[llvm-branch-commits] [llvm] 2e080eb - [SVE] Add support for scalable vectorization of loops with selects and cmps

David Sherwood via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Fri Jan 22 01:52:59 PST 2021


Author: David Sherwood
Date: 2021-01-22T09:48:13Z
New Revision: 2e080eb00ad76654313e0e119bb7fa0ffe2f9866

URL: https://github.com/llvm/llvm-project/commit/2e080eb00ad76654313e0e119bb7fa0ffe2f9866
DIFF: https://github.com/llvm/llvm-project/commit/2e080eb00ad76654313e0e119bb7fa0ffe2f9866.diff

LOG: [SVE] Add support for scalable vectorization of loops with selects and cmps

I have removed an unnecessary assert in LoopVectorizationCostModel::getInstructionCost
that prevented a cost being calculated for select instructions when using
scalable vectors. In addition, I have changed AArch64TTIImpl::getCmpSelInstrCost
to only do special cost calculations for fixed width vectors and fall
back to the base version for scalable vectors.

I have added a simple cost model test for cmps and selects:

  test/Analysis/CostModel/sve-cmpsel.ll

and some simple tests that show we vectorize loops with cmp and select:

  test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll

Differential Revision: https://reviews.llvm.org/D95039

Added: 
    llvm/test/Analysis/CostModel/sve-cmpsel.ll
    llvm/test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll

Modified: 
    llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
    llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index ffa045846e59..7fda6b8fb602 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -707,7 +707,7 @@ int AArch64TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
   // We don't lower some vector selects well that are wider than the register
   // width.
-  if (ValTy->isVectorTy() && ISD == ISD::SELECT) {
+  if (isa<FixedVectorType>(ValTy) && ISD == ISD::SELECT) {
     // We would need this many instructions to hide the scalarization happening.
     const int AmortizationCost = 20;
 
@@ -749,6 +749,8 @@ int AArch64TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
         return Entry->Cost;
     }
   }
+  // The base case handles scalable vectors fine for now, since it treats the
+  // cost as 1 * legalization cost.
   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
 }
 

diff  --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 1bc4afeae5f9..9e157f3061b6 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7334,10 +7334,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF,
     const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
     bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
     Type *CondTy = SI->getCondition()->getType();
-    if (!ScalarCond) {
-      assert(!VF.isScalable() && "VF is assumed to be non scalable.");
+    if (!ScalarCond)
       CondTy = VectorType::get(CondTy, VF);
-    }
     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy,
                                   CmpInst::BAD_ICMP_PREDICATE, CostKind, I);
   }

diff  --git a/llvm/test/Analysis/CostModel/sve-cmpsel.ll b/llvm/test/Analysis/CostModel/sve-cmpsel.ll
new file mode 100644
index 000000000000..163c863c1ea3
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/sve-cmpsel.ll
@@ -0,0 +1,146 @@
+; RUN: opt -cost-model -analyze -mtriple=aarch64--linux-gnu -mattr=+sve  < %s 2>%t | FileCheck %s
+
+; RUN: FileCheck --check-prefix=WARN --allow-empty %s <%t
+
+; If this check fails please read test/CodeGen/AArch64/README for instructions on how to resolve it.
+; WARN-NOT: warning
+
+; Check icmp for legal integer vectors.
+define void @cmp_legal_int() {
+; CHECK-LABEL: 'cmp_legal_int'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = icmp ne <vscale x 2 x i64> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = icmp ne <vscale x 4 x i32> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = icmp ne <vscale x 8 x i16> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = icmp ne <vscale x 16 x i8> undef, undef
+  %1 = icmp ne <vscale x 2 x i64> undef, undef
+  %2 = icmp ne <vscale x 4 x i32> undef, undef
+  %3 = icmp ne <vscale x 8 x i16> undef, undef
+  %4 = icmp ne <vscale x 16 x i8> undef, undef
+  ret void
+}
+
+; Check icmp for an illegal integer vector.
+define <vscale x 4 x i1> @cmp_nxv4i64() {
+; CHECK-LABEL: 'cmp_nxv4i64'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = icmp ne <vscale x 4 x i64> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 4 x i1> %res
+  %res = icmp ne <vscale x 4 x i64> undef, undef
+  ret <vscale x 4 x i1> %res
+}
+
+; Check icmp for legal predicate vectors.
+define void @cmp_legal_pred() {
+; CHECK-LABEL: 'cmp_legal_pred'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = icmp ne <vscale x 2 x i1> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = icmp ne <vscale x 4 x i1> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = icmp ne <vscale x 8 x i1> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = icmp ne <vscale x 16 x i1> undef, undef
+  %1 = icmp ne <vscale x 2 x i1> undef, undef
+  %2 = icmp ne <vscale x 4 x i1> undef, undef
+  %3 = icmp ne <vscale x 8 x i1> undef, undef
+  %4 = icmp ne <vscale x 16 x i1> undef, undef
+  ret void
+}
+
+; Check icmp for an illegal predicate vector.
+define <vscale x 32 x i1> @cmp_nxv32i1() {
+; CHECK-LABEL: 'cmp_nxv32i1'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = icmp ne <vscale x 32 x i1> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 32 x i1> %res
+  %res = icmp ne <vscale x 32 x i1> undef, undef
+  ret <vscale x 32 x i1> %res
+}
+
+; Check fcmp for legal FP vectors
+define void @cmp_legal_fp() #0 {
+; CHECK-LABEL: 'cmp_legal_fp'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = fcmp oge <vscale x 2 x double> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = fcmp oge <vscale x 4 x float> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = fcmp oge <vscale x 8 x half> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = fcmp oge <vscale x 8 x bfloat> undef, undef
+  %1 = fcmp oge <vscale x 2 x double> undef, undef
+  %2 = fcmp oge <vscale x 4 x float> undef, undef
+  %3 = fcmp oge <vscale x 8 x half> undef, undef
+  %4 = fcmp oge <vscale x 8 x bfloat> undef, undef
+  ret void
+}
+
+; Check fcmp for an illegal FP vector
+define <vscale x 16 x i1> @cmp_nxv16f16() {
+; CHECK-LABEL: 'cmp_nxv16f16'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = fcmp oge <vscale x 16 x half> undef, undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 16 x i1> %res
+  %res = fcmp oge <vscale x 16 x half> undef, undef
+  ret <vscale x 16 x i1> %res
+}
+
+; Check select for legal integer vectors
+define void @sel_legal_int() {
+; CHECK-LABEL: 'sel_legal_int'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = select <vscale x 2 x i1> undef, <vscale x 2 x i64> undef, <vscale x 2 x i64> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = select <vscale x 4 x i1> undef, <vscale x 4 x i32> undef, <vscale x 4 x i32> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = select <vscale x 8 x i1> undef, <vscale x 8 x i16> undef, <vscale x 8 x i16> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = select <vscale x 16 x i1> undef, <vscale x 16 x i8> undef, <vscale x 16 x i8> undef
+  %1 = select <vscale x 2 x i1> undef, <vscale x 2 x i64> undef, <vscale x 2 x i64> undef
+  %2 = select <vscale x 4 x i1> undef, <vscale x 4 x i32> undef, <vscale x 4 x i32> undef
+  %3 = select <vscale x 8 x i1> undef, <vscale x 8 x i16> undef, <vscale x 8 x i16> undef
+  %4 = select <vscale x 16 x i1> undef, <vscale x 16 x i8> undef, <vscale x 16 x i8> undef
+  ret void
+}
+
+; Check select for an illegal integer vector
+define <vscale x 16 x i16> @sel_nxv16i16() {
+; CHECK-LABEL: 'sel_nxv16i16'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = select <vscale x 16 x i1> undef, <vscale x 16 x i16> undef, <vscale x 16 x i16> undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 16 x i16> %res
+  %res = select <vscale x 16 x i1> undef, <vscale x 16 x i16> undef, <vscale x 16 x i16> undef
+  ret <vscale x 16 x i16> %res
+}
+
+; Check select for a legal FP vector
+define void @sel_legal_fp() #0 {
+; CHECK-LABEL: 'sel_legal_fp'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = select <vscale x 2 x i1> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = select <vscale x 4 x i1> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = select <vscale x 8 x i1> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = select <vscale x 8 x i1> undef, <vscale x 8 x bfloat> undef, <vscale x 8 x bfloat> undef
+  %1 = select <vscale x 2 x i1> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef
+  %2 = select <vscale x 4 x i1> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef
+  %3 = select <vscale x 8 x i1> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef
+  %4 = select <vscale x 8 x i1> undef, <vscale x 8 x bfloat> undef, <vscale x 8 x bfloat> undef
+  ret void
+}
+
+; Check select for an illegal FP vector
+define <vscale x 8 x float> @sel_nxv8f32() {
+; CHECK-LABEL: 'sel_nxv8f32'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = select <vscale x 8 x i1> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 8 x float> %res
+  %res = select <vscale x 8 x i1> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef
+  ret <vscale x 8 x float> %res
+}
+
+; Check select for a legal predicate vector
+define void @sel_legal_pred() {
+; CHECK-LABEL: 'sel_legal_pred'
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %1 = select <vscale x 2 x i1> undef, <vscale x 2 x i1> undef, <vscale x 2 x i1> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %2 = select <vscale x 4 x i1> undef, <vscale x 4 x i1> undef, <vscale x 4 x i1> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %3 = select <vscale x 8 x i1> undef, <vscale x 8 x i1> undef, <vscale x 8 x i1> undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %4 = select <vscale x 16 x i1> undef, <vscale x 16 x i1> undef, <vscale x 16 x i1> undef
+  %1 = select <vscale x 2 x i1> undef, <vscale x 2 x i1> undef, <vscale x 2 x i1> undef
+  %2 = select <vscale x 4 x i1> undef, <vscale x 4 x i1> undef, <vscale x 4 x i1> undef
+  %3 = select <vscale x 8 x i1> undef, <vscale x 8 x i1> undef, <vscale x 8 x i1> undef
+  %4 = select <vscale x 16 x i1> undef, <vscale x 16 x i1> undef, <vscale x 16 x i1> undef
+  ret void
+}
+
+; Check select for an illegal predicate vector
+define <vscale x 32 x i1> @sel_nxv32i1() {
+; CHECK-LABEL: 'sel_nxv32i1'
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %res = select <vscale x 32 x i1> undef, <vscale x 32 x i1> undef, <vscale x 32 x i1> undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret <vscale x 32 x i1> %res
+  %res = select <vscale x 32 x i1> undef, <vscale x 32 x i1> undef, <vscale x 32 x i1> undef
+  ret <vscale x 32 x i1> %res
+}
+
+attributes #0 = { "target-features"="+sve,+bf16" }

diff  --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll
new file mode 100644
index 000000000000..f59b2bf85db3
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll
@@ -0,0 +1,78 @@
+; RUN: opt -loop-vectorize -dce -instcombine -mtriple aarch64-linux-gnu -mattr=+sve < %s -S 2>%t | FileCheck %s
+
+; RUN: FileCheck --check-prefix=WARN --allow-empty %s <%t
+
+; If this check fails please read test/CodeGen/AArch64/README for instructions on how to resolve it.
+; WARN-NOT: warning
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-unknown-linux-gnu"
+
+define void @cmpsel_i32(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i64 %n) {
+; CHECK-LABEL: @cmpsel_i32(
+; CHECK-NEXT:  entry:
+; CHECK:       vector.body:
+; CHECK:         [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, <vscale x 4 x i32>* {{.*}}, align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq <vscale x 4 x i32> [[WIDE_LOAD]], shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 0, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[TMP2:%.*]] = select <vscale x 4 x i1> [[TMP1]], <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 2, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer), <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 10, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK:         store <vscale x 4 x i32> [[TMP2]], <vscale x 4 x i32>* {{.*}}, align 4
+;
+entry:
+  %cmp7 = icmp sgt i64 %n, 0
+  br i1 %cmp7, label %for.body, label %for.end
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %tobool.not = icmp eq i32 %0, 0
+  %cond = select i1 %tobool.not, i32 2, i32 10
+  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
+  store i32 %cond, i32* %arrayidx2, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, %n
+  br i1 %exitcond.not, label %for.end.loopexit, label %for.body, !llvm.loop !0
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
+define void @cmpsel_f32(float* noalias nocapture %a, float* noalias nocapture readonly %b, i64 %n) {
+; CHECK-LABEL: @cmpsel_f32(
+; CHECK-NEXT:  entry:
+; CHECK:       vector.body:
+; CHECK:         [[WIDE_LOAD:%.*]] = load <vscale x 4 x float>, <vscale x 4 x float>* {{.*}}, align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = fcmp ogt <vscale x 4 x float> [[WIDE_LOAD]], shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float 3.000000e+00, i32 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    [[TMP2:%.*]] = select <vscale x 4 x i1> [[TMP1]], <vscale x 4 x float> shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float 1.000000e+01, i32 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer), <vscale x 4 x float> shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float 2.000000e+00, i32 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer)
+; CHECK:         store <vscale x 4 x float> [[TMP2]], <vscale x 4 x float>* {{.*}}, align 4
+
+entry:
+  %cmp8 = icmp sgt i64 %n, 0
+  br i1 %cmp8, label %for.body, label %for.end
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %b, i64 %indvars.iv
+  %0 = load float, float* %arrayidx, align 4
+  %cmp1 = fcmp ogt float %0, 3.000000e+00
+  %conv = select i1 %cmp1, float 1.000000e+01, float 2.000000e+00
+  %arrayidx3 = getelementptr inbounds float, float* %a, i64 %indvars.iv
+  store float %conv, float* %arrayidx3, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, %n
+  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !6
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+}
+
+!0 = distinct !{!0, !1, !2, !3, !4, !5}
+!1 = !{!"llvm.loop.mustprogress"}
+!2 = !{!"llvm.loop.vectorize.width", i32 4}
+!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+!4 = !{!"llvm.loop.interleave.count", i32 1}
+!5 = !{!"llvm.loop.vectorize.enable", i1 true}
+!6 = distinct !{!6, !1, !2, !3, !4, !5}


        


More information about the llvm-branch-commits mailing list