[llvm-branch-commits] [llvm] PR for llvm/llvm-project#79800 (PR #79813)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Jan 29 03:52:04 PST 2024


https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/79813

resolves llvm/llvm-project#79800

>From 0f9256b72bcda1b0444bd302aa22ede428c73a54 Mon Sep 17 00:00:00 2001
From: David Sherwood <57997763+david-arm at users.noreply.github.com>
Date: Fri, 26 Jan 2024 14:43:48 +0000
Subject: [PATCH] [LoopVectorize] Refine runtime memory check costs when there
 is an outer loop (#76034)

When we generate runtime memory checks for an inner loop it's
possible that these checks are invariant in the outer loop and
so will get hoisted out. In such cases, the effective cost of
the checks should reduce to reflect the outer loop trip count.

This fixes a 25% performance regression introduced by commit

49b0e6dcc296792b577ae8f0f674e61a0929b99d

when building the SPEC2017 x264 benchmark with PGO, where we
decided the inner loop trip count wasn't high enough to warrant
the (incorrect) high cost of the runtime checks. Also, when
runtime memory checks consist entirely of diff checks these are
likely to be outer loop invariant.

(cherry picked from commit 962fbafecf4730ba84a3b9fd7a662a5c30bb2c7c)
---
 .../Transforms/Vectorize/LoopVectorize.cpp    |  62 ++++-
 .../AArch64/low_trip_memcheck_cost.ll         | 217 ++++++++++++++++++
 2 files changed, 273 insertions(+), 6 deletions(-)
 create mode 100644 llvm/test/Transforms/LoopVectorize/AArch64/low_trip_memcheck_cost.ll

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 6ca93e15719fb27..dd596c567cd4824 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -1957,6 +1957,8 @@ class GeneratedRTChecks {
   bool CostTooHigh = false;
   const bool AddBranchWeights;
 
+  Loop *OuterLoop = nullptr;
+
 public:
   GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI,
                     TargetTransformInfo *TTI, const DataLayout &DL,
@@ -2053,6 +2055,9 @@ class GeneratedRTChecks {
       DT->eraseNode(SCEVCheckBlock);
       LI->removeBlock(SCEVCheckBlock);
     }
+
+    // Outer loop is used as part of the later cost calculations.
+    OuterLoop = L->getParentLoop();
   }
 
   InstructionCost getCost() {
@@ -2076,16 +2081,61 @@ class GeneratedRTChecks {
         LLVM_DEBUG(dbgs() << "  " << C << "  for " << I << "\n");
         RTCheckCost += C;
       }
-    if (MemCheckBlock)
+    if (MemCheckBlock) {
+      InstructionCost MemCheckCost = 0;
       for (Instruction &I : *MemCheckBlock) {
         if (MemCheckBlock->getTerminator() == &I)
           continue;
         InstructionCost C =
             TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput);
         LLVM_DEBUG(dbgs() << "  " << C << "  for " << I << "\n");
-        RTCheckCost += C;
+        MemCheckCost += C;
       }
 
+      // If the runtime memory checks are being created inside an outer loop
+      // we should find out if these checks are outer loop invariant. If so,
+      // the checks will likely be hoisted out and so the effective cost will
+      // reduce according to the outer loop trip count.
+      if (OuterLoop) {
+        ScalarEvolution *SE = MemCheckExp.getSE();
+        // TODO: If profitable, we could refine this further by analysing every
+        // individual memory check, since there could be a mixture of loop
+        // variant and invariant checks that mean the final condition is
+        // variant.
+        const SCEV *Cond = SE->getSCEV(MemRuntimeCheckCond);
+        if (SE->isLoopInvariant(Cond, OuterLoop)) {
+          // It seems reasonable to assume that we can reduce the effective
+          // cost of the checks even when we know nothing about the trip
+          // count. Assume that the outer loop executes at least twice.
+          unsigned BestTripCount = 2;
+
+          // If exact trip count is known use that.
+          if (unsigned SmallTC = SE->getSmallConstantTripCount(OuterLoop))
+            BestTripCount = SmallTC;
+          else if (LoopVectorizeWithBlockFrequency) {
+            // Else use profile data if available.
+            if (auto EstimatedTC = getLoopEstimatedTripCount(OuterLoop))
+              BestTripCount = *EstimatedTC;
+          }
+
+          InstructionCost NewMemCheckCost = MemCheckCost / BestTripCount;
+
+          // Let's ensure the cost is always at least 1.
+          NewMemCheckCost = std::max(*NewMemCheckCost.getValue(),
+                                     (InstructionCost::CostType)1);
+
+          LLVM_DEBUG(dbgs()
+                     << "We expect runtime memory checks to be hoisted "
+                     << "out of the outer loop. Cost reduced from "
+                     << MemCheckCost << " to " << NewMemCheckCost << '\n');
+
+          MemCheckCost = NewMemCheckCost;
+        }
+      }
+
+      RTCheckCost += MemCheckCost;
+    }
+
     if (SCEVCheckBlock || MemCheckBlock)
       LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCost
                         << "\n");
@@ -2144,8 +2194,8 @@ class GeneratedRTChecks {
 
     BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock);
     // Create new preheader for vector loop.
-    if (auto *PL = LI->getLoopFor(LoopVectorPreHeader))
-      PL->addBasicBlockToLoop(SCEVCheckBlock, *LI);
+    if (OuterLoop)
+      OuterLoop->addBasicBlockToLoop(SCEVCheckBlock, *LI);
 
     SCEVCheckBlock->getTerminator()->eraseFromParent();
     SCEVCheckBlock->moveBefore(LoopVectorPreHeader);
@@ -2179,8 +2229,8 @@ class GeneratedRTChecks {
     DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock);
     MemCheckBlock->moveBefore(LoopVectorPreHeader);
 
-    if (auto *PL = LI->getLoopFor(LoopVectorPreHeader))
-      PL->addBasicBlockToLoop(MemCheckBlock, *LI);
+    if (OuterLoop)
+      OuterLoop->addBasicBlockToLoop(MemCheckBlock, *LI);
 
     BranchInst &BI =
         *BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond);
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/low_trip_memcheck_cost.ll b/llvm/test/Transforms/LoopVectorize/AArch64/low_trip_memcheck_cost.ll
new file mode 100644
index 000000000000000..8a796bb3065b196
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/low_trip_memcheck_cost.ll
@@ -0,0 +1,217 @@
+; REQUIRES: asserts
+; RUN: opt -p loop-vectorize -debug-only=loop-vectorize -S -disable-output < %s 2>&1 | FileCheck %s
+
+target triple = "aarch64-unknown-linux-gnu"
+
+define void @no_outer_loop(ptr nocapture noundef %a, ptr nocapture noundef readonly %b, i64 noundef %off, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'no_outer_loop'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK-NOT:  We expect runtime memory checks to be hoisted out of the outer loop.
+; CHECK:      Total cost of runtime checks: 4
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:16
+entry:
+  br label %inner.loop
+
+inner.loop:
+  %inner.iv = phi i64 [ 0, %entry ], [ %inner.iv.next, %inner.loop ]
+  %add.us = add nuw nsw i64 %inner.iv, %off
+  %arrayidx.us = getelementptr inbounds i8, ptr %b, i64 %add.us
+  %0 = load i8, ptr %arrayidx.us, align 1
+  %arrayidx7.us = getelementptr inbounds i8, ptr %a, i64 %add.us
+  %1 = load i8, ptr %arrayidx7.us, align 1
+  %add9.us = add i8 %1, %0
+  store i8 %add9.us, ptr %arrayidx7.us, align 1
+  %inner.iv.next = add nuw nsw i64 %inner.iv, 1
+  %exitcond.not = icmp eq i64 %inner.iv.next, %n
+  br i1 %exitcond.not, label %inner.exit, label %inner.loop
+
+inner.exit:
+  ret void
+}
+
+define void @outer_no_tc(ptr nocapture noundef %a, ptr nocapture noundef readonly %b, i64 noundef %m, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'outer_no_tc'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK:      We expect runtime memory checks to be hoisted out of the outer loop. Cost reduced from 6 to 3
+; CHECK:      Total cost of runtime checks: 3
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:16
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %outer.iv = phi i64 [ %outer.iv.next, %inner.exit ], [ 0, %entry ]
+  %mul.us = mul nsw i64 %outer.iv, %n
+  br label %inner.loop
+
+inner.loop:
+  %inner.iv = phi i64 [ 0, %outer.loop ], [ %inner.iv.next, %inner.loop ]
+  %add.us = add nuw nsw i64 %inner.iv, %mul.us
+  %arrayidx.us = getelementptr inbounds i8, ptr %b, i64 %add.us
+  %0 = load i8, ptr %arrayidx.us, align 1
+  %arrayidx7.us = getelementptr inbounds i8, ptr %a, i64 %add.us
+  %1 = load i8, ptr %arrayidx7.us, align 1
+  %add9.us = add i8 %1, %0
+  store i8 %add9.us, ptr %arrayidx7.us, align 1
+  %inner.iv.next = add nuw nsw i64 %inner.iv, 1
+  %exitcond.not = icmp eq i64 %inner.iv.next, %n
+  br i1 %exitcond.not, label %inner.exit, label %inner.loop
+
+inner.exit:
+  %outer.iv.next = add nuw nsw i64 %outer.iv, 1
+  %exitcond27.not = icmp eq i64 %outer.iv.next, %m
+  br i1 %exitcond27.not, label %outer.exit, label %outer.loop
+
+outer.exit:
+  ret void
+}
+
+
+define void @outer_known_tc3(ptr nocapture noundef %a, ptr nocapture noundef readonly %b, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'outer_known_tc3'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK:      We expect runtime memory checks to be hoisted out of the outer loop. Cost reduced from 6 to 2
+; CHECK:      Total cost of runtime checks: 2
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:16
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %outer.iv = phi i64 [ %outer.iv.next, %inner.exit ], [ 0, %entry ]
+  %mul.us = mul nsw i64 %outer.iv, %n
+  br label %inner.loop
+
+inner.loop:
+  %inner.iv = phi i64 [ 0, %outer.loop ], [ %inner.iv.next, %inner.loop ]
+  %add.us = add nuw nsw i64 %inner.iv, %mul.us
+  %arrayidx.us = getelementptr inbounds i8, ptr %b, i64 %add.us
+  %0 = load i8, ptr %arrayidx.us, align 1
+  %arrayidx7.us = getelementptr inbounds i8, ptr %a, i64 %add.us
+  %1 = load i8, ptr %arrayidx7.us, align 1
+  %add9.us = add i8 %1, %0
+  store i8 %add9.us, ptr %arrayidx7.us, align 1
+  %inner.iv.next = add nuw nsw i64 %inner.iv, 1
+  %exitcond.not = icmp eq i64 %inner.iv.next, %n
+  br i1 %exitcond.not, label %inner.exit, label %inner.loop
+
+inner.exit:
+  %outer.iv.next = add nuw nsw i64 %outer.iv, 1
+  %exitcond26.not = icmp eq i64 %outer.iv.next, 3
+  br i1 %exitcond26.not, label %outer.exit, label %outer.loop
+
+outer.exit:
+  ret void
+}
+
+
+define void @outer_known_tc64(ptr nocapture noundef %a, ptr nocapture noundef readonly %b, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'outer_known_tc64'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK:      We expect runtime memory checks to be hoisted out of the outer loop. Cost reduced from 6 to 1
+; CHECK:      Total cost of runtime checks: 1
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:16
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %outer.iv = phi i64 [ %outer.iv.next, %inner.exit ], [ 0, %entry ]
+  %mul.us = mul nsw i64 %outer.iv, %n
+  br label %inner.loop
+
+inner.loop:
+  %inner.iv = phi i64 [ 0, %outer.loop ], [ %inner.iv.next, %inner.loop ]
+  %add.us = add nuw nsw i64 %inner.iv, %mul.us
+  %arrayidx.us = getelementptr inbounds i8, ptr %b, i64 %add.us
+  %0 = load i8, ptr %arrayidx.us, align 1
+  %arrayidx7.us = getelementptr inbounds i8, ptr %a, i64 %add.us
+  %1 = load i8, ptr %arrayidx7.us, align 1
+  %add9.us = add i8 %1, %0
+  store i8 %add9.us, ptr %arrayidx7.us, align 1
+  %inner.iv.next = add nuw nsw i64 %inner.iv, 1
+  %exitcond.not = icmp eq i64 %inner.iv.next, %n
+  br i1 %exitcond.not, label %inner.exit, label %inner.loop
+
+inner.exit:
+  %outer.iv.next = add nuw nsw i64 %outer.iv, 1
+  %exitcond26.not = icmp eq i64 %outer.iv.next, 64
+  br i1 %exitcond26.not, label %outer.exit, label %outer.loop
+
+outer.exit:
+  ret void
+}
+
+
+define void @outer_pgo_3(ptr nocapture noundef %a, ptr nocapture noundef readonly %b, i64 noundef %m, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'outer_pgo_3'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK:      We expect runtime memory checks to be hoisted out of the outer loop. Cost reduced from 6 to 2
+; CHECK:      Total cost of runtime checks: 2
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:16
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %outer.iv = phi i64 [ %outer.iv.next, %inner.exit ], [ 0, %entry ]
+  %mul.us = mul nsw i64 %outer.iv, %n
+  br label %inner.loop
+
+inner.loop:
+  %inner.iv = phi i64 [ 0, %outer.loop ], [ %inner.iv.next, %inner.loop ]
+  %add.us = add nuw nsw i64 %inner.iv, %mul.us
+  %arrayidx.us = getelementptr inbounds i8, ptr %b, i64 %add.us
+  %0 = load i8, ptr %arrayidx.us, align 1
+  %arrayidx7.us = getelementptr inbounds i8, ptr %a, i64 %add.us
+  %1 = load i8, ptr %arrayidx7.us, align 1
+  %add9.us = add i8 %1, %0
+  store i8 %add9.us, ptr %arrayidx7.us, align 1
+  %inner.iv.next = add nuw nsw i64 %inner.iv, 1
+  %exitcond.not = icmp eq i64 %inner.iv.next, %n
+  br i1 %exitcond.not, label %inner.exit, label %inner.loop
+
+inner.exit:
+  %outer.iv.next = add nuw nsw i64 %outer.iv, 1
+  %exitcond26.not = icmp eq i64 %outer.iv.next, %m
+  br i1 %exitcond26.not, label %outer.exit, label %outer.loop, !prof !0
+
+outer.exit:
+  ret void
+}
+
+
+define void @outer_known_tc3_full_range_checks(ptr nocapture noundef %dst, ptr nocapture noundef readonly %src, i64 noundef %n) {
+; CHECK-LABEL: LV: Checking a loop in 'outer_known_tc3_full_range_checks'
+; CHECK:      Calculating cost of runtime checks:
+; CHECK:      We expect runtime memory checks to be hoisted out of the outer loop. Cost reduced from 6 to 2
+; CHECK:      Total cost of runtime checks: 2
+; CHECK-NEXT: LV: Minimum required TC for runtime checks to be profitable:4
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %outer.iv = phi i64 [ 0, %entry ], [ %outer.iv.next, %inner.exit ]
+  %0 = mul nsw i64 %outer.iv, %n
+  br label %inner.loop
+
+inner.loop:
+  %iv.inner = phi i64 [ 0, %outer.loop ], [ %iv.inner.next, %inner.loop ]
+  %1 = add nuw nsw i64 %iv.inner, %0
+  %arrayidx.us = getelementptr inbounds i32, ptr %src, i64 %1
+  %2 = load i32, ptr %arrayidx.us, align 4
+  %arrayidx8.us = getelementptr inbounds i32, ptr %dst, i64 %1
+  %3 = load i32, ptr %arrayidx8.us, align 4
+  %add9.us = add nsw i32 %3, %2
+  store i32 %add9.us, ptr %arrayidx8.us, align 4
+  %iv.inner.next = add nuw nsw i64 %iv.inner, 1
+  %inner.exit.cond = icmp eq i64 %iv.inner.next, %n
+  br i1 %inner.exit.cond, label %inner.exit, label %inner.loop
+
+inner.exit:
+  %outer.iv.next = add nuw nsw i64 %outer.iv, 1
+  %outer.exit.cond = icmp eq i64 %outer.iv.next, 3
+  br i1 %outer.exit.cond, label %outer.exit, label %outer.loop
+
+outer.exit:
+  ret void
+}
+
+
+!0 = !{!"branch_weights", i32 10, i32 20}



More information about the llvm-branch-commits mailing list