[PATCH] D158674: [CodeGen] Fix unconditional branch duplication issue in bbsections

[Daniel Hoekwater via Phabricator via llvm-commits]

dhoekwater created this revision.
dhoekwater added a reviewer: rahmanl.
Herald added subscribers: pengfei, hiraditya, kristof.beyls.
Herald added a project: All.
dhoekwater requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

If an end section basic block ends in an unconditional branch to its
fallthrough, BasicBlockSections will duplicate the unconditional branch.
This doesn't break x86, but it is a (slight) size optimization and more
importantly prevents AArch64 builds from breaking.

Ex:

  bb1 (bbsections Hot):
    jmp bb2
  
  bb2 (bbsections Cold):
    /* do work... */

After running sortBasicBlocksAndUpdateBranches():

  bb1 (bbsections Hot):
    jmp bb2
    jmp bb2
  
  bb2 (bbsections Cold):
    /* do work... */


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D158674

Files:
  llvm/lib/CodeGen/BasicBlockSections.cpp
  llvm/test/CodeGen/Generic/machine-function-splitter.ll


Index: llvm/test/CodeGen/Generic/machine-function-splitter.ll
===================================================================
--- llvm/test/CodeGen/Generic/machine-function-splitter.ll
+++ llvm/test/CodeGen/Generic/machine-function-splitter.ll
@@ -6,6 +6,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-psi-cutoff=0 -mfs-count-threshold=2000 | FileCheck %s --dump-input=always -check-prefixes=MFS-OPTS1,MFS-OPTS1-X86
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-psi-cutoff=950000 | FileCheck %s -check-prefixes=MFS-OPTS2,MFS-OPTS2-X86
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-split-ehcode | FileCheck %s -check-prefixes=MFS-EH-SPLIT,MFS-EH-SPLIT-X86
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -O0 -mfs-psi-cutoff=0 -mfs-count-threshold=4000 | FileCheck %s -check-prefixes=MFS-O0,MFS-O0-X86
 
 ; COM: Machine function splitting with AFDO profiles
 ; RUN: sed 's/InstrProf/SampleProfile/g' %s > %t.ll
@@ -447,6 +448,25 @@
   ret void
 }
 
+define void @foo16(i1 zeroext %0) nounwind !prof !14 !section_prefix !15 {
+;; Check that an unconditional branch is only appended to a block
+;; if it would fall through to the wrong block otherwise.
+; MFS-O0-LABEL:               foo16
+; MFS-O0-X86:                 jmp
+; MFS-O0-X86-NOT:             jmp
+; MFS-O0-X86:                 .{{LBB[a-zA-Z0-9_]*:}}
+  %2 = call i32 @baz()
+  br i1 false, label %3, label %5, !prof !25
+
+3:                                                ; preds = %1
+  %4 = call i32 @bar()
+  unreachable
+
+5:                                                ; preds = %1
+  %6 = tail call i32 @qux()
+  ret void
+}
+
 declare i32 @bar()
 declare i32 @baz()
 declare i32 @bam()
Index: llvm/lib/CodeGen/BasicBlockSections.cpp
===================================================================
--- llvm/lib/CodeGen/BasicBlockSections.cpp
+++ llvm/lib/CodeGen/BasicBlockSections.cpp
@@ -260,7 +260,8 @@
   [[maybe_unused]] const MachineBasicBlock *EntryBlock = &MF.front();
   SmallVector<MachineBasicBlock *> PreLayoutFallThroughs(MF.getNumBlockIDs());
   for (auto &MBB : MF)
-    PreLayoutFallThroughs[MBB.getNumber()] = MBB.getFallThrough();
+    PreLayoutFallThroughs[MBB.getNumber()] =
+        MBB.getFallThrough(/*JumpToFallThrough=*/false);
 
   MF.sort(MBBCmp);
   assert(&MF.front() == EntryBlock &&


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D158674.552892.patch
Type: text/x-patch
Size: 2429 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230823/8c09d403/attachment.bin>