[BranchFolding] Kill common hoisted debug instructions

branch-folder hoists common instructions from TBB and FBB into their
pred. Without this patch it achieves this by splicing the instructions from TBB
and deleting the common ones in FBB. That moves the debug locations and debug
instructions from TBB into the pred without modification, which is not
ideal. Debug locations are handled in #140063.

This patch handles debug instructions - in the simplest way possible, which is
to just kill (undef) them. We kill and hoist the ones in FBB as well as TBB
because otherwise the fact there's an assignment on the code path is deleted
(which might lead to a prior location extending further than it should).

There's possibly something we could do to preserve some variable locations in
some cases, but this is the easiest not-incorrect thing to do.

Note I had to replace the constant DBG_VALUEs to use registers in the test- it
turns out setDebugValueUndef doesn't undef constant DBG_VALUEs... which feels
wrong to me, but isn't something I want to touch right now.

Author: OCHyams

llvm/lib/CodeGen/BranchFolding.cpp

@@ -25,6 +25,7 @@
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/BranchFoldingPass.h"
+#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
 #include "llvm/CodeGen/MBFIWrapper.h"
 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
@@ -2072,25 +2073,56 @@ bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
     return false;
 
   // Hoist the instructions from [T.begin, TIB) and then delete [F.begin, FIB).
-  // Merge the debug locations. FIXME: We should do something with the
-  // debug instructions too (from BOTH branches).
+  // Merge the debug locations, and hoist and kill the debug instructions from
+  // both branches. FIXME: We could probably try harder to preserve some debug
+  // instructions (but at least this isn't producing wrong locations).
   {
+    MachineIRBuilder MIRBuilder(*MBB, Loc);
+    auto HoistAndKillDbgInstr =
+        [&MIRBuilder](MachineBasicBlock::iterator DI,
+                      MachineBasicBlock::iterator InsertBefore) {
+          assert(DI->isDebugInstr() && "Expected a debug instruction");
+          if (DI->isDebugRef()) {
+            MIRBuilder.setDebugLoc(DI->getDebugLoc());
+            MIRBuilder.buildDirectDbgValue(
+                0, DI->getDebugVariable(),
+                DI->getDebugExpression());
+            return;
+          }
+
+          DI->setDebugValueUndef();
+          DI->moveBefore(&*InsertBefore);
+        };
+
     // TIB and FIB point to the end of the regions to hoist/merge in TBB and
     // FBB.
     MachineBasicBlock::iterator FE = FIB;
     MachineBasicBlock::iterator FI = FBB->begin();
     for (MachineBasicBlock::iterator TI :
          make_early_inc_range(make_range(TBB->begin(), TIB))) {
-      // Move debug instructions and pseudo probes without modifying them.
-      // FIXME: This is the wrong thing to do for debug locations, which
-      // should at least be killed.
-      if (TI->isDebugOrPseudoInstr()) {
+
+      // Hoist and kill debug instructions from FBB. Skip over pseudo
+      // instructions. After this loop FI points to the next non-debug
+      // instruction to hoist (checked in assert after the TBB debug
+      // instruction handling code).
+      while (FI->isDebugOrPseudoInstr()) {
+        assert(FI != FE && "Unexpected end of FBB range");
+        MachineBasicBlock::iterator FINext = std::next(FI) ;
+        HoistAndKillDbgInstr(FI, Loc);
+        FI = FINext;
+      }
+
+      // Move pseudo probes without modifying them.
+      if (TI->isPseudoProbe()) {
         TI->moveBefore(&*Loc);
         continue;
       }
+      // Kill debug instructions before moving.
+      if (TI->isDebugInstr()) {
+        HoistAndKillDbgInstr(TI, Loc);
+        continue;
+      }
 
-      // Get the next non-meta instruction in FBB.
-      FI = skipDebugInstructionsForward(FI, FE, false);
       assert(TI->isIdenticalTo(*FI, MachineInstr::CheckKillDead) &&
              "Expected non-debug lockstep");
 
@@ -2101,6 +2133,7 @@ bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
       ++FI;
     }
   }
+  assert(FIB->getParent() == FBB && "Unexpectedly moved FIB?");
   FBB->erase(FBB->begin(), FIB);
 
   if (UpdateLiveIns)

llvm/test/DebugInfo/X86/branch-folder-dbg.mir

@@ -1,16 +1,21 @@
 # RUN: llc %s --start-before=branch-folder --stop-after=branch-folder -o - | FileCheck %s
 
 ## Check that common instructions hoisted from `if.then` and `if.else` into
-## common pred `entry` get merged debug locations.
-
-## FIXME: The debug instructions handling here is wrong.
+## common pred `entry` get merged debug locations. The debug instructions from
+## both branches should get hoisted and killed.
+##
+## The MIR debug instructions have been modified by hand in order to check they
+## can be killed.
 
 # CHECK: bb.0
 # CHECK:      CALL64pcrel32 @f, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
-## --- Start splice from bb.2.if.else ---
-# CHECK-NEXT: DBG_VALUE 2, $noreg, ![[#]], !DIExpression(),  debug-location ![[#]]
-# CHECK-NEXT: $edi = MOV32r0 implicit-def dead $eflags,  debug-location !DILocation(line: 0, scope: ![[#]])
-## --- End splice --------------
+## --- Start splice from bb.2.if.else (and debug instructions from bb.1.if.then) ---
+# CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[#]], !DIExpression(),  debug-location ![[#]]
+# CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[#]], !DIExpression(),  debug-location ![[#]]
+# CHECK-NEXT: $edi = MOV32r0 implicit-def dead $eflags, debug-instr-number 2, debug-location !DILocation(line: 0, scope: ![[#]])
+# CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_arg, 0),  debug-location ![[#]]
+# CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_arg, 0),  debug-location ![[#]]
+## --- End splice ------------------------------------------------------------------
 # CHECK-NEXT: TEST64rr killed renamable $rax, renamable $rax, implicit-def $eflags
 # CHECK-NEXT: JCC_1 %bb.2, 9, implicit killed $eflags
 # CHECK: bb.1
@@ -73,6 +78,8 @@
 ...
 ---
 name:            g
+tracksRegLiveness: true
+isSSA: false
 body:             |
   bb.0 (%ir-block.0):
     successors: %bb.1(0x40000000), %bb.2(0x40000000)
@@ -86,22 +93,21 @@ body:             |
 
   bb.1.if.then:
     successors: %bb.3(0x80000000)
-
-    DBG_VALUE 0, $noreg, !11, !DIExpression(),  debug-location !13
-    $edi = MOV32r0 implicit-def dead $eflags,  debug-location !14
+    DBG_VALUE $esi, $noreg, !11, !DIExpression(),  debug-location !13
+    $edi = MOV32r0 implicit-def dead $eflags, debug-instr-number 1, debug-location !14
+    DBG_INSTR_REF !11, !DIExpression(DW_OP_LLVM_arg, 0), dbg-instr-ref(1, 0),  debug-location !13
     $esi = MOV32r0 implicit-def dead $eflags,  debug-location !14
     CALL64pcrel32 target-flags(x86-plt) @_Z3fooii, csr_64, implicit $rsp, implicit $ssp, implicit killed $edi, implicit killed $esi, implicit-def $rsp, implicit-def $ssp,  debug-location !14
-    DBG_VALUE 1, $noreg, !11, !DIExpression(),  debug-location !13
     JMP_1 %bb.3,  debug-location !15
 
   bb.2.if.else:
     successors: %bb.3(0x80000000)
 
-    DBG_VALUE 2, $noreg, !11, !DIExpression(),  debug-location !13
-    $edi = MOV32r0 implicit-def dead $eflags,  debug-location !16
+    DBG_VALUE $esp, $noreg, !11, !DIExpression(), debug-location !13
+    $edi = MOV32r0 implicit-def dead $eflags, debug-instr-number 2, debug-location !16
+    DBG_INSTR_REF !11, !DIExpression(DW_OP_LLVM_arg, 0), dbg-instr-ref(2, 0),  debug-location !13
     $esi = MOV32ri 1,  debug-location !16
     CALL64pcrel32 target-flags(x86-plt) @_Z3barii, csr_64, implicit $rsp, implicit $ssp, implicit killed $edi, implicit killed $esi, implicit-def $rsp, implicit-def $ssp,  debug-location !16
-    DBG_VALUE 3, $noreg, !11, !DIExpression(),  debug-location !13
 
   bb.3.if.end:
     $eax = MOV32ri 2