https://github.com/ritter-x2a updated
https://github.com/llvm/llvm-project/pull/146075
>From 452008111a34c815b38242272063654393261921 Mon Sep 17 00:00:00 2001
From: Fabian Ritter <fabian.rit...@amd.com>
Date: Fri, 27 Jun 2025 04:23:50 -0400
Subject: [PATCH 1/3] [AMDGPU][SDAG] DAGCombine PTRADD -> disjoint OR
If we can't fold a PTRADD's offset into its users, lowering them to
disjoint ORs is preferable: Often, a 32-bit OR instruction suffices
where we'd otherwise use a pair of 32-bit additions with carry.
This needs to be a DAGCombine (and not a selection rule) because its
main purpose is to enable subsequent DAGCombines for bitwise operations.
We don't want to just turn PTRADDs into disjoint ORs whenever that's
sound because this transform loses the information that the operation
implements pointer arithmetic, which we will soon need to fold offsets
into FLAT instructions. Currently, disjoint ORs can still be used for
offset folding, so that part of the logic can't be tested.
The PR contains a hacky workaround for a situation where an AssertAlign
operand of a PTRADD is not DAGCombined before the PTRADD, causing the
PTRADD to be turned into a disjoint OR although reassociating it with
the operand of the AssertAlign would be better. This wouldn't be a
problem if the DAGCombiner ensured that a node is only processed after
all its operands have been processed.
For SWDEV-516125.
---
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 35 ++++++++++++
.../AMDGPU/ptradd-sdag-optimizations.ll | 56 ++++++++++++++++++-
2 files changed, 90 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 822bab88c8a09..71230078edc69 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15136,6 +15136,41 @@ SDValue SITargetLowering::performPtrAddCombine(SDNode
*N,
return Folded;
}
+ // Transform (ptradd a, b) -> (or disjoint a, b) if it is equivalent and if
+ // that transformation can't block an offset folding at any use of the
ptradd.
+ // This should be done late, after legalization, so that it doesn't block
+ // other ptradd combines that could enable more offset folding.
+ bool HasIntermediateAssertAlign =
+ N0->getOpcode() == ISD::AssertAlign && N0->getOperand(0)->isAnyAdd();
+ // This is a hack to work around an ordering problem for DAGs like this:
+ // (ptradd (AssertAlign (ptradd p, c1), k), c2)
+ // If the outer ptradd is handled first by the DAGCombiner, it can be
+ // transformed into a disjoint or. Then, when the generic AssertAlign combine
+ // pushes the AssertAlign through the inner ptradd, it's too late for the
+ // ptradd reassociation to trigger.
+ if (!DCI.isBeforeLegalizeOps() && !HasIntermediateAssertAlign &&
+ DAG.haveNoCommonBitsSet(N0, N1)) {
+ bool TransformCanBreakAddrMode = any_of(N->users(), [&](SDNode *User) {
+ if (auto *LoadStore = dyn_cast<MemSDNode>(User);
+ LoadStore && LoadStore->getBasePtr().getNode() == N) {
+ unsigned AS = LoadStore->getAddressSpace();
+ // Currently, we only really need ptradds to fold offsets into flat
+ // memory instructions.
+ if (AS != AMDGPUAS::FLAT_ADDRESS)
+ return false;
+ TargetLoweringBase::AddrMode AM;
+ AM.HasBaseReg = true;
+ EVT VT = LoadStore->getMemoryVT();
+ Type *AccessTy = VT.getTypeForEVT(*DAG.getContext());
+ return isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS);
+ }
+ return false;
+ });
+
+ if (!TransformCanBreakAddrMode)
+ return DAG.getNode(ISD::OR, DL, VT, N0, N1, SDNodeFlags::Disjoint);
+ }
+
if (N1.getOpcode() != ISD::ADD || !N1.hasOneUse())
return SDValue();
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
index 893deb35fe822..64e041103a563 100644
--- a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
+++ b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
@@ -100,7 +100,7 @@ define void @baseptr_null(i64 %offset, i8 %v) {
; Taken from implicit-kernarg-backend-usage.ll, tests the PTRADD handling in
the
; assertalign DAG combine.
-define amdgpu_kernel void @llvm_amdgcn_queue_ptr(ptr addrspace(1) %ptr) #0 {
+define amdgpu_kernel void @llvm_amdgcn_queue_ptr(ptr addrspace(1) %ptr) {
; GFX942-LABEL: llvm_amdgcn_queue_ptr:
; GFX942: ; %bb.0:
; GFX942-NEXT: v_mov_b32_e32 v2, 0
@@ -416,6 +416,60 @@ entry:
ret void
}
+; Check that ptradds can be lowered to disjoint ORs.
+define ptr @gep_disjoint_or(ptr %base) {
+; GFX942-LABEL: gep_disjoint_or:
+; GFX942: ; %bb.0:
+; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT: v_and_or_b32 v0, v0, -16, 4
+; GFX942-NEXT: s_setpc_b64 s[30:31]
+ %p = call ptr @llvm.ptrmask(ptr %base, i64 s0xf0)
+ %gep = getelementptr nuw inbounds i8, ptr %p, i64 4
+ ret ptr %gep
+}
+
+; Check that AssertAlign nodes between ptradd nodes don't block offset folding,
+; taken from preload-implicit-kernargs.ll
+define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg
%out) #0 {
+; GFX942_PTRADD-LABEL: random_incorrect_offset:
+; GFX942_PTRADD: ; %bb.1:
+; GFX942_PTRADD-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX942_PTRADD-NEXT: s_waitcnt lgkmcnt(0)
+; GFX942_PTRADD-NEXT: s_branch .LBB21_0
+; GFX942_PTRADD-NEXT: .p2align 8
+; GFX942_PTRADD-NEXT: ; %bb.2:
+; GFX942_PTRADD-NEXT: .LBB21_0:
+; GFX942_PTRADD-NEXT: s_load_dword s0, s[0:1], 0xa
+; GFX942_PTRADD-NEXT: v_mov_b32_e32 v0, 0
+; GFX942_PTRADD-NEXT: s_waitcnt lgkmcnt(0)
+; GFX942_PTRADD-NEXT: v_mov_b32_e32 v1, s0
+; GFX942_PTRADD-NEXT: global_store_dword v0, v1, s[2:3]
+; GFX942_PTRADD-NEXT: s_endpgm
+;
+; GFX942_LEGACY-LABEL: random_incorrect_offset:
+; GFX942_LEGACY: ; %bb.1:
+; GFX942_LEGACY-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX942_LEGACY-NEXT: s_waitcnt lgkmcnt(0)
+; GFX942_LEGACY-NEXT: s_branch .LBB21_0
+; GFX942_LEGACY-NEXT: .p2align 8
+; GFX942_LEGACY-NEXT: ; %bb.2:
+; GFX942_LEGACY-NEXT: .LBB21_0:
+; GFX942_LEGACY-NEXT: s_mov_b32 s4, 8
+; GFX942_LEGACY-NEXT: s_load_dword s0, s[0:1], s4 offset:0x2
+; GFX942_LEGACY-NEXT: v_mov_b32_e32 v0, 0
+; GFX942_LEGACY-NEXT: s_waitcnt lgkmcnt(0)
+; GFX942_LEGACY-NEXT: v_mov_b32_e32 v1, s0
+; GFX942_LEGACY-NEXT: global_store_dword v0, v1, s[2:3]
+; GFX942_LEGACY-NEXT: s_endpgm
+ %imp_arg_ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
+ %gep = getelementptr i8, ptr addrspace(4) %imp_arg_ptr, i32 2
+ %load = load i32, ptr addrspace(4) %gep
+ store i32 %load, ptr addrspace(1) %out
+ ret void
+}
+
declare void @llvm.memcpy.p0.p4.i64(ptr noalias nocapture writeonly, ptr
addrspace(4) noalias nocapture readonly, i64, i1 immarg)
!0 = !{}
+
+attributes #0 = { "amdgpu-agpr-alloc"="0" "amdgpu-no-completion-action"
"amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr"
"amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-lds-kernel-id"
"amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x"
"amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x"
"amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z"
"uniform-work-group-size"="false" }
>From 607f982f6bba48aa2b9c72ff34cb61257f22f0e0 Mon Sep 17 00:00:00 2001
From: Fabian Ritter <fabian.rit...@amd.com>
Date: Mon, 30 Jun 2025 02:39:22 -0400
Subject: [PATCH 2/3] Remove unnecessary attributes in test
---
.../AMDGPU/ptradd-sdag-optimizations.ll | 18 ++++++++----------
1 file changed, 8 insertions(+), 10 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
index 64e041103a563..a03bd97309c6b 100644
--- a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
+++ b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
@@ -430,36 +430,36 @@ define ptr @gep_disjoint_or(ptr %base) {
; Check that AssertAlign nodes between ptradd nodes don't block offset folding,
; taken from preload-implicit-kernargs.ll
-define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg
%out) #0 {
+define amdgpu_kernel void @random_incorrect_offset(ptr addrspace(1) inreg
%out) {
; GFX942_PTRADD-LABEL: random_incorrect_offset:
; GFX942_PTRADD: ; %bb.1:
-; GFX942_PTRADD-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX942_PTRADD-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
; GFX942_PTRADD-NEXT: s_waitcnt lgkmcnt(0)
; GFX942_PTRADD-NEXT: s_branch .LBB21_0
; GFX942_PTRADD-NEXT: .p2align 8
; GFX942_PTRADD-NEXT: ; %bb.2:
; GFX942_PTRADD-NEXT: .LBB21_0:
-; GFX942_PTRADD-NEXT: s_load_dword s0, s[0:1], 0xa
+; GFX942_PTRADD-NEXT: s_load_dword s0, s[4:5], 0xa
; GFX942_PTRADD-NEXT: v_mov_b32_e32 v0, 0
; GFX942_PTRADD-NEXT: s_waitcnt lgkmcnt(0)
; GFX942_PTRADD-NEXT: v_mov_b32_e32 v1, s0
-; GFX942_PTRADD-NEXT: global_store_dword v0, v1, s[2:3]
+; GFX942_PTRADD-NEXT: global_store_dword v0, v1, s[8:9]
; GFX942_PTRADD-NEXT: s_endpgm
;
; GFX942_LEGACY-LABEL: random_incorrect_offset:
; GFX942_LEGACY: ; %bb.1:
-; GFX942_LEGACY-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX942_LEGACY-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
; GFX942_LEGACY-NEXT: s_waitcnt lgkmcnt(0)
; GFX942_LEGACY-NEXT: s_branch .LBB21_0
; GFX942_LEGACY-NEXT: .p2align 8
; GFX942_LEGACY-NEXT: ; %bb.2:
; GFX942_LEGACY-NEXT: .LBB21_0:
-; GFX942_LEGACY-NEXT: s_mov_b32 s4, 8
-; GFX942_LEGACY-NEXT: s_load_dword s0, s[0:1], s4 offset:0x2
+; GFX942_LEGACY-NEXT: s_mov_b32 s0, 8
+; GFX942_LEGACY-NEXT: s_load_dword s0, s[4:5], s0 offset:0x2
; GFX942_LEGACY-NEXT: v_mov_b32_e32 v0, 0
; GFX942_LEGACY-NEXT: s_waitcnt lgkmcnt(0)
; GFX942_LEGACY-NEXT: v_mov_b32_e32 v1, s0
-; GFX942_LEGACY-NEXT: global_store_dword v0, v1, s[2:3]
+; GFX942_LEGACY-NEXT: global_store_dword v0, v1, s[8:9]
; GFX942_LEGACY-NEXT: s_endpgm
%imp_arg_ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
%gep = getelementptr i8, ptr addrspace(4) %imp_arg_ptr, i32 2
@@ -471,5 +471,3 @@ define amdgpu_kernel void @random_incorrect_offset(ptr
addrspace(1) inreg %out)
declare void @llvm.memcpy.p0.p4.i64(ptr noalias nocapture writeonly, ptr
addrspace(4) noalias nocapture readonly, i64, i1 immarg)
!0 = !{}
-
-attributes #0 = { "amdgpu-agpr-alloc"="0" "amdgpu-no-completion-action"
"amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr"
"amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-lds-kernel-id"
"amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x"
"amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x"
"amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z"
"uniform-work-group-size"="false" }
>From f0f708e3914879153c1b770fa3c1b1661e8903ca Mon Sep 17 00:00:00 2001
From: Fabian Ritter <fabian.rit...@amd.com>
Date: Mon, 30 Jun 2025 09:28:34 -0400
Subject: [PATCH 3/3] Move ptradd -> disjoint OR combine to generic combines
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 27 ++++++++++++++
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 35 -------------------
2 files changed, 27 insertions(+), 35 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 3626ac45a4860..8a8abf5654cd4 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -2766,6 +2766,33 @@ SDValue DAGCombiner::visitPTRADD(SDNode *N) {
}
}
+ // Transform (ptradd a, b) -> (or disjoint a, b) if it is equivalent and if
+ // that transformation can't block an offset folding at any use of the
ptradd.
+ // This should be done late, after legalization, so that it doesn't block
+ // other ptradd combines that could enable more offset folding.
+ if (LegalOperations && DAG.haveNoCommonBitsSet(N0, N1)) {
+ bool TransformCanBreakAddrMode = false;
+ if (auto *C = dyn_cast<ConstantSDNode>(N1)) {
+ TargetLoweringBase::AddrMode AM;
+ AM.HasBaseReg = true;
+ AM.BaseOffs = C->getSExtValue();
+ TransformCanBreakAddrMode = any_of(N->users(), [&](SDNode *User) {
+ if (auto *LoadStore = dyn_cast<MemSDNode>(User);
+ LoadStore && LoadStore->getBasePtr().getNode() == N) {
+ unsigned AS = LoadStore->getAddressSpace();
+ EVT AccessVT = LoadStore->getMemoryVT();
+ Type *AccessTy = AccessVT.getTypeForEVT(*DAG.getContext());
+ return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy,
+ AS);
+ }
+ return false;
+ });
+ }
+
+ if (!TransformCanBreakAddrMode)
+ return DAG.getNode(ISD::OR, DL, PtrVT, N0, N1, SDNodeFlags::Disjoint);
+ }
+
return SDValue();
}
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 71230078edc69..822bab88c8a09 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15136,41 +15136,6 @@ SDValue SITargetLowering::performPtrAddCombine(SDNode
*N,
return Folded;
}
- // Transform (ptradd a, b) -> (or disjoint a, b) if it is equivalent and if
- // that transformation can't block an offset folding at any use of the
ptradd.
- // This should be done late, after legalization, so that it doesn't block
- // other ptradd combines that could enable more offset folding.
- bool HasIntermediateAssertAlign =
- N0->getOpcode() == ISD::AssertAlign && N0->getOperand(0)->isAnyAdd();
- // This is a hack to work around an ordering problem for DAGs like this:
- // (ptradd (AssertAlign (ptradd p, c1), k), c2)
- // If the outer ptradd is handled first by the DAGCombiner, it can be
- // transformed into a disjoint or. Then, when the generic AssertAlign combine
- // pushes the AssertAlign through the inner ptradd, it's too late for the
- // ptradd reassociation to trigger.
- if (!DCI.isBeforeLegalizeOps() && !HasIntermediateAssertAlign &&
- DAG.haveNoCommonBitsSet(N0, N1)) {
- bool TransformCanBreakAddrMode = any_of(N->users(), [&](SDNode *User) {
- if (auto *LoadStore = dyn_cast<MemSDNode>(User);
- LoadStore && LoadStore->getBasePtr().getNode() == N) {
- unsigned AS = LoadStore->getAddressSpace();
- // Currently, we only really need ptradds to fold offsets into flat
- // memory instructions.
- if (AS != AMDGPUAS::FLAT_ADDRESS)
- return false;
- TargetLoweringBase::AddrMode AM;
- AM.HasBaseReg = true;
- EVT VT = LoadStore->getMemoryVT();
- Type *AccessTy = VT.getTypeForEVT(*DAG.getContext());
- return isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS);
- }
- return false;
- });
-
- if (!TransformCanBreakAddrMode)
- return DAG.getNode(ISD::OR, DL, VT, N0, N1, SDNodeFlags::Disjoint);
- }
-
if (N1.getOpcode() != ISD::ADD || !N1.hasOneUse())
return SDValue();
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
