[llvm] 538a83e - RegisterCoalescer: Add undef flags in removePartialRedundancy (#75152)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Dec 13 00:02:57 PST 2023
Author: Matt Arsenault
Date: 2023-12-13T15:02:53+07:00
New Revision: 538a83e4b9ddaa6cc9e8680fa97b2e33cd03192d
URL: https://github.com/llvm/llvm-project/commit/538a83e4b9ddaa6cc9e8680fa97b2e33cd03192d
DIFF: https://github.com/llvm/llvm-project/commit/538a83e4b9ddaa6cc9e8680fa97b2e33cd03192d.diff
LOG: RegisterCoalescer: Add undef flags in removePartialRedundancy (#75152)
If the copy being hoisted was undef, we have the same problems that
eliminateUndefCopy needs to solve. We would effectively be introducing a
new live out implicit_def. We need to add an undef flag to avoid
artificially introducing a live through undef value. Previously, the
verifier would fail due to the dead def inside the loop providing the
live in value for the %1 use.
Added:
llvm/test/CodeGen/X86/coalescer-partial-redundancy-clear-dead-flag-undef-copy.mir
Modified:
llvm/lib/CodeGen/RegisterCoalescer.cpp
Removed:
################################################################################
diff --git a/llvm/lib/CodeGen/RegisterCoalescer.cpp b/llvm/lib/CodeGen/RegisterCoalescer.cpp
index c067d87a9fd81..2ef91e78f3370 100644
--- a/llvm/lib/CodeGen/RegisterCoalescer.cpp
+++ b/llvm/lib/CodeGen/RegisterCoalescer.cpp
@@ -1201,6 +1201,8 @@ bool RegisterCoalescer::removePartialRedundancy(const CoalescerPair &CP,
<< printMBBReference(MBB) << '\t' << CopyMI);
}
+ const bool IsUndefCopy = CopyMI.getOperand(1).isUndef();
+
// Remove CopyMI.
// Note: This is fine to remove the copy before updating the live-ranges.
// While updating the live-ranges, we only look at slot indices and
@@ -1214,6 +1216,19 @@ bool RegisterCoalescer::removePartialRedundancy(const CoalescerPair &CP,
LIS->pruneValue(*static_cast<LiveRange *>(&IntB), CopyIdx.getRegSlot(),
&EndPoints);
BValNo->markUnused();
+
+ if (IsUndefCopy) {
+ // We're introducing an undef phi def, and need to set undef on any users of
+ // the previously local def to avoid artifically extending the lifetime
+ // through the block.
+ for (MachineOperand &MO : MRI->use_nodbg_operands(IntB.reg())) {
+ const MachineInstr &MI = *MO.getParent();
+ SlotIndex UseIdx = LIS->getInstructionIndex(MI);
+ if (!IntB.liveAt(UseIdx))
+ MO.setIsUndef(true);
+ }
+ }
+
// Extend IntB to the EndPoints of its original live interval.
LIS->extendToIndices(IntB, EndPoints);
diff --git a/llvm/test/CodeGen/X86/coalescer-partial-redundancy-clear-dead-flag-undef-copy.mir b/llvm/test/CodeGen/X86/coalescer-partial-redundancy-clear-dead-flag-undef-copy.mir
new file mode 100644
index 0000000000000..5f33be0bc1555
--- /dev/null
+++ b/llvm/test/CodeGen/X86/coalescer-partial-redundancy-clear-dead-flag-undef-copy.mir
@@ -0,0 +1,47 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
+# RUN: llc -mtriple=x86_64-pc-linux-gnu -run-pass=register-coalescer -verify-coalescing -o - %s | FileCheck %s
+
+# Check for "Live range continues after dead def flag".
+
+# There are 2 copies of undef, but the registers also appear to be
+# live due to block live outs, and thus were not deleted as
+# eliminateUndefCopy only considered the live range, and not the undef
+# flag.
+#
+# removePartialRedundancy would move the COPY undef %0 in bb.1 to
+# bb.0. The live range of %1 would then be extended to be live out of
+# %bb.1 for the backedge phi. This would then fail the verifier, since
+# the dead flag was no longer valid. This was fixed by directly
+# considering the undef flag to avoid considering this special case.
+
+---
+name: partial_redundancy_coalesce_undef_copy_live_out
+tracksRegLiveness: true
+body: |
+ ; CHECK-LABEL: name: partial_redundancy_coalesce_undef_copy_live_out
+ ; CHECK: bb.0:
+ ; CHECK-NEXT: successors: %bb.1(0x80000000)
+ ; CHECK-NEXT: liveins: $rdi
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $rdi
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.1:
+ ; CHECK-NEXT: successors: %bb.1(0x80000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: dead [[XOR32ri:%[0-9]+]]:gr32 = XOR32ri undef [[XOR32ri]], 1, implicit-def dead $eflags
+ ; CHECK-NEXT: dead [[MOV32rr:%[0-9]+]]:gr32 = MOV32rr [[COPY]]
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = IMPLICIT_DEF
+ ; CHECK-NEXT: JMP_1 %bb.1
+ bb.0:
+ liveins: $rdi
+
+ %0:gr32 = COPY $rdi
+
+ bb.1:
+ %1:gr32 = COPY undef %0
+ dead %1:gr32 = XOR32ri %1, 1, implicit-def dead $eflags
+ dead %2:gr32 = MOV32rr killed %0
+ %0:gr32 = COPY killed undef %1
+ JMP_1 %bb.1
+
+...
More information about the llvm-commits
mailing list