[PATCH] D149080: [AMDGPU] Fix negative offset values interpretation in getMemOperandsWithOffset for DS

[Janek van Oirschot via Phabricator via llvm-commits]

This revision was landed with ongoing or failed builds.
This revision was automatically updated to reflect the committed changes.
Closed by commit rG124acb7ca31a: [AMDGPU] Fix negative offset values interpretation in getMemOperandsWithOffset… (authored by JanekvO).

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D149080/new/

https://reviews.llvm.org/D149080

Files:
  llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
  llvm/test/CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll
  llvm/test/CodeGen/AMDGPU/triv-disjoint-mem-access-neg-offset.mir


Index: llvm/test/CodeGen/AMDGPU/triv-disjoint-mem-access-neg-offset.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/triv-disjoint-mem-access-neg-offset.mir
@@ -0,0 +1,37 @@
+# RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -enable-misched --debug-only=machine-scheduler -run-pass=machine-scheduler -o /dev/null %s 2>&1 | FileCheck %s
+# REQUIRES: asserts
+
+# Make sure handling of unsigned immediate values interpreted as negative values
+# still works for SIInstrInfo::areMemAccessesTriviallyDisjoint.
+
+# LABEL: {{^}}no_reorder_flat_load_local_store_local_load:
+# CHECK:      SU(5):   %5:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+# CHECK:      SU(6):   DS_WRITE_B128_gfx9 %5:vgpr_32, %4:vreg_128, 512, 0, implicit $exec
+# CHECK:      SU(7):   %6:vreg_64 = DS_READ2_B32_gfx9 %5:vgpr_32, -127, -126, 0, implicit $exec
+# CHECK:      Predecessors:
+# CHECK-NEXT: SU(6): Ord
+# CHECK-NEXT: SU(5): Data
+---
+name:            no_reorder_flat_load_local_store_local_load
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1
+
+    %1:sgpr_64(p4) = COPY $sgpr0_sgpr1
+    %4:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %1(p4), 36, 0
+    %5:sreg_64_xexec = S_LOAD_DWORDX2_IMM %1(p4), 44, 0
+    %7:vreg_64 = COPY %5
+    %6:vreg_128 = FLAT_LOAD_DWORDX4 killed %7, 0, 0, implicit $exec, implicit $flat_scr
+    %8:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    DS_WRITE_B128_gfx9 %8, killed %6, 512, 0, implicit $exec
+    %9:vreg_64 = DS_READ2_B32_gfx9 %8, -127, -126, 0, implicit $exec
+    %10:sreg_32 = COPY %9.sub1
+    %11:sreg_32 = COPY %9.sub0
+    %12:vgpr_32 = COPY %4
+    %13:vgpr_32 = COPY %11
+    %14:vgpr_32 = COPY %10
+    DS_WRITE2_B32_gfx9 killed %12, killed %13, killed %14, 0, 1, 0, implicit $exec
+    S_ENDPGM 0
+
+...
Index: llvm/test/CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll
+++ llvm/test/CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll
@@ -1,10 +1,25 @@
 ; RUN: llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=bonaire -enable-amdgpu-aa=0 -verify-machineinstrs -enable-misched -enable-aa-sched-mi < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CI %s
 ; RUN: llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=gfx900 -enable-amdgpu-aa=0 -verify-machineinstrs -enable-misched -enable-aa-sched-mi < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9 %s
 
+%struct.lds = type { [64 x ptr], [16 x i8] }
+ at stored_lds_struct = addrspace(3) global %struct.lds addrspace(3) *undef, align 16
 @stored_lds_ptr = addrspace(3) global ptr addrspace(3) undef, align 4
 @stored_constant_ptr = addrspace(3) global ptr addrspace(4) undef, align 8
 @stored_global_ptr = addrspace(3) global ptr addrspace(1) undef, align 8
 
+; GCN-LABEL: {{^}}no_reorder_flat_load_local_store_local_load:
+; GCN: flat_load_dwordx4
+; GCN: ds_write_b128 {{v[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}} offset:512
+; GCN: ds_read2_b32 {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset0:129 offset1:130
+define amdgpu_kernel void @no_reorder_flat_load_local_store_local_load(ptr addrspace(3) %out, ptr %fptr) #0 {
+  %ptr1 = getelementptr %struct.lds, ptr addrspace(3) @stored_lds_struct, i32 0, i32 1
+  %ptr2 = getelementptr %struct.lds, ptr addrspace(3) @stored_lds_struct, i32 0, i32 1, i32 4
+  call void @llvm.memcpy.p3.p0(ptr addrspace(3) align 16 %ptr1, ptr align 8 %fptr, i64 16, i1 false)
+  %vector_load = load <2 x i32>, ptr addrspace(3) %ptr2, align 4
+  store <2 x i32> %vector_load, ptr addrspace(3) %out, align 4
+  ret void
+}
+
 ; GCN-LABEL: {{^}}reorder_local_load_global_store_local_load:
 ; CI: ds_read2_b32 {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset0:1 offset1:3
 ; CI: buffer_store_dword
@@ -312,6 +327,7 @@
   ret void
 }
 
+declare void @llvm.memcpy.p3.p0(ptr addrspace(3), ptr, i64, i1)
 declare void @llvm.amdgcn.s.barrier() #1
 declare i32 @llvm.amdgcn.workitem.id.x() #2
 declare void @llvm.amdgcn.struct.tbuffer.store.v4i32(<4 x i32>, <4 x i32>, i32, i32, i32, i32 immarg, i32 immarg) #3
Index: llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -329,8 +329,8 @@
       const MachineOperand *Offset1Op =
           getNamedOperand(LdSt, AMDGPU::OpName::offset1);
 
-      unsigned Offset0 = Offset0Op->getImm();
-      unsigned Offset1 = Offset1Op->getImm();
+      unsigned Offset0 = Offset0Op->getImm() & 0xff;
+      unsigned Offset1 = Offset1Op->getImm() & 0xff;
       if (Offset0 + 1 != Offset1)
         return false;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D149080.517141.patch
Type: text/x-patch
Size: 4739 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230426/0feca7f0/attachment.bin>