[llvm-branch-commits] [llvm] [SelectionDAG] Split vector types for atomic load (PR #120640)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Tue May 27 10:34:44 PDT 2025
https://github.com/jofrn updated https://github.com/llvm/llvm-project/pull/120640
>From f91634795e4dbfd6e081f0b096b872411051ff0f Mon Sep 17 00:00:00 2001
From: jofrn <jofernau at amd.com>
Date: Thu, 19 Dec 2024 16:25:55 -0500
Subject: [PATCH] [SelectionDAG] Split vector types for atomic load
Vector types that aren't widened are split
so that a single ATOMIC_LOAD is issued for the entire vector at once.
This change utilizes the load vectorization infrastructure in
SelectionDAG in order to group the vectors. This enables SelectionDAG
to translate vectors with type bfloat,half.
commit-id:3a045357
---
llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h | 1 +
.../SelectionDAG/LegalizeVectorTypes.cpp | 37 ++++++++
llvm/test/CodeGen/X86/atomic-load-store.ll | 86 +++++++++++++++++++
3 files changed, 124 insertions(+)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
index b6e018ba0e454..b15f70cbec1cd 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
@@ -965,6 +965,7 @@ class LLVM_LIBRARY_VISIBILITY DAGTypeLegalizer {
void SplitVecRes_FPOp_MultiType(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_IS_FPCLASS(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, SDValue &Hi);
+ void SplitVecRes_ATOMIC_LOAD(AtomicSDNode *LD, SDValue &Lo, SDValue &Hi);
void SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo, SDValue &Hi);
void SplitVecRes_VP_LOAD(VPLoadSDNode *LD, SDValue &Lo, SDValue &Hi);
void SplitVecRes_VP_STRIDED_LOAD(VPStridedLoadSDNode *SLD, SDValue &Lo,
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 42763aab5bb55..af48a5e216803 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -1161,6 +1161,9 @@ void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {
SplitVecRes_STEP_VECTOR(N, Lo, Hi);
break;
case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;
+ case ISD::ATOMIC_LOAD:
+ SplitVecRes_ATOMIC_LOAD(cast<AtomicSDNode>(N), Lo, Hi);
+ break;
case ISD::LOAD:
SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi);
break;
@@ -1414,6 +1417,40 @@ void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {
SetSplitVector(SDValue(N, ResNo), Lo, Hi);
}
+void DAGTypeLegalizer::SplitVecRes_ATOMIC_LOAD(AtomicSDNode *LD, SDValue &Lo,
+ SDValue &Hi) {
+ assert(LD->getExtensionType() == ISD::NON_EXTLOAD &&
+ "Extended load during type legalization!");
+ SDLoc dl(LD);
+ EVT VT = LD->getValueType(0);
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
+
+ SDValue Ch = LD->getChain();
+ SDValue Ptr = LD->getBasePtr();
+
+ EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
+ EVT MemIntVT =
+ EVT::getIntegerVT(*DAG.getContext(), LD->getMemoryVT().getSizeInBits());
+ SDValue ALD = DAG.getAtomicLoad(ISD::NON_EXTLOAD, dl, MemIntVT, IntVT, Ch,
+ Ptr, LD->getMemOperand());
+
+ EVT LoIntVT = EVT::getIntegerVT(*DAG.getContext(), LoVT.getSizeInBits());
+ EVT HiIntVT = EVT::getIntegerVT(*DAG.getContext(), HiVT.getSizeInBits());
+ SDValue ExtractLo = DAG.getNode(ISD::TRUNCATE, dl, LoIntVT, ALD);
+ SDValue ExtractHi =
+ DAG.getNode(ISD::SRL, dl, IntVT, ALD,
+ DAG.getIntPtrConstant(VT.getSizeInBits() / 2, dl));
+ ExtractHi = DAG.getNode(ISD::TRUNCATE, dl, HiIntVT, ExtractHi);
+
+ Lo = DAG.getBitcast(LoVT, ExtractLo);
+ Hi = DAG.getBitcast(HiVT, ExtractHi);
+
+ // Legalize the chain result - switch anything that used the old chain to
+ // use the new one.
+ ReplaceValueWith(SDValue(LD, 1), ALD.getValue(1));
+}
+
void DAGTypeLegalizer::IncrementPointer(MemSDNode *N, EVT MemVT,
MachinePointerInfo &MPI, SDValue &Ptr,
uint64_t *ScaledOffset) {
diff --git a/llvm/test/CodeGen/X86/atomic-load-store.ll b/llvm/test/CodeGen/X86/atomic-load-store.ll
index 535a87316e162..039edcbf83544 100644
--- a/llvm/test/CodeGen/X86/atomic-load-store.ll
+++ b/llvm/test/CodeGen/X86/atomic-load-store.ll
@@ -374,6 +374,74 @@ define <2 x float> @atomic_vec2_float_align(ptr %x) {
ret <2 x float> %ret
}
+define <2 x half> @atomic_vec2_half(ptr %x) {
+; CHECK-O3-LABEL: atomic_vec2_half:
+; CHECK-O3: # %bb.0:
+; CHECK-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-O3-NEXT: retq
+;
+; CHECK-SSE-O3-LABEL: atomic_vec2_half:
+; CHECK-SSE-O3: # %bb.0:
+; CHECK-SSE-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-SSE-O3-NEXT: retq
+;
+; CHECK-AVX-O3-LABEL: atomic_vec2_half:
+; CHECK-AVX-O3: # %bb.0:
+; CHECK-AVX-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-AVX-O3-NEXT: retq
+;
+; CHECK-O0-LABEL: atomic_vec2_half:
+; CHECK-O0: # %bb.0:
+; CHECK-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-O0-NEXT: retq
+;
+; CHECK-SSE-O0-LABEL: atomic_vec2_half:
+; CHECK-SSE-O0: # %bb.0:
+; CHECK-SSE-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-SSE-O0-NEXT: retq
+;
+; CHECK-AVX-O0-LABEL: atomic_vec2_half:
+; CHECK-AVX-O0: # %bb.0:
+; CHECK-AVX-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-AVX-O0-NEXT: retq
+ %ret = load atomic <2 x half>, ptr %x acquire, align 4
+ ret <2 x half> %ret
+}
+
+define <2 x bfloat> @atomic_vec2_bfloat(ptr %x) {
+; CHECK-O3-LABEL: atomic_vec2_bfloat:
+; CHECK-O3: # %bb.0:
+; CHECK-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-O3-NEXT: retq
+;
+; CHECK-SSE-O3-LABEL: atomic_vec2_bfloat:
+; CHECK-SSE-O3: # %bb.0:
+; CHECK-SSE-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-SSE-O3-NEXT: retq
+;
+; CHECK-AVX-O3-LABEL: atomic_vec2_bfloat:
+; CHECK-AVX-O3: # %bb.0:
+; CHECK-AVX-O3-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-AVX-O3-NEXT: retq
+;
+; CHECK-O0-LABEL: atomic_vec2_bfloat:
+; CHECK-O0: # %bb.0:
+; CHECK-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-O0-NEXT: retq
+;
+; CHECK-SSE-O0-LABEL: atomic_vec2_bfloat:
+; CHECK-SSE-O0: # %bb.0:
+; CHECK-SSE-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-SSE-O0-NEXT: retq
+;
+; CHECK-AVX-O0-LABEL: atomic_vec2_bfloat:
+; CHECK-AVX-O0: # %bb.0:
+; CHECK-AVX-O0-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; CHECK-AVX-O0-NEXT: retq
+ %ret = load atomic <2 x bfloat>, ptr %x acquire, align 4
+ ret <2 x bfloat> %ret
+}
+
define <1 x ptr> @atomic_vec1_ptr(ptr %x) nounwind {
; CHECK-O3-LABEL: atomic_vec1_ptr:
; CHECK-O3: # %bb.0:
@@ -835,6 +903,24 @@ define <4 x i16> @atomic_vec4_i16(ptr %x) nounwind {
ret <4 x i16> %ret
}
+define <4 x half> @atomic_vec4_half(ptr %x) nounwind {
+; CHECK-LABEL: atomic_vec4_half:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movq (%rdi), %xmm0
+; CHECK-NEXT: retq
+ %ret = load atomic <4 x half>, ptr %x acquire, align 8
+ ret <4 x half> %ret
+}
+
+define <4 x bfloat> @atomic_vec4_bfloat(ptr %x) nounwind {
+; CHECK-LABEL: atomic_vec4_bfloat:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movq (%rdi), %xmm0
+; CHECK-NEXT: retq
+ %ret = load atomic <4 x bfloat>, ptr %x acquire, align 8
+ ret <4 x bfloat> %ret
+}
+
define <4 x float> @atomic_vec4_float_align(ptr %x) nounwind {
; CHECK-O3-LABEL: atomic_vec4_float_align:
; CHECK-O3: # %bb.0:
More information about the llvm-branch-commits
mailing list