[llvm] [MachineOutliner] Preserve instruction bundles (PR #106402)

[Simon Tatham via llvm-commits]

https://github.com/statham-arm created https://github.com/llvm/llvm-project/pull/106402

When the machine outliner copies instructions from a source function into an outlined function, it was doing it using `CloneMachineInstr`, which is documented as not preserving the interior of any instruction bundle. So outlining code that includes an instruction bundle would fail, because in the outlined version, the bundle would be empty, so instructions would go missing in the move.

This occurs when any bundled instruction appears in the outlined code, so there was no need to construct an unusual test case: I've just copied a function from the existing `stp-opt-with-renaming.mir`, which happens to contain an SVE instruction bundle. Including two identical copies of that function makes the outliner merge them, and then we check that it didn't destroy the interior of the bundle in the process.

>From 86e618aba4a364af67a99b67315211193af0389c Mon Sep 17 00:00:00 2001
From: Simon Tatham <simon.tatham at arm.com>
Date: Wed, 28 Aug 2024 11:18:27 +0100
Subject: [PATCH] [MachineOutliner] Preserve instruction bundles

When the machine outliner copies instructions from a source function
into an outlined function, it was doing it using `CloneMachineInstr`,
which is documented as not preserving the interior of any instruction
bundle. So outlining code that includes an instruction bundle would
fail, because in the outlined version, the bundle would be empty, so
instructions would go missing in the move.

This occurs when any bundled instruction appears in the outlined code,
so there was no need to construct an unusual test case: I've just
copied a function from the existing `stp-opt-with-renaming.mir`, which
happens to contain an SVE instruction bundle. Including two identical
copies of that function makes the outliner merge them, and then we
check that it didn't destroy the interior of the bundle in the process.
---
 llvm/lib/CodeGen/MachineOutliner.cpp          |  7 ++-
 .../AArch64/machine-outliner-bundle.mir       | 54 +++++++++++++++++++
 2 files changed, 57 insertions(+), 4 deletions(-)
 create mode 100644 llvm/test/CodeGen/AArch64/machine-outliner-bundle.mir

diff --git a/llvm/lib/CodeGen/MachineOutliner.cpp b/llvm/lib/CodeGen/MachineOutliner.cpp
index 42f410c277179b..9678534cffce09 100644
--- a/llvm/lib/CodeGen/MachineOutliner.cpp
+++ b/llvm/lib/CodeGen/MachineOutliner.cpp
@@ -763,10 +763,9 @@ MachineFunction *MachineOutliner::createOutlinedFunction(
       BuildMI(MBB, MBB.end(), DL, TII.get(TargetOpcode::CFI_INSTRUCTION))
           .addCFIIndex(MF.addFrameInst(CFI));
     } else {
-      MachineInstr *NewMI = MF.CloneMachineInstr(&MI);
-      NewMI->dropMemRefs(MF);
-      NewMI->setDebugLoc(DL);
-      MBB.insert(MBB.end(), NewMI);
+      MachineInstr &NewMI = MF.cloneMachineInstrBundle(MBB, MBB.end(), MI);
+      NewMI.dropMemRefs(MF);
+      NewMI.setDebugLoc(DL);
     }
   }
 
diff --git a/llvm/test/CodeGen/AArch64/machine-outliner-bundle.mir b/llvm/test/CodeGen/AArch64/machine-outliner-bundle.mir
new file mode 100644
index 00000000000000..1dd5b0811bdfb6
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/machine-outliner-bundle.mir
@@ -0,0 +1,54 @@
+# RUN: llc -mtriple=aarch64 -run-pass=machine-outliner \
+# RUN: -verify-machineinstrs %s -o - | FileCheck %s
+
+# CHECK: name: OUTLINED_FUNCTION_0
+# CHECK-NOT: name:
+# CHECK: BUNDLE implicit-def $z3, implicit-def $q3, implicit-def $d3, implicit-def $s3, implicit-def $h3, implicit-def $b3, implicit $z19, implicit $p0, implicit $z16 {
+# CHECK:     $z3 = MOVPRFX_ZZ $z19
+# CHECK:     $z3 = FMUL_ZPmZ_S renamable $p0, killed $z3, renamable $z16
+# CHECK: }
+
+---
+name:            bundled
+alignment:       4
+tracksRegLiveness: true
+frameInfo:
+  maxAlignment:    1
+  maxCallFrameSize: 0
+machineFunctionInfo:
+  hasRedZone:      false
+body:             |
+  bb.0.entry:
+    liveins: $z3, $z19, $p0, $z16
+    renamable $q0 = LDRQui $sp, 1 :: (load 16)
+    STRSui renamable $s0, $sp, 9, implicit killed $q0 :: (store (s32))
+    BUNDLE implicit-def $z3, implicit-def $q3, implicit-def $d3, implicit-def $s3, implicit-def $h3, implicit-def $b3, implicit $z19, implicit $p0, implicit $z16 {
+      $z3 = MOVPRFX_ZZ $z19
+      $z3 = FMUL_ZPmZ_S renamable $p0, killed $z3, renamable $z16
+    }
+    renamable $q0 = LDRQui $sp, 0 :: (load 16, align 32)
+    STRSui renamable $s0, $sp, 10, implicit killed $q0 :: (store (s32))
+    RET undef $lr
+...
+---
+name:            bundled_clone
+alignment:       4
+tracksRegLiveness: true
+frameInfo:
+  maxAlignment:    1
+  maxCallFrameSize: 0
+machineFunctionInfo:
+  hasRedZone:      false
+body:             |
+  bb.0.entry:
+    liveins: $z3, $z19, $p0, $z16
+    renamable $q0 = LDRQui $sp, 1 :: (load 16)
+    STRSui renamable $s0, $sp, 9, implicit killed $q0 :: (store (s32))
+    BUNDLE implicit-def $z3, implicit-def $q3, implicit-def $d3, implicit-def $s3, implicit-def $h3, implicit-def $b3, implicit $z19, implicit $p0, implicit $z16 {
+      $z3 = MOVPRFX_ZZ $z19
+      $z3 = FMUL_ZPmZ_S renamable $p0, killed $z3, renamable $z16
+    }
+    renamable $q0 = LDRQui $sp, 0 :: (load 16, align 32)
+    STRSui renamable $s0, $sp, 10, implicit killed $q0 :: (store (s32))
+    RET undef $lr
+...