[llvm] [SLP]Do not require external uses for roots and single use for other instructions in computeMinimumValueSizes. (PR #72679)

[Alexey Bataev via llvm-commits]

https://github.com/alexey-bataev created https://github.com/llvm/llvm-project/pull/72679

After changes, that does not require support from InstCombine, we can
drop some extra requirements for values-to-be-demoted. No need to check
for external uses for roots/other instructions, just check that the
  no non-vectorized insertelement instruction, which may require
  widening.


>From 17d5ac0e52ca593e70e3608fbe6278bda1aa0581 Mon Sep 17 00:00:00 2001
From: Alexey Bataev <a.bataev at outlook.com>
Date: Fri, 17 Nov 2023 08:43:06 -0800
Subject: [PATCH] [SLP]Do not require external uses for roots and single use
 for other instructions in computeMinimumValueSizes.

After changes, that does not require support from InstCombine, we can
drop some extra requirements for values-to-be-demoted. No need to check
for external uses for roots/other instructions, just check that the
  no non-vectorized insertelement instruction, which may require
  widening.
---
 .../Transforms/Vectorize/SLPVectorizer.cpp    | 14 +++++-------
 .../X86/root-trunc-extract-reuse.ll           | 22 +++++++++----------
 2 files changed, 16 insertions(+), 20 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index bff3e49b14bc6bc..42e4540eb8ecbd7 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -13078,10 +13078,13 @@ bool BoUpSLP::collectValuesToDemote(Value *V,
     return true;
   }
 
-  // If the value is not a vectorized instruction in the expression with only
-  // one use, it cannot be demoted.
+  // If the value is not a vectorized instruction in the expression and not the
+  // used by the insertelement instruction, it cannot be demoted.
   auto *I = dyn_cast<Instruction>(V);
-  if (!I || !I->hasOneUse() || !getTreeEntry(I) || !Visited.insert(I).second)
+  if (!I || !getTreeEntry(I) || !Visited.insert(I).second ||
+      all_of(I->users(), [&](User *U) {
+        return isa<InsertElementInst>(U) && !getTreeEntry(U);
+      }))
     return false;
 
   switch (I->getOpcode()) {
@@ -13140,11 +13143,6 @@ bool BoUpSLP::collectValuesToDemote(Value *V,
 }
 
 void BoUpSLP::computeMinimumValueSizes() {
-  // If there are no external uses, the expression tree must be rooted by a
-  // store. We can't demote in-memory values, so there is nothing to do here.
-  if (ExternalUses.empty())
-    return;
-
   // We only attempt to truncate integer expressions.
   auto &TreeRoot = VectorizableTree[0]->Scalars;
   auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType());
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/root-trunc-extract-reuse.ll b/llvm/test/Transforms/SLPVectorizer/X86/root-trunc-extract-reuse.ll
index f48528e502b8cf1..af46b4f576234b2 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/root-trunc-extract-reuse.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/root-trunc-extract-reuse.ll
@@ -8,20 +8,18 @@ define i1 @test() {
 ; CHECK:       then:
 ; CHECK-NEXT:    br label [[ELSE]]
 ; CHECK:       else:
-; CHECK-NEXT:    [[TMP0:%.*]] = phi <2 x i1> [ zeroinitializer, [[THEN]] ], [ zeroinitializer, [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[TMP1:%.*]] = zext <2 x i1> [[TMP0]] to <2 x i32>
-; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <2 x i1> [[TMP0]], i32 0
-; CHECK-NEXT:    [[TMP3:%.*]] = zext i1 [[TMP2]] to i32
-; CHECK-NEXT:    [[BF_CAST162:%.*]] = and i32 [[TMP3]], 0
-; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <2 x i32> zeroinitializer, <2 x i32> [[TMP1]], <2 x i32> <i32 3, i32 1>
-; CHECK-NEXT:    [[T13:%.*]] = and <2 x i32> [[TMP4]], zeroinitializer
+; CHECK-NEXT:    [[TMP0:%.*]] = phi <2 x i32> [ zeroinitializer, [[THEN]] ], [ zeroinitializer, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x i32> [[TMP0]], i32 0
+; CHECK-NEXT:    [[BF_CAST162:%.*]] = and i32 [[TMP1]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <2 x i32> zeroinitializer, <2 x i32> [[TMP0]], <2 x i32> <i32 3, i32 1>
+; CHECK-NEXT:    [[T13:%.*]] = and <2 x i32> [[TMP2]], zeroinitializer
 ; CHECK-NEXT:    br label [[ELSE1:%.*]]
 ; CHECK:       else1:
-; CHECK-NEXT:    [[TMP5:%.*]] = shufflevector <2 x i32> [[T13]], <2 x i32> poison, <2 x i32> <i32 poison, i32 0>
-; CHECK-NEXT:    [[TMP6:%.*]] = insertelement <2 x i32> [[TMP5]], i32 [[BF_CAST162]], i32 0
-; CHECK-NEXT:    [[TMP7:%.*]] = icmp ugt <2 x i32> [[TMP6]], zeroinitializer
-; CHECK-NEXT:    [[TMP8:%.*]] = extractelement <2 x i1> [[TMP7]], i32 1
-; CHECK-NEXT:    ret i1 [[TMP8]]
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <2 x i32> [[T13]], <2 x i32> poison, <2 x i32> <i32 poison, i32 0>
+; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <2 x i32> [[TMP3]], i32 [[BF_CAST162]], i32 0
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ugt <2 x i32> [[TMP4]], zeroinitializer
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x i1> [[TMP5]], i32 1
+; CHECK-NEXT:    ret i1 [[TMP6]]
 ;
 entry:
   br i1 false, label %then, label %else