[ARM] Remove kill flags between VCMP and insertion point

When we fold a VCMP into a VPST instruction any kill flags between the
old VCMP position and the new insertion point need to be removed, in
order to keep the verifier happy.

Differential Revision: https://reviews.llvm.org/D90964

diff --git a/llvm/lib/Target/ARM/MVEVPTBlockPass.cpp b/llvm/lib/Target/ARM/MVEVPTBlockPass.cpp
index dc769ae..232db61 100644 (file)
--- a/llvm/lib/Target/ARM/MVEVPTBlockPass.cpp
+++ b/llvm/lib/Target/ARM/MVEVPTBlockPass.cpp
@@ -270,6 +270,15 @@ bool MVEVPTBlock::InsertVPTBlocks(MachineBasicBlock &Block) {
       MIBuilder.add(VCMP->getOperand(1));
       MIBuilder.add(VCMP->getOperand(2));
       MIBuilder.add(VCMP->getOperand(3));
+
+      // We need to remove any kill flags between the original VCMP and the new
+      // insertion point.
+      for (MachineInstr &MI :
+           make_range(VCMP->getIterator(), MI->getIterator())) {
+        MI.clearRegisterKills(VCMP->getOperand(1).getReg(), TRI);
+        MI.clearRegisterKills(VCMP->getOperand(2).getReg(), TRI);
+      }
+
       VCMP->eraseFromParent();
     } else {
       MIBuilder = BuildMI(Block, MI, DL, TII->get(ARM::MVE_VPST));

diff --git a/llvm/test/CodeGen/Thumb2/mve-vpt-block-kill.mir b/llvm/test/CodeGen/Thumb2/mve-vpt-block-kill.mir
new file mode 100644 (file)
index 0000000..b222b9f
--- /dev/null
+++ b/llvm/test/CodeGen/Thumb2/mve-vpt-block-kill.mir
@@ -0,0 +1,38 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp -run-pass arm-mve-vpt %s -o - -verify-machineinstrs | FileCheck %s
+
+# Check we remove kill flags when combining VCMP into a VPST
+
+--- |
+  define i32 @b() {
+    ret i32 undef
+  }
+
+...
+---
+name:            b
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $r0, $r1, $r4, $r6, $r7, $lr
+
+    ; CHECK-LABEL: name: b
+    ; CHECK: liveins: $r0, $r1, $r4, $r6, $r7, $lr
+    ; CHECK: $sp = frame-setup t2STMDB_UPD $sp, 14 /* CC::al */, $noreg, killed $r4, killed $r6, killed $r7, killed $lr
+    ; CHECK: renamable $q0 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q0
+    ; CHECK: renamable $q1 = MVE_VMOVimmi32 1, 0, $noreg, undef renamable $q1
+    ; CHECK: renamable $q2 = MVE_VADD_qr_i32 renamable $q1, renamable $r1, 0, $noreg, undef renamable $q2
+    ; CHECK: BUNDLE implicit-def dead $vpr, implicit-def $q0, implicit-def $d0, implicit-def $s0, implicit-def $s1, implicit-def $d1, implicit-def $s2, implicit-def $s3, implicit killed $q0, implicit $q1, implicit killed $q2 {
+    ; CHECK:   MVE_VPTv4u32 8, renamable $q0, renamable $q1, 8, implicit-def $vpr
+    ; CHECK:   renamable $q0 = MVE_VORR killed renamable $q0, killed renamable $q2, 1, internal killed renamable $vpr, renamable $q0
+    ; CHECK: }
+    ; CHECK: $sp = frame-destroy t2LDMIA_RET $sp, 14 /* CC::al */, $noreg, def $r4, def $r6, def $r7, def $pc, implicit undef $r0
+    $sp = frame-setup t2STMDB_UPD $sp, 14 /* CC::al */, $noreg, killed $r4, killed $r6, killed $r7, killed $lr
+    renamable $q0 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q0
+    renamable $q1 = MVE_VMOVimmi32 1, 0, $noreg, undef renamable $q1
+    renamable $vpr = MVE_VCMPu32 renamable $q0, renamable $q1, 8, 0, $noreg
+    renamable $q2 = MVE_VADD_qr_i32 killed renamable $q1, renamable $r1, 0, $noreg, undef renamable $q2
+    renamable $q0 = MVE_VORR killed renamable $q0, killed renamable $q2, 1, killed renamable $vpr, renamable $q0
+    $sp = frame-destroy t2LDMIA_RET $sp, 14 /* CC::al */, $noreg, def $r4, def $r6, def $r7, def $pc, implicit undef $r0
+
+...