[AMDGPU][SDAG] Handle ISD::PTRADD in SelectionDAGAddressAnalysis

[ritter-x2a]

This is used in a bunch of memory-related transforms.

For SWDEV-516125.

[ritter-x2a]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• [AMDGPU][SDAG] Handle ISD::PTRADD in SelectionDAGAddressAnalysis #142778 👈 (View in Graphite)
• [AMDGPU][SDAG] Add test for ISD::PTRADD handling in SelectionDAGAddressAnalysis #142777
• [AMDGPU][SDAG] Add ISD::PTRADD DAG combines #142739
• [AMDGPU][SDAG] Add tests for ISD::PTRADD DAG combines #142738
• [AMDGPU][SDAG] Initial support for ISD::PTRADD #141725
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: Fabian Ritter (ritter-x2a)

Changes

This is used in a bunch of memory-related transforms.

For SWDEV-516125.

---

Full diff: https://github.com/llvm/llvm-project/pull/142778.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAGAddressAnalysis.cpp (+3-3)
• (modified) llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll (+8-20)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGAddressAnalysis.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGAddressAnalysis.cpp
index f2ab88851b780..da92aaa860b2b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGAddressAnalysis.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGAddressAnalysis.cpp
@@ -231,6 +231,7 @@ static BaseIndexOffset matchLSNode(const LSBaseSDNode *N,
         }
       break;
     case ISD::ADD:
+    case ISD::PTRADD:
       if (auto *C = dyn_cast<ConstantSDNode>(Base->getOperand(1))) {
         Offset += C->getSExtValue();
         Base = DAG.getTargetLoweringInfo().unwrapAddress(Base->getOperand(0));
@@ -259,7 +260,7 @@ static BaseIndexOffset matchLSNode(const LSBaseSDNode *N,
     break;
   }
 
-  if (Base->getOpcode() == ISD::ADD) {
+  if (Base->isAnyAdd()) {
     // TODO: The following code appears to be needless as it just
     //       bails on some Ptrs early, reducing the cases where we
     //       find equivalence. We should be able to remove this.
@@ -282,8 +283,7 @@ static BaseIndexOffset matchLSNode(const LSBaseSDNode *N,
     }
 
     // Check if Index Offset pattern
-    if (Index->getOpcode() != ISD::ADD ||
-        !isa<ConstantSDNode>(Index->getOperand(1)))
+    if (!Index->isAnyAdd() || !isa<ConstantSDNode>(Index->getOperand(1)))
       return BaseIndexOffset(PotentialBase, Index, Offset, IsIndexSignExt);
 
     Offset += cast<ConstantSDNode>(Index->getOperand(1))->getSExtValue();
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
index bce59307446ce..1069339774894 100644
--- a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
+++ b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
@@ -136,26 +136,14 @@ declare noalias ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
 ; Taken from memcpy-param-combinations.ll, tests PTRADD handling in
 ; SelectionDAGAddressAnalysis.
 define void @memcpy_p1_p4_sz16_align_1_1(ptr addrspace(1) align 1 %dst, ptr addrspace(4) align 1 readonly %src) {
-; GFX942_PTRADD-LABEL: memcpy_p1_p4_sz16_align_1_1:
-; GFX942_PTRADD:       ; %bb.0: ; %entry
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
-; GFX942_PTRADD-NEXT:    global_store_dwordx2 v[0:1], v[4:5], off
-; GFX942_PTRADD-NEXT:    global_load_dwordx2 v[2:3], v[2:3], off offset:8
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
-; GFX942_PTRADD-NEXT:    global_store_dwordx2 v[0:1], v[2:3], off offset:8
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: memcpy_p1_p4_sz16_align_1_1:
-; GFX942_LEGACY:       ; %bb.0: ; %entry
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    global_load_dwordx4 v[2:5], v[2:3], off
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
-; GFX942_LEGACY-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: memcpy_p1_p4_sz16_align_1_1:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    global_load_dwordx4 v[2:5], v[2:3], off
+; GFX942-NEXT:    s_waitcnt vmcnt(0)
+; GFX942-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off
+; GFX942-NEXT:    s_waitcnt vmcnt(0)
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
 entry:
   tail call void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noundef nonnull align 1 %dst, ptr addrspace(4) noundef nonnull align 1 %src, i64 16, i1 false)
   ret void