From d1d0e207b699ec995dc8078e87fda4125fe720b7 Mon Sep 17 00:00:00 2001
From: Jessica Paquette <jpaquette@apple.com>
Date: Wed, 24 Jul 2019 17:18:51 +0000
Subject: [PATCH] [AArch64][GlobalISel] Make vector dup optimization look at
 last elt of ZeroVec

Fix an off-by-one error which made us not look at the last element of the
zero vector. This caused a miscompile in 188.ammp.

Differential Revision: https://reviews.llvm.org/D65168

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@366930 91177308-0d34-0410-b5e6-96231b3b80d8
---
 .../AArch64/AArch64InstructionSelector.cpp    |  2 +-
 .../AArch64/GlobalISel/opt-shuffle-splat.mir  | 40 +++++++++++++++++++
 2 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/lib/Target/AArch64/AArch64InstructionSelector.cpp b/lib/Target/AArch64/AArch64InstructionSelector.cpp
index 892d469d232..3bce4df88da 100644
--- a/lib/Target/AArch64/AArch64InstructionSelector.cpp
+++ b/lib/Target/AArch64/AArch64InstructionSelector.cpp
@@ -3523,7 +3523,7 @@ bool AArch64InstructionSelector::tryOptVectorDup(MachineInstr &I) const {
   int64_t Zero = 0;
   if (!mi_match(ZeroVec->getOperand(1).getReg(), MRI, m_ICst(Zero)) || Zero)
     return false;
-  for (unsigned i = 1, e = ZeroVec->getNumOperands() - 1; i < e; ++i) {
+  for (unsigned i = 1, e = ZeroVec->getNumOperands(); i < e; ++i) {
     if (ZeroVec->getOperand(i).getReg() != ZeroVec->getOperand(1).getReg())
       return false; // This wasn't an all zeros vector.
   }
diff --git a/test/CodeGen/AArch64/GlobalISel/opt-shuffle-splat.mir b/test/CodeGen/AArch64/GlobalISel/opt-shuffle-splat.mir
index 6c831e13c7f..b62f569dc4d 100644
--- a/test/CodeGen/AArch64/GlobalISel/opt-shuffle-splat.mir
+++ b/test/CodeGen/AArch64/GlobalISel/opt-shuffle-splat.mir
@@ -120,6 +120,14 @@ body:             |
 
     ; This test is exactly the same as splat_2xf64, except it adds two copies.
     ; These copies shouldn't get in the way of matching the dup pattern.
+    ; CHECK-LABEL: name: splat_2xf64_copies
+    ; CHECK: liveins: $d0
+    ; CHECK: [[COPY:%[0-9]+]]:fpr64 = COPY $d0
+    ; CHECK: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF
+    ; CHECK: [[INSERT_SUBREG:%[0-9]+]]:fpr128 = INSERT_SUBREG [[DEF]], [[COPY]], %subreg.dsub
+    ; CHECK: [[DUPv2i64lane:%[0-9]+]]:fpr128 = DUPv2i64lane [[INSERT_SUBREG]], 0
+    ; CHECK: $q0 = COPY [[DUPv2i64lane]]
+    ; CHECK: RET_ReallyLR implicit $q0
     %0:fpr(s64) = COPY $d0
     %2:fpr(<2 x s64>) = G_IMPLICIT_DEF
     %6:fpr(<2 x s64>) = COPY %2
@@ -130,3 +138,35 @@ body:             |
     %4:fpr(<2 x s64>) = G_SHUFFLE_VECTOR %7(<2 x s64>), %2, %5(<2 x s32>)
     $q0 = COPY %4(<2 x s64>)
     RET_ReallyLR implicit $q0
+
+...
+---
+name:            not_all_zeros
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+    ; Make sure that we don't do the optimization when it's not all zeroes.
+    ; CHECK-LABEL: name: not_all_zeros
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF
+    ; CHECK: [[INSvi64gpr:%[0-9]+]]:fpr128 = INSvi64gpr [[DEF]], 0, [[COPY]]
+    ; CHECK: [[ADRP:%[0-9]+]]:gpr64common = ADRP target-flags(aarch64-page) %const.0
+    ; CHECK: [[LDRQui:%[0-9]+]]:fpr128 = LDRQui [[ADRP]], target-flags(aarch64-pageoff, aarch64-nc) %const.0
+    ; CHECK: [[REG_SEQUENCE:%[0-9]+]]:qq = REG_SEQUENCE [[INSvi64gpr]], %subreg.qsub0, [[DEF]], %subreg.qsub1
+    ; CHECK: [[TBLv16i8Two:%[0-9]+]]:fpr128 = TBLv16i8Two [[REG_SEQUENCE]], [[LDRQui]]
+    ; CHECK: $q0 = COPY [[TBLv16i8Two]]
+    ; CHECK: RET_ReallyLR implicit $q0
+    %0:gpr(s64) = COPY $x0
+    %2:fpr(<2 x s64>) = G_IMPLICIT_DEF
+    %3:gpr(s32) = G_CONSTANT i32 0
+    %6:gpr(s32) = G_CONSTANT i32 1
+    %5:fpr(<2 x s32>) = G_BUILD_VECTOR %3(s32), %6(s32)
+    %1:fpr(<2 x s64>) = G_INSERT_VECTOR_ELT %2, %0(s64), %3(s32)
+    %4:fpr(<2 x s64>) = G_SHUFFLE_VECTOR %1(<2 x s64>), %2, %5(<2 x s32>)
+    $q0 = COPY %4(<2 x s64>)
+    RET_ReallyLR implicit $q0
-- 
2.40.0