWInst::UMax: {
if (AMDGPU::isFlatGlobalAddrSpace(AS) ||
AS == AMDGPUAS::BUFFER_FAT_POINTER) {
- // Always expand system scope min/max atomics.
- if (HasSystemScope)
+ if (HasSystemScope && !Subtarget->hasEmulatedSystemScopeAtomics())
return AtomicExpansionKind::CmpXChg;
}
diff --git a/llvm/test/CodeGen/AMDGPU/atomics-system-scope.ll b/llvm/test/CodeGen/AMDGPU/atomics-system-scope.ll
new file mode 100644
index 0000000000000..5fc9f4a0f8038
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/atomics-system-scope.ll
@@ -0,0 +1,1486 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN:llc -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck --check-prefix=GFX1250 %s
+
+define float @global_system_atomic_fadd_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_system_atomic_fadd_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val monotonic
+ ret float %result
+}
+
+define float @global_one_as_atomic_fadd_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_one_as_atomic_fadd_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic
+ ret float %result
+}
+
+define double @global_system_atomic_fadd_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_system_atomic_fadd_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val monotonic
+ ret double %result
+}
+
+define double @global_one_as_atomic_fadd_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_one_as_atomic_fadd_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic
+ ret double %result
+}
+
+define float @global_system_atomic_fmin_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_system_atomic_fmin_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val monotonic
+ ret float %result
+}
+
+define float @global_one_as_atomic_fmin_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_one_as_atomic_fmin_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic
+ ret float %result
+}
+
+define double @global_system_atomic_fmin_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_system_atomic_fmin_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_num_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val monotonic
+ ret double %result
+}
+
+define double @global_one_as_atomic_fmin_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_one_as_atomic_fmin_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_num_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic
+ ret double %result
+}
+
+define float @global_system_atomic_fmax_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_system_atomic_fmax_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val monotonic
+ ret float %result
+}
+
+define float @global_one_as_atomic_fmax_f32(ptr addrspace(1) %ptr, float %val) {
+; GFX1250-LABEL: global_one_as_atomic_fmax_f32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("one-as") monotonic
+ ret float %result
+}
+
+define double @global_system_atomic_fmax_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_system_atomic_fmax_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_num_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val monotonic
+ ret double %result
+}
+
+define double @global_one_as_atomic_fmax_f64(ptr addrspace(1) %ptr, double %val) {
+; GFX1250-LABEL: global_one_as_atomic_fmax_f64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_num_f64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("one-as") monotonic
+ ret double %result
+}
+
+define i32 @global_one_as_atomic_min_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_one_as_atomic_min_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_i32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw min ptr addrspace(1) %ptr, i32 %val syncscope("one-as") monotonic
+ ret i32 %result
+}
+
+define i32 @global_system_atomic_min_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_system_atomic_min_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_i32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw min ptr addrspace(1) %ptr, i32 %val monotonic
+ ret i32 %result
+}
+
+define i32 @global_one_as_atomic_max_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_one_as_atomic_max_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_i32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw max ptr addrspace(1) %ptr, i32 %val syncscope("one-as") monotonic
+ ret i32 %result
+}
+
+define i32 @global_system_atomic_max_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_system_atomic_max_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_i32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw max ptr addrspace(1) %ptr, i32 %val monotonic
+ ret i32 %result
+}
+
+define i32 @global_one_as_atomic_umin_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_one_as_atomic_umin_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umin ptr addrspace(1) %ptr, i32 %val syncscope("one-as") monotonic
+ ret i32 %result
+}
+
+define i32 @global_system_atomic_umin_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_system_atomic_umin_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umin ptr addrspace(1) %ptr, i32 %val monotonic
+ ret i32 %result
+}
+
+define i32 @global_one_as_atomic_umax_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_one_as_atomic_umax_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umax ptr addrspace(1) %ptr, i32 %val syncscope("one-as") monotonic
+ ret i32 %result
+}
+
+define i32 @global_system_atomic_umax_i32(ptr addrspace(1) %ptr, i32 %val) {
+; GFX1250-LABEL: global_system_atomic_umax_i32:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umax ptr addrspace(1) %ptr, i32 %val monotonic
+ ret i32 %result
+}
+
+define i64 @global_one_as_atomic_min_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_one_as_atomic_min_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_i64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw min ptr addrspace(1) %ptr, i64 %val syncscope("one-as") monotonic
+ ret i64 %result
+}
+
+define i64 @global_system_atomic_min_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_system_atomic_min_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_i64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw min ptr addrspace(1) %ptr, i64 %val monotonic
+ ret i64 %result
+}
+
+define i64 @global_one_as_atomic_max_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_one_as_atomic_max_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_i64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw max ptr addrspace(1) %ptr, i64 %val syncscope("one-as") monotonic
+ ret i64 %result
+}
+
+define i64 @global_system_atomic_max_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_system_atomic_max_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_i64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw max ptr addrspace(1) %ptr, i64 %val monotonic
+ ret i64 %result
+}
+
+define i64 @global_one_as_atomic_umin_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_one_as_atomic_umin_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_u64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umin ptr addrspace(1) %ptr, i64 %val syncscope("one-as") monotonic
+ ret i64 %result
+}
+
+define i64 @global_system_atomic_umin_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_system_atomic_umin_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_min_u64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umin ptr addrspace(1) %ptr, i64 %val monotonic
+ ret i64 %result
+}
+
+define i64 @global_one_as_atomic_umax_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_one_as_atomic_umax_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_u64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umax ptr addrspace(1) %ptr, i64 %val syncscope("one-as") monotonic
+ ret i64 %result
+}
+
+define i64 @global_system_atomic_umax_i64(ptr addrspace(1) %ptr, i64 %val) {
+; GFX1250-LABEL: global_system_atomic_umax_i64:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_atomic_max_u64 v[0:1], v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw umax ptr addrspace(1) %ptr, i64 %val monotonic
+ ret i64 %result
+}
+
+define i16 @global_one_as_atomic_min_i16(ptr addrspace(1) %ptr, i16 %val) {
+; GFX1250-LABEL: global_one_as_atomic_min_i16:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: v_mov_b32_e32 v3, v0
+; GFX1250-NEXT: s_mov_b32 s0, 0
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX1250-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX1250-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX1250-NEXT: global_load_b32 v5, v[0:1], off
+; GFX1250-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_not_b32_e32 v4, v4
+; GFX1250-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX1250-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: v_mov_b32_e32 v7, v5
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX1250-NEXT: v_min_i16 v5, v5, v2
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX1250-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX1250-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[6:7], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v7
+; GFX1250-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX1250-NEXT: s_wait_xcnt 0x0
+; GFX1250-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX1250-NEXT: s_cbranch_execnz .LBB28_1
+; GFX1250-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX1250-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX1250-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %result = atomicrmw min ptr addrspace(1) %ptr, i16 %val syncscope("one-as") monotonic
+ ret i16 %resu...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/152369
More information about the llvm-branch-commits
mailing list