[llvm] e351b9b - [EarlyIfCvt] Don't if-convert if condition has only loop-invariant ops.

Florian Hahn via llvm-commits llvm-commits at lists.llvm.org
Fri May 12 11:21:26 PDT 2023


Author: Florian Hahn
Date: 2023-05-12T19:21:03+01:00
New Revision: e351b9b66da088e5905ad3238ccf11e777943ead

URL: https://github.com/llvm/llvm-project/commit/e351b9b66da088e5905ad3238ccf11e777943ead
DIFF: https://github.com/llvm/llvm-project/commit/e351b9b66da088e5905ad3238ccf11e777943ead.diff

LOG: [EarlyIfCvt] Don't if-convert if condition has only loop-invariant ops.

This patch adds a heuristic to skip if-conversion if the condition has a
high chance of being predictable.

If the condition is in a loop, consider it predictable if the condition
itself or all its operands are loop-invariant. E.g. this considers a load
from a loop-invariant address predictable; we were unable to prove that it
doesn't alias any of the memory-writes in the loop, but it is likely to
read to same value multiple times.

This is a relatively crude heuristic, but it helps to prevent excessive
if-conversion in multiple workloads in practice.

Reviewed By: apostolakis

Differential Revision: https://reviews.llvm.org/D141639

Added: 
    

Modified: 
    llvm/lib/CodeGen/EarlyIfConversion.cpp
    llvm/test/CodeGen/AArch64/early-ifcvt-likely-predictable.mir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/EarlyIfConversion.cpp b/llvm/lib/CodeGen/EarlyIfConversion.cpp
index 86c3d96b1bf0..30e26d40e229 100644
--- a/llvm/lib/CodeGen/EarlyIfConversion.cpp
+++ b/llvm/lib/CodeGen/EarlyIfConversion.cpp
@@ -119,10 +119,10 @@ class SSAIfConv {
 
   SmallVector<PHIInfo, 8> PHIs;
 
-private:
   /// The branch condition determined by analyzeBranch.
   SmallVector<MachineOperand, 4> Cond;
 
+private:
   /// Instructions in Head that define values used by the conditional blocks.
   /// The hoisted instructions must be inserted after these instructions.
   SmallPtrSet<MachineInstr*, 8> InsertAfter;
@@ -876,6 +876,38 @@ bool EarlyIfConverter::shouldConvertIf() {
   if (Stress)
     return true;
 
+  // Do not try to if-convert if the condition has a high chance of being
+  // predictable.
+  MachineLoop *CurrentLoop = Loops->getLoopFor(IfConv.Head);
+  // If the condition is in a loop, consider it predictable if the condition
+  // itself or all its operands are loop-invariant. E.g. this considers a load
+  // from a loop-invariant address predictable; we were unable to prove that it
+  // doesn't alias any of the memory-writes in the loop, but it is likely to
+  // read to same value multiple times.
+  if (CurrentLoop && any_of(IfConv.Cond, [&](MachineOperand &MO) {
+        if (!MO.isReg() || !MO.isUse())
+          return false;
+        Register Reg = MO.getReg();
+        if (Register::isPhysicalRegister(Reg))
+          return false;
+
+        MachineInstr *Def = MRI->getVRegDef(Reg);
+        return CurrentLoop->isLoopInvariant(*Def) ||
+               all_of(Def->operands(), [&](MachineOperand &Op) {
+                 if (Op.isImm())
+                   return true;
+                 if (!MO.isReg() || !MO.isUse())
+                   return false;
+                 Register Reg = MO.getReg();
+                 if (Register::isPhysicalRegister(Reg))
+                   return false;
+
+                 MachineInstr *Def = MRI->getVRegDef(Reg);
+                 return CurrentLoop->isLoopInvariant(*Def);
+               });
+      }))
+    return false;
+
   if (!MinInstr)
     MinInstr = Traces->getEnsemble(MachineTraceStrategy::TS_MinInstrCount);
 

diff  --git a/llvm/test/CodeGen/AArch64/early-ifcvt-likely-predictable.mir b/llvm/test/CodeGen/AArch64/early-ifcvt-likely-predictable.mir
index 2c3670ab236b..425a23214871 100644
--- a/llvm/test/CodeGen/AArch64/early-ifcvt-likely-predictable.mir
+++ b/llvm/test/CodeGen/AArch64/early-ifcvt-likely-predictable.mir
@@ -33,27 +33,35 @@ body:             |
   ; CHECK-NEXT:   [[COPY3:%[0-9]+]]:gpr64common = COPY $x0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.1:
-  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT:   successors: %bb.3(0x30000000), %bb.2(0x50000000)
   ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[LDRBBui:%[0-9]+]]:gpr32common = LDRBBui [[COPY3]], 0 :: (load (s8))
+  ; CHECK-NEXT:   [[LDRBBui:%[0-9]+]]:gpr32 = LDRBBui [[COPY3]], 0 :: (load (s8))
   ; CHECK-NEXT:   [[COPY4:%[0-9]+]]:gpr32all = COPY $wzr
-  ; CHECK-NEXT:   [[COPY5:%[0-9]+]]:gpr32 = COPY [[COPY4]]
+  ; CHECK-NEXT:   [[COPY5:%[0-9]+]]:gpr32all = COPY [[COPY4]]
+  ; CHECK-NEXT:   CBZW killed [[LDRBBui]], %bb.3
+  ; CHECK-NEXT:   B %bb.2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[SUBSWri:%[0-9]+]]:gpr32 = SUBSWri [[COPY1]], 4080, 12, implicit-def $nzcv
   ; CHECK-NEXT:   [[MOVi32imm:%[0-9]+]]:gpr32 = MOVi32imm 16711680
   ; CHECK-NEXT:   [[CSELWr:%[0-9]+]]:gpr32common = CSELWr [[COPY1]], killed [[MOVi32imm]], 11, implicit $nzcv
   ; CHECK-NEXT:   [[SUBSWri1:%[0-9]+]]:gpr32 = SUBSWri [[CSELWr]], 0, 0, implicit-def $nzcv
   ; CHECK-NEXT:   [[COPY6:%[0-9]+]]:gpr32 = COPY $wzr
   ; CHECK-NEXT:   [[CSELWr1:%[0-9]+]]:gpr32 = CSELWr [[CSELWr]], [[COPY6]], 12, implicit $nzcv
-  ; CHECK-NEXT:   [[COPY7:%[0-9]+]]:gpr32 = COPY [[CSELWr1]]
+  ; CHECK-NEXT:   [[COPY7:%[0-9]+]]:gpr32all = COPY [[CSELWr1]]
   ; CHECK-NEXT:   [[SUBSWri2:%[0-9]+]]:gpr32 = SUBSWri [[COPY1]], 0, 0, implicit-def $nzcv
   ; CHECK-NEXT:   [[CSELWr2:%[0-9]+]]:gpr32 = CSELWr [[COPY1]], [[COPY6]], 12, implicit $nzcv
-  ; CHECK-NEXT:   [[COPY8:%[0-9]+]]:gpr32 = COPY [[CSELWr2]]
-  ; CHECK-NEXT:   $wzr = SUBSWri [[LDRBBui]], 0, 0, implicit-def $nzcv
-  ; CHECK-NEXT:   [[CSELWr3:%[0-9]+]]:gpr32 = CSELWr [[COPY5]], [[COPY7]], 0, implicit $nzcv
-  ; CHECK-NEXT:   $wzr = SUBSWri [[LDRBBui]], 0, 0, implicit-def $nzcv
-  ; CHECK-NEXT:   [[CSELWr4:%[0-9]+]]:gpr32 = CSELWr [[COPY5]], [[COPY8]], 0, implicit $nzcv
-  ; CHECK-NEXT:   STRBBui [[CSELWr4]], [[COPY2]], 0 :: (store (s8))
-  ; CHECK-NEXT:   STRBBui [[CSELWr3]], [[COPY]], 0 :: (store (s8))
+  ; CHECK-NEXT:   [[COPY8:%[0-9]+]]:gpr32all = COPY [[CSELWr2]]
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.3:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:gpr32 = PHI [[COPY5]], %bb.1, [[COPY7]], %bb.2
+  ; CHECK-NEXT:   [[PHI1:%[0-9]+]]:gpr32 = PHI [[COPY5]], %bb.1, [[COPY8]], %bb.2
+  ; CHECK-NEXT:   STRBBui [[PHI1]], [[COPY2]], 0 :: (store (s8))
+  ; CHECK-NEXT:   STRBBui [[PHI]], [[COPY]], 0 :: (store (s8))
   ; CHECK-NEXT:   B %bb.1
   bb.0:
     liveins: $x0, $x1, $w2, $x3
@@ -108,17 +116,26 @@ body:             |
   ; CHECK-NEXT:   [[COPY3:%[0-9]+]]:gpr64common = COPY $x0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.1:
-  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT:   successors: %bb.3(0x30000000), %bb.2(0x50000000)
   ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[LDRBBui:%[0-9]+]]:gpr32common = LDRBBui [[COPY3]], 0 :: (load (s8))
+  ; CHECK-NEXT:   [[LDRBBui:%[0-9]+]]:gpr32 = LDRBBui [[COPY3]], 0 :: (load (s8))
   ; CHECK-NEXT:   [[COPY4:%[0-9]+]]:gpr32all = COPY $wzr
   ; CHECK-NEXT:   [[COPY5:%[0-9]+]]:gpr32all = COPY [[COPY4]]
-  ; CHECK-NEXT:   [[COPY6:%[0-9]+]]:gpr32 = COPY [[LDRBBui]]
+  ; CHECK-NEXT:   [[COPY6:%[0-9]+]]:gpr32all = COPY [[LDRBBui]]
+  ; CHECK-NEXT:   CBZW killed [[LDRBBui]], %bb.3
+  ; CHECK-NEXT:   B %bb.2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
+  ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[SUBSWri:%[0-9]+]]:gpr32 = SUBSWri [[COPY1]], 4080, 12, implicit-def $nzcv
-  ; CHECK-NEXT:   [[COPY7:%[0-9]+]]:gpr32 = COPY [[SUBSWri]]
-  ; CHECK-NEXT:   $wzr = SUBSWri [[LDRBBui]], 0, 0, implicit-def $nzcv
-  ; CHECK-NEXT:   [[CSELWr:%[0-9]+]]:gpr32 = CSELWr [[COPY6]], [[COPY7]], 0, implicit $nzcv
-  ; CHECK-NEXT:   STRBBui [[CSELWr]], [[COPY]], 0 :: (store (s8))
+  ; CHECK-NEXT:   [[COPY7:%[0-9]+]]:gpr32all = COPY [[SUBSWri]]
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.3:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:gpr32 = PHI [[COPY6]], %bb.1, [[COPY7]], %bb.2
+  ; CHECK-NEXT:   STRBBui [[PHI]], [[COPY]], 0 :: (store (s8))
   ; CHECK-NEXT:   B %bb.1
   bb.0:
     liveins: $x0, $x1, $w2, $x3


        


More information about the llvm-commits mailing list