[ARM] Don't block tail-predication from unrelated VPT blocks.

[davemgreen]

VPT blocks that do not produce an interesting 'output' (like a stored value or reduction result), do not need to be predicated on vctp for the whole loop to be tail-predicated. Just producing results for the valid tail predication lanes should be enough.

[llvmbot]

@llvm/pr-subscribers-backend-arm

Author: David Green (davemgreen)

Changes

VPT blocks that do not produce an interesting 'output' (like a stored value or reduction result), do not need to be predicated on vctp for the whole loop to be tail-predicated. Just producing results for the valid tail predication lanes should be enough.

---

Full diff: https://github.com/llvm/llvm-project/pull/94239.diff

2 Files Affected:

• (modified) llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp (+16-6)
• (modified) llvm/test/CodeGen/Thumb2/mve-tailpred-vptblock.ll (+5-20)

diff --git a/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp b/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
index a46c383115e2d..d7c89ca6c6a57 100644
--- a/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
+++ b/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
@@ -115,6 +115,12 @@ static bool shouldInspect(MachineInstr &MI) {
   return isDomainMVE(&MI) || isVectorPredicate(&MI) || hasVPRUse(MI);
 }
 
+static bool isHorizontalReduction(const MachineInstr &MI) {
+  const MCInstrDesc &MCID = MI.getDesc();
+  uint64_t Flags = MCID.TSFlags;
+  return (Flags & ARMII::HorizontalReduction) != 0;
+}
+
 namespace {
 
   using InstSet = SmallPtrSetImpl<MachineInstr *>;
@@ -275,6 +281,16 @@ namespace {
       if (VPT->getOpcode() == ARM::MVE_VPST)
         return false;
 
+      // If the VPT block does not define something that is an "output", then the
+      // tail-predicated version will just perform a subset of the original vpt
+      // block, where the last lanes should not be used.
+      if (isVPTOpcode(VPT->getOpcode()) &&
+          all_of(Block.getInsts(), [](const MachineInstr *MI) {
+            return !MI->mayStore() && !isHorizontalReduction(*MI) &&
+                   !isVCTP(MI);
+          }))
+        return true;
+
       auto IsOperandPredicated = [&](MachineInstr *MI, unsigned Idx) {
         MachineInstr *Op = RDA.getMIOperand(MI, MI->getOperand(Idx));
         return Op && PredicatedInsts.count(Op) && isPredicatedOnVCTP(Op);
@@ -813,12 +829,6 @@ static bool producesDoubleWidthResult(const MachineInstr &MI) {
   return (Flags & ARMII::DoubleWidthResult) != 0;
 }
 
-static bool isHorizontalReduction(const MachineInstr &MI) {
-  const MCInstrDesc &MCID = MI.getDesc();
-  uint64_t Flags = MCID.TSFlags;
-  return (Flags & ARMII::HorizontalReduction) != 0;
-}
-
 // Can this instruction generate a non-zero result when given only zeroed
 // operands? This allows us to know that, given operands with false bytes
 // zeroed by masked loads, that the result will also contain zeros in those
diff --git a/llvm/test/CodeGen/Thumb2/mve-tailpred-vptblock.ll b/llvm/test/CodeGen/Thumb2/mve-tailpred-vptblock.ll
index f9b3757bb6d2c..6392452e06cdd 100644
--- a/llvm/test/CodeGen/Thumb2/mve-tailpred-vptblock.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-tailpred-vptblock.ll
@@ -20,50 +20,35 @@ define void @convert_vptblock(ptr %pchTarget, i16 signext %iTargetStride, ptr %p
 ; CHECK-NEXT:    mov.w r8, #0
 ; CHECK-NEXT:    ldrd r4, r5, [sp, #88]
 ; CHECK-NEXT:    mov r7, r0
-; CHECK-NEXT:    cmp.w r10, #8
-; CHECK-NEXT:    mov.w r0, #1
-; CHECK-NEXT:    mov r3, r10
 ; CHECK-NEXT:    mov.w r11, #0
-; CHECK-NEXT:    it ge
-; CHECK-NEXT:    movge r3, #8
 ; CHECK-NEXT:    vidup.u16 q0, r8, #4
-; CHECK-NEXT:    sub.w r3, r10, r3
 ; CHECK-NEXT:    vmov.i32 q1, #0x0
-; CHECK-NEXT:    adds r3, #7
 ; CHECK-NEXT:    vmov.i16 q2, #0x100
 ; CHECK-NEXT:    vmov.i16 q3, #0xff
-; CHECK-NEXT:    add.w r9, r0, r3, lsr #3
 ; CHECK-NEXT:  .LBB0_2: @ %for.body
 ; CHECK-NEXT:    @ =>This Loop Header: Depth=1
 ; CHECK-NEXT:    @ Child Loop BB0_3 Depth 2
-; CHECK-NEXT:    mov r3, r10
 ; CHECK-NEXT:    vmov q4, q0
 ; CHECK-NEXT:    mov r6, r8
 ; CHECK-NEXT:    mov r0, r7
-; CHECK-NEXT:    dls lr, r9
+; CHECK-NEXT:    dlstp.16 lr, r10
 ; CHECK-NEXT:  .LBB0_3: @ %do.body
 ; CHECK-NEXT:    @ Parent Loop BB0_2 Depth=1
 ; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
-; CHECK-NEXT:    vctp.16 r3
-; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vldrbt.u16 q5, [r2, q4]
+; CHECK-NEXT:    vldrb.u16 q5, [r2, q4]
 ; CHECK-NEXT:    vmul.i16 q4, q5, r5
 ; CHECK-NEXT:    vshr.u16 q4, q4, #8
 ; CHECK-NEXT:    vsub.i16 q5, q2, q4
 ; CHECK-NEXT:    vpt.i16 eq, q4, q3
 ; CHECK-NEXT:    vmovt q5, q1
-; CHECK-NEXT:    vctp.16 r3
-; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vldrbt.u16 q6, [r0]
+; CHECK-NEXT:    vldrb.u16 q6, [r0]
 ; CHECK-NEXT:    vsub.i16 q4, q2, q5
-; CHECK-NEXT:    subs r3, #8
 ; CHECK-NEXT:    vmul.i16 q5, q5, q6
 ; CHECK-NEXT:    vmla.i16 q5, q4, r4
 ; CHECK-NEXT:    vshr.u16 q4, q5, #8
-; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vstrbt.16 q4, [r0], #8
+; CHECK-NEXT:    vstrb.16 q4, [r0], #8
 ; CHECK-NEXT:    vidup.u16 q4, r6, #4
-; CHECK-NEXT:    le lr, .LBB0_3
+; CHECK-NEXT:    letp lr, .LBB0_3
 ; CHECK-NEXT:  @ %bb.4: @ %do.end
 ; CHECK-NEXT:    @ in Loop: Header=BB0_2 Depth=1
 ; CHECK-NEXT:    add.w r0, r11, #1

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[sparker-arm]

Okay, cheers! LGTM