[PATCH] D151870: [ARM] Increase cost of unaligned double and float loads

[Sam Tebbs via Phabricator via llvm-commits]

samtebbs created this revision.
samtebbs added reviewers: dmgreen, SjoerdMeijer.
Herald added subscribers: hiraditya, kristof.beyls.
Herald added a project: All.
samtebbs requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Unaligned double and float loads produce a lot of stack stores and loads that are not accounted for in the cost model yet. This patch adjusts the cost model to account for those extra operations.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D151870

Files:
  llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
  llvm/test/Analysis/CostModel/ARM/load_store.ll


Index: llvm/test/Analysis/CostModel/ARM/load_store.ll
===================================================================
--- llvm/test/Analysis/CostModel/ARM/load_store.ll
+++ llvm/test/Analysis/CostModel/ARM/load_store.ll
@@ -70,7 +70,7 @@
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: store i64 undef, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: store i128 undef, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: store float undef, ptr undef, align 4
-; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: store double undef, ptr undef, align 4
+; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: store double undef, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: store <2 x i8> undef, ptr undef, align 1
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: store <2 x i16> undef, ptr undef, align 2
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: store <2 x i32> undef, ptr undef, align 4
@@ -96,7 +96,7 @@
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: store i64 undef, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: store i128 undef, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: store float undef, ptr undef, align 4
-; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: store double undef, ptr undef, align 4
+; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: store double undef, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: store <2 x i8> undef, ptr undef, align 1
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: store <2 x i16> undef, ptr undef, align 2
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: store <2 x i32> undef, ptr undef, align 4
@@ -257,7 +257,7 @@
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %4 = load i64, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %5 = load i128, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %6 = load float, ptr undef, align 4
-; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %7 = load double, ptr undef, align 4
+; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %7 = load double, ptr undef, align 4
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: %8 = load <2 x i8>, ptr undef, align 1
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: %9 = load <2 x i16>, ptr undef, align 2
 ; CHECK-MVE-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: %10 = load <2 x i32>, ptr undef, align 4
@@ -283,7 +283,7 @@
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %4 = load i64, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %5 = load i128, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %6 = load float, ptr undef, align 4
-; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %7 = load double, ptr undef, align 4
+; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %7 = load double, ptr undef, align 4
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %8 = load <2 x i8>, ptr undef, align 1
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %9 = load <2 x i16>, ptr undef, align 2
 ; CHECK-NEON-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %10 = load <2 x i32>, ptr undef, align 4
Index: llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
===================================================================
--- llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
+++ llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
@@ -1494,6 +1494,16 @@
       return ST->getMVEVectorCostFactor(CostKind);
   }
 
+  // Unaligned double loads introduce 3 extra stores (1 to the stack), 2 loads
+  // and a register move
+  // Unaligned float loads introduce an extra store and a register move
+  // The codegen for these unaligned loads could be improved, but these costs
+  // represent what is currently produced
+  if (ST->hasFPRegs64() && Src->isDoubleTy() && Alignment && Alignment.valueOrOne() < 8)
+    return 6;
+  else if (ST->hasFPRegs() && Src->isFloatTy() && Alignment && Alignment.valueOrOne() < 4)
+    return 2;
+
   int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
                      ? ST->getMVEVectorCostFactor(CostKind)
                      : 1;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D151870.527327.patch
Type: text/x-patch
Size: 5015 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230601/144d82cf/attachment.bin>