[PATCH] D150336: [LV][AArch64] Disable maximising bandwidth for streaming compatible sve

[Dinar Temirbulatov via Phabricator via llvm-commits]

dtemirbulatov created this revision.
dtemirbulatov added reviewers: sdesmalen, CarolineConcatto, hassnaa-arm, david-arm.
Herald added subscribers: ctetreau, hiraditya, kristof.beyls, tschuett.
Herald added a project: All.
dtemirbulatov requested review of this revision.
Herald added a subscriber: pcwang-thead.
Herald added a project: LLVM.

We noticed some runtime performance improvements by disabling maximising bandwidth for streaming compatible sve and here is the patch that disables that feature.


https://reviews.llvm.org/D150336

Files:
  llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
  llvm/test/Transforms/LoopVectorize/AArch64/streaming-compatible-sve-no-maximize-bandwidth.ll


Index: llvm/test/Transforms/LoopVectorize/AArch64/streaming-compatible-sve-no-maximize-bandwidth.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/LoopVectorize/AArch64/streaming-compatible-sve-no-maximize-bandwidth.ll
@@ -0,0 +1,68 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=loop-vectorize -debug-only=loop-vectorize -force-streaming-compatible-sve -mattr=+sve -scalable-vectorization=off -aarch64-sve-vector-bits-min=128 -S 2>&1 | FileCheck %s --check-prefix=SC_SVE
+; RUN: opt < %s -passes=loop-vectorize -debug-only=loop-vectorize -mattr=+sve -scalable-vectorization=off -aarch64-sve-vector-bits-min=128 -S -disable-output 2>&1 | FileCheck %s --check-prefix=NO_SC_SVE
+; SC_SVE: LV: Selecting VF: 2.
+; NO_SC_SVE: LV: Selecting VF: 8.
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-none-linux-gnu"
+
+define void @fxpAutoCorrelation() {
+; SC_SVE-LABEL: @fxpAutoCorrelation(
+; SC_SVE-NEXT:  entry:
+; SC_SVE-NEXT:    br i1 true, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; SC_SVE:       vector.ph:
+; SC_SVE-NEXT:    br label [[VECTOR_BODY:%.*]]
+; SC_SVE:       vector.body:
+; SC_SVE-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; SC_SVE-NEXT:    [[VEC_PHI:%.*]] = phi <2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP0:%.*]], [[VECTOR_BODY]] ]
+; SC_SVE-NEXT:    [[VEC_PHI1:%.*]] = phi <2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP1:%.*]], [[VECTOR_BODY]] ]
+; SC_SVE-NEXT:    [[TMP0]] = or <2 x i32> zeroinitializer, [[VEC_PHI]]
+; SC_SVE-NEXT:    [[TMP1]] = or <2 x i32> zeroinitializer, [[VEC_PHI1]]
+; SC_SVE-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; SC_SVE-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
+; SC_SVE-NEXT:    br i1 [[TMP2]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SC_SVE:       middle.block:
+; SC_SVE-NEXT:    [[BIN_RDX:%.*]] = or <2 x i32> [[TMP1]], [[TMP0]]
+; SC_SVE-NEXT:    [[TMP3:%.*]] = call i32 @llvm.vector.reduce.or.v2i32(<2 x i32> [[BIN_RDX]])
+; SC_SVE-NEXT:    [[CMP_N:%.*]] = icmp eq i64 0, 0
+; SC_SVE-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
+; SC_SVE:       scalar.ph:
+; SC_SVE-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 0, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
+; SC_SVE-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[TMP3]], [[MIDDLE_BLOCK]] ]
+; SC_SVE-NEXT:    br label [[FOR_BODY6:%.*]]
+; SC_SVE:       for.body6:
+; SC_SVE-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY6]] ]
+; SC_SVE-NEXT:    [[ACCUMULATOR_032:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[ADD12:%.*]], [[FOR_BODY6]] ]
+; SC_SVE-NEXT:    [[TMP4:%.*]] = load i16, ptr null, align 2
+; SC_SVE-NEXT:    [[CONV10:%.*]] = sext i16 0 to i32
+; SC_SVE-NEXT:    [[MUL:%.*]] = mul i32 [[CONV10]], 0
+; SC_SVE-NEXT:    [[ADD12]] = or i32 0, [[ACCUMULATOR_032]]
+; SC_SVE-NEXT:    [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
+; SC_SVE-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 0
+; SC_SVE-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY6]], !llvm.loop [[LOOP3:![0-9]+]]
+; SC_SVE:       for.end:
+; SC_SVE-NEXT:    [[ADD12_LCSSA:%.*]] = phi i32 [ [[ADD12]], [[FOR_BODY6]] ], [ [[TMP3]], [[MIDDLE_BLOCK]] ]
+; SC_SVE-NEXT:    [[TMP5:%.*]] = lshr i32 [[ADD12_LCSSA]], 0
+; SC_SVE-NEXT:    ret void
+;
+entry:
+  br label %for.body6
+
+for.body6:                                        ; preds = %for.body6, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body6 ]
+  %Accumulator.032 = phi i32 [ 0, %entry ], [ %add12, %for.body6 ]
+  %0 = load i16, ptr null, align 2
+  %conv10 = sext i16 0 to i32
+  %mul = mul i32 %conv10, 0
+  %add12 = or i32 0, %Accumulator.032
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, 0
+  br i1 %exitcond.not, label %for.end, label %for.body6
+
+for.end:                                          ; preds = %for.body6
+  %1 = lshr i32 %add12, 0
+  ret void
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; NO_SC_SVE: {{.*}}
Index: llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
===================================================================
--- llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -153,7 +153,8 @@
 bool AArch64TTIImpl::shouldMaximizeVectorBandwidth(
     TargetTransformInfo::RegisterKind K) const {
   assert(K != TargetTransformInfo::RGK_Scalar);
-  return K == TargetTransformInfo::RGK_FixedWidthVector;
+  return (K == TargetTransformInfo::RGK_FixedWidthVector &&
+          !ST->forceStreamingCompatibleSVE());
 }
 
 /// Calculate the cost of materializing a 64-bit value. This helper


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D150336.521185.patch
Type: text/x-patch
Size: 5032 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230511/fb2ecfae/attachment.bin>