[VPlan] Use VPTypeAnalysis to get the step type of widen pointer induction

[Mel-Chen]

This patch uses VPTypeAnalysis to determine its type since the induction step is not always a live-in value in the VPlan and may be defined by a recipe.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-vectorizers

Author: Mel Chen (Mel-Chen)

Changes

This patch uses VPTypeAnalysis to determine its type since the induction step is not always a live-in value in the VPlan and may be defined by a recipe.

---

Full diff: https://github.com/llvm/llvm-project/pull/147925.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+2-2)
• (added) llvm/test/Transforms/LoopVectorize/ivopt-widen-ptr-induction.ll (+75)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 02cea8620d271..53eee046b18da 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -854,8 +854,8 @@ optimizeLatchExitInductionUser(VPlan &Plan, VPTypeAnalysis &TypeInfo,
   if (ScalarTy->isIntegerTy())
     return B.createNaryOp(Instruction::Sub, {EndValue, Step}, {}, "ind.escape");
   if (ScalarTy->isPointerTy()) {
-    auto *Zero = Plan.getOrAddLiveIn(
-        ConstantInt::get(Step->getLiveInIRValue()->getType(), 0));
+    Type *StepTy = TypeInfo.inferScalarType(Step);
+    auto *Zero = Plan.getOrAddLiveIn(ConstantInt::get(StepTy, 0));
     return B.createPtrAdd(EndValue,
                           B.createNaryOp(Instruction::Sub, {Zero, Step}), {},
                           "ind.escape");
diff --git a/llvm/test/Transforms/LoopVectorize/ivopt-widen-ptr-induction.ll b/llvm/test/Transforms/LoopVectorize/ivopt-widen-ptr-induction.ll
new file mode 100644
index 0000000000000..e24c112e9e65c
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/ivopt-widen-ptr-induction.ll
@@ -0,0 +1,75 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -lv-strided-pointer-ivs -S < %s | FileCheck %s
+
+define i64 @widen_ptr_indvar(ptr noalias %a, ptr noalias %b ,i64 %stride, i64 %n) {
+; CHECK-LABEL: define i64 @widen_ptr_indvar(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]], i64 [[STRIDE:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[TMP9:%.*]] = add i64 [[N]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = shl i64 [[STRIDE]], 3
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP9]], 4
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP9]], 4
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP9]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TMP2:%.*]] = mul i64 [[N_VEC]], [[TMP1]]
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr null, i64 [[TMP2]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr i64, ptr [[A]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[TMP4]], i32 0
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i64>, ptr [[TMP5]], align 8
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <4 x i64> [[WIDE_LOAD]], i32 3
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP9]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP8:%.*]] = sub i64 0, [[TMP1]]
+; CHECK-NEXT:    [[IND_ESCAPE:%.*]] = getelementptr i8, ptr [[TMP3]], i64 [[TMP8]]
+; CHECK-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[TMP3]], %[[MIDDLE_BLOCK]] ], [ null, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[FOR_BODY:.*]]
+; CHECK:       [[FOR_BODY]]:
+; CHECK-NEXT:    [[INDVAR:%.*]] = phi i64 [ [[INDVAR_NEXT:%.*]], %[[FOR_BODY]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ]
+; CHECK-NEXT:    [[PTR_IV:%.*]] = phi ptr [ [[PTR_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], %[[SCALAR_PH]] ]
+; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr i64, ptr [[A]], i64 [[INDVAR]]
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr [[ARRAYIDX]], align 8
+; CHECK-NEXT:    [[PTR_IV_NEXT]] = getelementptr i64, ptr [[PTR_IV]], i64 [[STRIDE]]
+; CHECK-NEXT:    [[INDVAR_NEXT]] = add i64 [[INDVAR]], 1
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVAR]], [[N]]
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label %[[EXIT]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    [[PTR_IV_LCSSA:%.*]] = phi ptr [ [[PTR_IV]], %[[FOR_BODY]] ], [ [[IND_ESCAPE]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[DOTLCSSA:%.*]] = phi i64 [ [[TMP0]], %[[FOR_BODY]] ], [ [[TMP7]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[CAST_PTR:%.*]] = ptrtoint ptr [[PTR_IV_LCSSA]] to i64
+; CHECK-NEXT:    [[RESULT:%.*]] = add i64 [[CAST_PTR]], [[DOTLCSSA]]
+; CHECK-NEXT:    ret i64 [[RESULT]]
+;
+entry:
+  br label %for.body
+
+for.body:
+  %indvar = phi i64 [ %indvar.next, %for.body ], [ 0, %entry ]
+  %ptr.iv = phi ptr [ %ptr.iv.next, %for.body ], [ null, %entry ]
+  %arrayidx = getelementptr i64, ptr %a, i64 %indvar
+  %0 = load i64, ptr %arrayidx, align 8
+  %ptr.iv.next = getelementptr i64, ptr %ptr.iv, i64 %stride
+  %indvar.next = add i64 %indvar, 1
+  %exitcond.not = icmp eq i64 %indvar, %n
+  br i1 %exitcond.not, label %exit, label %for.body
+
+exit:
+  %cast.ptr = ptrtoint ptr %ptr.iv to i64
+  %result = add i64 %cast.ptr, %0
+  ret i64 %result
+}
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+;.

[david-arm]

Is it possible to also have a test case where the widened pointer induction is used inside the loop?

[Mel-Chen]

d8e8414

[david-arm]

Is it worth having a test case for a use of %ptr.iv.next too?

[Mel-Chen]

d8e8414

[fhahn]

Hmm, wondering what the longer-term plan for lv-strided-pointer-ivs should be and if anybody is using it?

[Mel-Chen]

Hmm, wondering what the longer-term plan for lv-strided-pointer-ivs should be and if anybody is using it?

We've had this feature enabled by default in our internal branch for a long time to support RISC-V's strided segment accesses. That's why our tests were able to catch this bug.

[Mel-Chen]

ping

[fhahn]

LGTM, thanks