[PATCH] D87398: AMDGPU: Fix inserting waitcnts before kill uses

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, kerbowa.
Herald added subscribers: hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, jvesely, kzhuravl.
Herald added a project: LLVM.
arsenm requested review of this revision.
Herald added a subscriber: wdng.

https://reviews.llvm.org/D87398

Files:
  llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
  llvm/test/CodeGen/AMDGPU/waitcnt-meta-instructions.mir


Index: llvm/test/CodeGen/AMDGPU/waitcnt-meta-instructions.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/waitcnt-meta-instructions.mir
@@ -0,0 +1,66 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=si-insert-waitcnts -o - %s | FileCheck -check-prefix=GCN %s
+
+# Make sure no waitcnt is inserted for meta instruction uses.
+
+---
+
+name: waitcnt_kill
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GCN-LABEL: name: waitcnt_kill
+    ; GCN: S_WAITCNT 0
+    ; GCN: $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    ; GCN: KILL $vgpr0
+    $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    KILL $vgpr0
+...
+
+---
+
+name: waitcnt_implicit_def
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GCN-LABEL: name: waitcnt_implicit_def
+    ; GCN: S_WAITCNT 0
+    ; GCN: $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    ; GCN: $vgpr0 = IMPLICIT_DEF
+    $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+...
+
+---
+
+name: waitcnt_eh_label
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2
+    ; GCN-LABEL: name: waitcnt_eh_label
+    ; GCN: S_WAITCNT 0
+    ; GCN: $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    ; GCN: EH_LABEL <mcsymbol Ltmp0>, implicit $vgpr0
+    $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    EH_LABEL <mcsymbol Ltmp0>, implicit $vgpr0
+
+...
+
+---
+
+name: waitcnt_cfi
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2
+    ; GCN-LABEL: name: waitcnt_cfi
+    ; GCN: S_WAITCNT 0
+    ; GCN: $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    ; GCN: CFI_INSTRUCTION offset $vgpr0_lo16, 16
+    $vgpr0 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, 0, 0, implicit $exec
+    CFI_INSTRUCTION offset $vgpr0, 16
+
+...
Index: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -855,7 +855,7 @@
   setForceEmitWaitcnt();
   bool IsForceEmitWaitcnt = isForceEmitWaitcnt();
 
-  if (MI.isDebugInstr())
+  if (MI.isMetaInstruction())
     return false;
 
   AMDGPU::Waitcnt Wait;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D87398.290768.patch
Type: text/x-patch
Size: 2477 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200909/be064147/attachment.bin>