[PATCH] D84477: AMDGPU: Don't assume there is only one terminator copy

[Matt Arsenault via Phabricator via llvm-commits]

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

This would stop on the first in reverse order, failing the verifier if
there were more earlier in the block.


https://reviews.llvm.org/D84477

Files:
  llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
  llvm/test/CodeGen/AMDGPU/optimize-exec-masking-strip-terminator-bits.mir


Index: llvm/test/CodeGen/AMDGPU/optimize-exec-masking-strip-terminator-bits.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/optimize-exec-masking-strip-terminator-bits.mir
@@ -0,0 +1,44 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs -run-pass=si-optimize-exec-masking -o - %s | FileCheck %s
+
+# There are multiple _term pseudos here. Starting from the end of the
+# block, they all need to be converted to regular copies removed in
+# order to avoid making the verifier unhappy.
+
+---
+name:  multi_term_pseudos
+tracksRegLiveness: true
+body:             |
+  ; CHECK-LABEL: name: multi_term_pseudos
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9
+  ; CHECK:   $exec = COPY killed renamable $sgpr4_sgpr5
+  ; CHECK:   renamable $sgpr10_sgpr11 = COPY killed renamable $sgpr6_sgpr7, implicit $exec
+  ; CHECK:   renamable $sgpr12_sgpr13 = COPY killed renamable $sgpr8_sgpr9, implicit $exec
+  ; CHECK:   S_CBRANCH_EXECZ %bb.1, implicit $exec
+  ; CHECK:   S_BRANCH %bb.2
+  ; CHECK: bb.1:
+  ; CHECK:   liveins: $sgpr12_sgpr13
+  ; CHECK:   S_ENDPGM 0, implicit $sgpr12_sgpr13
+  ; CHECK: bb.2:
+  ; CHECK:   liveins: $sgpr12_sgpr13
+  ; CHECK:   S_ENDPGM 0, implicit $sgpr12_sgpr13
+  bb.0:
+    successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9
+    $exec = S_MOV_B64_term killed renamable $sgpr4_sgpr5
+    renamable $sgpr10_sgpr11 = S_MOV_B64_term killed renamable $sgpr6_sgpr7, implicit $exec
+    renamable $sgpr12_sgpr13 = S_MOV_B64_term killed renamable $sgpr8_sgpr9, implicit $exec
+    S_CBRANCH_EXECZ %bb.1, implicit $exec
+    S_BRANCH %bb.2
+
+  bb.1:
+    liveins: $sgpr12_sgpr13
+    S_ENDPGM 0, implicit $sgpr12_sgpr13
+
+  bb.2:
+    liveins: $sgpr12_sgpr13
+    S_ENDPGM 0, implicit $sgpr12_sgpr13
+
+...
Index: llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
+++ llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
@@ -224,15 +224,22 @@
   const SIInstrInfo &TII,
   MachineBasicBlock &MBB) {
   MachineBasicBlock::reverse_iterator I = MBB.rbegin(), E = MBB.rend();
+
+  bool Seen = false;
+  MachineBasicBlock::reverse_iterator FirstNonTerm = I;
   for (; I != E; ++I) {
     if (!I->isTerminator())
-      return I;
+      return FirstNonTerm;
 
-    if (removeTerminatorBit(TII, *I))
-      return I;
+    if (removeTerminatorBit(TII, *I)) {
+      if (!Seen) {
+        FirstNonTerm = I;
+        Seen = true;
+      }
+    }
   }
 
-  return E;
+  return FirstNonTerm;
 }
 
 static MachineBasicBlock::reverse_iterator findExecCopy(
@@ -291,6 +298,7 @@
     if (I == E)
       continue;
 
+    // TODO: It's possible to see other terminator copies after the exec copy.
     Register CopyToExec = isCopyToExec(*I, ST);
     if (!CopyToExec.isValid())
       continue;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D84477.280293.patch
Type: text/x-patch
Size: 3123 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200723/ee795c6b/attachment.bin>