[RISCV] Allow non-loop invariant steps in RISCVGatherScatterLowering

The motivation for this is to allow us to match strided accesses that are emitted from the loop vectorizer with EVL tail folding (see llvm#122232)

In these loops the step isn't loop invariant and is based off of @llvm.experimental.get.vector.length.

We can relax this as long as we make sure to construct the updates after the definition inside the loop, instead of the preheader.

I presume the restriction was previously added so that the step would dominate the insertion point in the preheader. I can't think of why it wouldn't be safe to calculate it in the loop otherwise.

Author: lukel97

llvm/lib/Target/RISCV/RISCVGatherScatterLowering.cpp

@@ -212,10 +212,6 @@ bool RISCVGatherScatterLowering::matchStridedRecurrence(Value *Index, Loop *L,
     assert(Phi->getIncomingValue(IncrementingBlock) == Inc &&
            "Expected one operand of phi to be Inc");
 
-    // Only proceed if the step is loop invariant.
-    if (!L->isLoopInvariant(Step))
-      return false;
-
     // Step should be a splat.
     Step = getSplatValue(Step);
     if (!Step)
@@ -311,18 +307,31 @@ bool RISCVGatherScatterLowering::matchStridedRecurrence(Value *Index, Loop *L,
   }
   case Instruction::Mul: {
     Start = Builder.CreateMul(Start, SplatOp, "start");
-    Step = Builder.CreateMul(Step, SplatOp, "step");
     Stride = Builder.CreateMul(Stride, SplatOp, "stride");
     break;
   }
   case Instruction::Shl: {
     Start = Builder.CreateShl(Start, SplatOp, "start");
-    Step = Builder.CreateShl(Step, SplatOp, "step");
     Stride = Builder.CreateShl(Stride, SplatOp, "stride");
     break;
   }
   }
 
+  // Adjust the step value after its definition if it's an instruction.
+  if (auto *StepI = dyn_cast<Instruction>(Step))
+    Builder.SetInsertPoint(*StepI->getInsertionPointAfterDef());
+
+  switch (BO->getOpcode()) {
+  default:
+    break;
+  case Instruction::Mul:
+    Step = Builder.CreateMul(Step, SplatOp, "step");
+    break;
+  case Instruction::Shl:
+    Step = Builder.CreateShl(Step, SplatOp, "step");
+    break;
+  }
+
   Inc->setOperand(StepIndex, Step);
   BasePtr->setIncomingValue(StartBlock, Start);
   return true;

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-load-store.ll

@@ -320,8 +320,8 @@ for.cond.cleanup:                                 ; preds = %vector.body
 define void @gather_unknown_pow2(ptr noalias nocapture %A, ptr noalias nocapture readonly %B, i64 %shift) {
 ; CHECK-LABEL: @gather_unknown_pow2(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[STEP:%.*]] = shl i64 8, [[SHIFT:%.*]]
-; CHECK-NEXT:    [[STRIDE:%.*]] = shl i64 1, [[SHIFT]]
+; CHECK-NEXT:    [[STRIDE:%.*]] = shl i64 1, [[SHIFT:%.*]]
+; CHECK-NEXT:    [[STEP:%.*]] = shl i64 8, [[SHIFT]]
 ; CHECK-NEXT:    [[TMP0:%.*]] = mul i64 [[STRIDE]], 4
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:

llvm/test/CodeGen/RISCV/rvv/strided-load-store.ll

@@ -403,23 +403,22 @@ define <vscale x 1 x i64> @gather_loop_variant_step(ptr %a, i32 %len) {
 ; CHECK-LABEL: @gather_loop_variant_step(
 ; CHECK-NEXT:  vector.ph:
 ; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[LEN:%.*]] to i64
-; CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 1 x i64> @llvm.stepvector.nxv1i64()
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[VEC_IND_SCALAR:%.*]] = phi i64 [ 0, [[VECTOR_PH:%.*]] ], [ [[VEC_IND_NEXT_SCALAR:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 1 x i64> [ [[TMP0]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_IND_SCALAR1:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT_SCALAR1:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[ACCUM:%.*]] = phi <vscale x 1 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[ACCUM_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[ELEMS:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[VEC_IND_SCALAR]]
 ; CHECK-NEXT:    [[EVL:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[ELEMS]], i32 1, i1 true)
 ; CHECK-NEXT:    [[EVL_ZEXT:%.*]] = zext i32 [[EVL]] to i64
-; CHECK-NEXT:    [[OFFSET:%.*]] = shl <vscale x 1 x i64> [[VEC_IND]], splat (i64 4)
-; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_FOO:%.*]], ptr [[A:%.*]], <vscale x 1 x i64> [[OFFSET]], i32 3
-; CHECK-NEXT:    [[GATHER:%.*]] = call <vscale x 1 x i64> @llvm.masked.gather.nxv1i64.nxv1p0(<vscale x 1 x ptr> [[TMP1]], i32 8, <vscale x 1 x i1> splat (i1 true), <vscale x 1 x i64> undef)
+; CHECK-NEXT:    [[STEP:%.*]] = shl i64 [[EVL_ZEXT]], 4
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr [[STRUCT_FOO:%.*]], ptr [[A:%.*]], i64 [[VEC_IND_SCALAR1]], i32 3
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.vscale.i32()
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 1 x i64> @llvm.experimental.vp.strided.load.nxv1i64.p0.i64(ptr [[TMP0]], i64 256, <vscale x 1 x i1> splat (i1 true), i32 [[TMP1]])
+; CHECK-NEXT:    [[GATHER:%.*]] = call <vscale x 1 x i64> @llvm.vp.select.nxv1i64(<vscale x 1 x i1> splat (i1 true), <vscale x 1 x i64> [[TMP2]], <vscale x 1 x i64> undef, i32 [[TMP1]])
 ; CHECK-NEXT:    [[ACCUM_NEXT]] = add <vscale x 1 x i64> [[ACCUM]], [[GATHER]]
 ; CHECK-NEXT:    [[VEC_IND_NEXT_SCALAR]] = add nuw i64 [[VEC_IND_SCALAR]], [[EVL_ZEXT]]
-; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 1 x i64> poison, i64 [[EVL_ZEXT]], i64 0
-; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 1 x i64> [[DOTSPLATINSERT]], <vscale x 1 x i64> poison, <vscale x 1 x i32> zeroinitializer
-; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 1 x i64> [[VEC_IND]], [[DOTSPLAT]]
+; CHECK-NEXT:    [[VEC_IND_NEXT_SCALAR1]] = add i64 [[VEC_IND_SCALAR1]], [[STEP]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i64 [[VEC_IND_NEXT_SCALAR]], [[WIDE_TRIP_COUNT]]
 ; CHECK-NEXT:    br i1 [[TMP3]], label [[FOR_COND_CLEANUP:%.*]], label [[VECTOR_BODY]]
 ; CHECK:       for.cond.cleanup: