[CodeGen] Let RDA recompute live-ins.

[mgudim]

We plan to use ReachingDefinitionsAnalysis in
PrologEpilogInsertion. In some tests this will be a problem because ReachingDefinitionAnalysis will have to run on a function that does not have live-in info. This PR lets ReachingDefinitionsAnalysis recompute live-ins.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Mikhail Gudim (mgudim)

Changes

We plan to use ReachingDefinitionsAnalysis in
PrologEpilogInsertion. In some tests this will be a problem because ReachingDefinitionAnalysis will have to run on a function that does not have live-in info. This PR lets ReachingDefinitionsAnalysis recompute live-ins.

---

Full diff: https://github.com/llvm/llvm-project/pull/166773.diff

3 Files Affected:

• (modified) llvm/lib/CodeGen/ReachingDefAnalysis.cpp (+12)
• (modified) llvm/test/CodeGen/RISCV/pr53662.mir (+4)
• (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-emergency-slot.mir (+1-1)

diff --git a/llvm/lib/CodeGen/ReachingDefAnalysis.cpp b/llvm/lib/CodeGen/ReachingDefAnalysis.cpp
index 40a89078bcf59..b4b7e0b64ab4e 100644
--- a/llvm/lib/CodeGen/ReachingDefAnalysis.cpp
+++ b/llvm/lib/CodeGen/ReachingDefAnalysis.cpp
@@ -7,8 +7,10 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/ReachingDefAnalysis.h"
+#include "llvm/ADT/PostOrderIterator.h"
 #include "llvm/ADT/SetOperations.h"
 #include "llvm/ADT/SmallSet.h"
+#include "llvm/CodeGen/LivePhysRegs.h"
 #include "llvm/CodeGen/LiveRegUnits.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
@@ -287,6 +289,16 @@ void ReachingDefInfo::run(MachineFunction &mf) {
   TRI = STI.getRegisterInfo();
   TII = STI.getInstrInfo();
   LLVM_DEBUG(dbgs() << "********** REACHING DEFINITION ANALYSIS **********\n");
+
+  MachineFunctionProperties &Props = MF->getProperties();
+  if (!Props.hasTracksLiveness()) {
+    Props.setTracksLiveness();
+
+    SmallVector<MachineBasicBlock *> AllMBBsInPostOrder;
+    for (MachineBasicBlock *MBB : post_order(MF))
+      AllMBBsInPostOrder.push_back(MBB);
+    fullyRecomputeLiveIns(AllMBBsInPostOrder);
+  }
   init();
   traverse();
 }
diff --git a/llvm/test/CodeGen/RISCV/pr53662.mir b/llvm/test/CodeGen/RISCV/pr53662.mir
index dccad40368111..834bcbc1cf82c 100644
--- a/llvm/test/CodeGen/RISCV/pr53662.mir
+++ b/llvm/test/CodeGen/RISCV/pr53662.mir
@@ -18,15 +18,19 @@ body:             |
   ; CHECK-LABEL: name: b
   ; CHECK: bb.0:
   ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT:   liveins: $x10
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   PseudoBR %bb.1
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.1:
   ; CHECK-NEXT:   successors: %bb.2(0x80000000)
+  ; CHECK-NEXT:   liveins: $x10
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   DBG_VALUE $noreg
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   liveins: $x10
+  ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   PseudoRET implicit killed $x10
   bb.0 :
     PseudoBR %bb.1
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-emergency-slot.mir b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-emergency-slot.mir
index c728fcb8d8b0d..44f60a43a2790 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-emergency-slot.mir
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-emergency-slot.mir
@@ -47,7 +47,7 @@ body:             |
 
     SD $x10, %stack.0, 0
     SD $x10, %stack.2, 0
-    dead renamable $x15 = PseudoVSETIVLI 1, 72, implicit-def $vl, implicit-def $vtype
+    renamable $x15 = PseudoVSETIVLI 1, 72, implicit-def $vl, implicit-def $vtype
     VS1R_V killed renamable $v25, %stack.1 :: (store (<vscale x 1 x s64>) into %stack.1, align 8)
     ; This is here just to make all the eligible registers live at this point.
     ; This way when we replace the frame index %stack.1 with its actual address

[mgudim]

I removed dead here because it was wrong - $x15 is used later.

[mgudim]

I am not sure that this is a proper way to solve the problem. My concern is that RDA is an analysis pass, but it may modify MIR by adding live-ins for every block. I suspect it could be a problem.

Please suggest other approaches if you have any ideas!

[lenary]

IIRC, RDA doesn't track function live-ins particularly well. ReachingDefAnalysis::getGlobalReachingDefs(...) doesn't model function live-ins at all, so you end up needing to check live-in-ness of a register separately.

I've worked around this in the past, but this issue would give me pause for using RDA in PEI, before the issue is resolved.