[PATCH] D44849: AMDGPU: Fix FP restore from being reordered with stack ops

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added a reviewer: rampitec.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, kzhuravl.

In a function, s5 is used as the frame base SGPR. Before it was
possible if there were stack operations immediately after
the call to be moved before the FP restore copy after the call.

      

I'm not 100% satisfied with the solution of adding an implicit
use to the ADJCALLSTACK* pseudo, but I'm not sure what would
be better.


https://reviews.llvm.org/D44849

Files:
  lib/Target/AMDGPU/SIISelLowering.cpp
  test/CodeGen/AMDGPU/call-preserved-registers.ll


Index: test/CodeGen/AMDGPU/call-preserved-registers.ll
===================================================================
--- test/CodeGen/AMDGPU/call-preserved-registers.ll
+++ test/CodeGen/AMDGPU/call-preserved-registers.ll
@@ -34,9 +34,9 @@
 ; GCN: s_mov_b32 s33, s5
 ; GCN-NEXT: s_swappc_b64
 ; GCN-NEXT: s_mov_b32 s5, s33
+; GCN-NEXT: s_mov_b32 s33, s5
 ; GCN-NEXT: ;;#ASMSTART
 ; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_mov_b32 s33, s5
 ; GCN-NEXT: s_swappc_b64
 ; GCN-NEXT: s_mov_b32 s5, s33
 ; GCN: v_readlane_b32 s37, v32, 4
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -3351,8 +3351,13 @@
   case AMDGPU::ADJCALLSTACKDOWN: {
     const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
     MachineInstrBuilder MIB(*MF, &MI);
+
+    // Add an implicit use of the frame offset reg to prevent the restore copy
+    // inserted after the call from being reorderd after stack operations in the
+    // the caller's frame.
     MIB.addReg(Info->getStackPtrOffsetReg(), RegState::ImplicitDefine)
-        .addReg(Info->getStackPtrOffsetReg(), RegState::Implicit);
+        .addReg(Info->getStackPtrOffsetReg(), RegState::Implicit)
+        .addReg(Info->getFrameOffsetReg(), RegState::Implicit);
     return BB;
   }
   case AMDGPU::SI_CALL_ISEL:


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D44849.139662.patch
Type: text/x-patch
Size: 1417 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180323/d5f0e014/attachment.bin>