[ARM] Fix vector saddsat costs.

It turns out the vectorizer calls the getIntrinsicInstrCost functions
with a scalar return type and vector VF. This updates the costmodel to
handle that, still producing the correct vector costs.

A vectorizer test is added to show it vectorizing at the correct factor
again.

Author: davemgreen

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

@@ -1531,8 +1531,13 @@ int ARMTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
   case Intrinsic::usub_sat: {
     if (!ST->hasMVEIntegerOps())
       break;
+    // Get the Return type, either directly of from ICA.ReturnType and ICA.VF.
+    Type *VT = ICA.getReturnType();
+    if (!VT->isVectorTy() && !ICA.getVectorFactor().isScalar())
+      VT = VectorType::get(VT, ICA.getVectorFactor());
+
     std::pair<int, MVT> LT =
-        TLI->getTypeLegalizationCost(DL, ICA.getReturnType());
+        TLI->getTypeLegalizationCost(DL, VT);
     if (LT.second == MVT::v4i32 || LT.second == MVT::v8i16 ||
         LT.second == MVT::v16i8) {
       // This is a base cost of 1 for the vadd, plus 3 extract shifts if we

llvm/test/Transforms/LoopVectorize/ARM/mve-saddsatcost.ll

@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -loop-vectorize -instcombine -simplifycfg < %s -S -o - | FileCheck %s --check-prefix=CHECK
+
+target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
+target triple = "thumbv8.1m.main-arm-none-eabi"
+
+define void @arm_offset_q15(i16* nocapture readonly %pSrc, i16 signext %offset, i16* nocapture noalias %pDst, i32 %blockSize) #0 {
+; CHECK-LABEL: @arm_offset_q15(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP_NOT6:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP_NOT6]], label [[WHILE_END:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[BLOCKSIZE]], 7
+; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -8
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT8:%.*]] = insertelement <8 x i16> poison, i16 [[OFFSET:%.*]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT9:%.*]] = shufflevector <8 x i16> [[BROADCAST_SPLATINSERT8]], <8 x i16> poison, <8 x i32> zeroinitializer
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i16, i16* [[PSRC:%.*]], i32 [[INDEX]]
+; CHECK-NEXT:    [[NEXT_GEP5:%.*]] = getelementptr i16, i16* [[PDST:%.*]], i32 [[INDEX]]
+; CHECK-NEXT:    [[ACTIVE_LANE_MASK:%.*]] = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 [[INDEX]], i32 [[BLOCKSIZE]])
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast i16* [[NEXT_GEP]] to <8 x i16>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* [[TMP0]], i32 2, <8 x i1> [[ACTIVE_LANE_MASK]], <8 x i16> poison)
+; CHECK-NEXT:    [[TMP1:%.*]] = call <8 x i16> @llvm.sadd.sat.v8i16(<8 x i16> [[WIDE_MASKED_LOAD]], <8 x i16> [[BROADCAST_SPLAT9]])
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i16* [[NEXT_GEP5]] to <8 x i16>*
+; CHECK-NEXT:    call void @llvm.masked.store.v8i16.p0v8i16(<8 x i16> [[TMP1]], <8 x i16>* [[TMP2]], i32 2, <8 x i1> [[ACTIVE_LANE_MASK]])
+; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP3]], label [[WHILE_END]], label [[VECTOR_BODY]], [[LOOP0:!llvm.loop !.*]]
+; CHECK:       while.end:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %cmp.not6 = icmp eq i32 %blockSize, 0
+  br i1 %cmp.not6, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %blkCnt.09 = phi i32 [ %dec, %while.body ], [ %blockSize, %entry ]
+  %pSrc.addr.08 = phi i16* [ %incdec.ptr, %while.body ], [ %pSrc, %entry ]
+  %pDst.addr.07 = phi i16* [ %incdec.ptr3, %while.body ], [ %pDst, %entry ]
+  %incdec.ptr = getelementptr inbounds i16, i16* %pSrc.addr.08, i32 1
+  %0 = load i16, i16* %pSrc.addr.08, align 2
+  %1 = tail call i16 @llvm.sadd.sat.i16(i16 %0, i16 %offset)
+  %incdec.ptr3 = getelementptr inbounds i16, i16* %pDst.addr.07, i32 1
+  store i16 %1, i16* %pDst.addr.07, align 2
+  %dec = add i32 %blkCnt.09, -1
+  %cmp.not = icmp eq i32 %dec, 0
+  br i1 %cmp.not, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare i16 @llvm.sadd.sat.i16(i16, i16)
+
+attributes #0 = { "target-features"="+mve" }