[VPlan] Implement mayHaveSideEffects/mayWriteToMemory for VPInterleav…

[arcbbb]

…eRecipe

This helps VPlanTransforms::removeDeadRecipes to work on VPInterleaveRecipe

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-backend-arm

Author: Shih-Po Hung (arcbbb)

Changes

…eRecipe

This helps VPlanTransforms::removeDeadRecipes to work on VPInterleaveRecipe

---

Full diff: https://github.com/llvm/llvm-project/pull/71360.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp (+4)
• (modified) llvm/test/CodeGen/ARM/loopvectorize_pr33804.ll (+10-2)
• (modified) llvm/test/Transforms/LoopVectorize/AArch64/loopvectorize_pr33804_double.ll (+10-2)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 6b3218dca1b18b0..f8325c645540d56 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -44,6 +44,8 @@ extern cl::opt<bool> EnableVPlanNativePath;
 
 bool VPRecipeBase::mayWriteToMemory() const {
   switch (getVPDefID()) {
+  case VPInterleaveSC:
+    return cast<VPInterleaveRecipe>(this)->getNumStoreOperands() > 0;
   case VPWidenMemoryInstructionSC: {
     return cast<VPWidenMemoryInstructionRecipe>(this)->isStore();
   }
@@ -147,6 +149,8 @@ bool VPRecipeBase::mayHaveSideEffects() const {
            "underlying instruction has side-effects");
     return false;
   }
+  case VPInterleaveSC:
+    return mayWriteToMemory();
   case VPWidenMemoryInstructionSC:
     assert(cast<VPWidenMemoryInstructionRecipe>(this)
                    ->getIngredient()
diff --git a/llvm/test/CodeGen/ARM/loopvectorize_pr33804.ll b/llvm/test/CodeGen/ARM/loopvectorize_pr33804.ll
index 540cbbfe96e8123..547cb4fe0c91825 100644
--- a/llvm/test/CodeGen/ARM/loopvectorize_pr33804.ll
+++ b/llvm/test/CodeGen/ARM/loopvectorize_pr33804.ll
@@ -64,7 +64,7 @@ for.end22.i.i:                                    ; preds = %for.body14.i.i
 ; CHECK-LABEL: @cvCalcEMD3
 ; CHECK: vector.body
 ; CHECK: inttoptr <{{[0-9]+}} x i32>
-define void @cvCalcEMD3() local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
+define void @cvCalcEMD3(ptr %dst) local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
 entry:
   br label %for.body14.i.i
 
@@ -74,6 +74,10 @@ for.body14.i.i:                                   ; preds = %for.body14.i.i, %en
   %loadf = load float, ptr %arrayidx15.i.i1427, align 4
   %next19.i.i = getelementptr inbounds %struct.CvNode1D, ptr undef, i32 %i.1424.i.i, i32 1
   %loadp = load ptr, ptr %next19.i.i, align 4
+  %dst.ptr = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i
+  %dst.ptr.1 = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i, i32 1
+  store float %loadf, ptr %dst.ptr, align 4
+  store ptr %loadp, ptr %dst.ptr.1, align 4
   %inc21.i.i = add nuw nsw i32 %i.1424.i.i, 1
   %exitcond438.i.i = icmp eq i32 %inc21.i.i, 0
   br i1 %exitcond438.i.i, label %for.end22.i.i, label %for.body14.i.i
@@ -87,7 +91,7 @@ for.end22.i.i:                                    ; preds = %for.body14.i.i
 ; CHECK-LABEL: @cvCalcEMD3_2
 ; CHECK: vector.body
 ; CHECK: ptrtoint <{{[0-9]+}} x ptr>
-define void @cvCalcEMD3_2() local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
+define void @cvCalcEMD3_2(ptr %dst) local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
 entry:
   br label %for.body14.i.i
 
@@ -98,6 +102,10 @@ for.body14.i.i:                                   ; preds = %for.body14.i.i, %en
   %arrayidx15.i.i1427 = getelementptr inbounds %struct.CvNode1D2, ptr undef, i32 %i.1424.i.i
   %val.i.i = getelementptr inbounds %struct.CvNode1D2, ptr %arrayidx15.i.i1427, i32 0, i32 1
   %loadf = load float, ptr %val.i.i, align 4
+  %dst.ptr = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i
+  %dst.ptr.1 = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i, i32 1
+  store float %loadf, ptr %dst.ptr, align 4
+  store ptr %loadp, ptr %dst.ptr.1, align 4
   %inc21.i.i = add nuw nsw i32 %i.1424.i.i, 1
   %exitcond438.i.i = icmp eq i32 %inc21.i.i, 0
   br i1 %exitcond438.i.i, label %for.end22.i.i, label %for.body14.i.i
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/loopvectorize_pr33804_double.ll b/llvm/test/Transforms/LoopVectorize/AArch64/loopvectorize_pr33804_double.ll
index 263da076fbb4e13..b7f4eaef9240894 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/loopvectorize_pr33804_double.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/loopvectorize_pr33804_double.ll
@@ -64,7 +64,7 @@ for.end22.i.i:                                    ; preds = %for.body14.i.i
 ; CHECK-LABEL: @cvCalcEMD3
 ; CHECK: vector.body
 ; CHECK: inttoptr <{{[0-9]+}} x i64>
-define void @cvCalcEMD3() local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
+define void @cvCalcEMD3(ptr %dst) local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
 entry:
   br label %for.body14.i.i
 
@@ -74,6 +74,10 @@ for.body14.i.i:                                   ; preds = %for.body14.i.i, %en
   %load_d = load double, ptr %arrayidx15.i.i1427, align 4
   %next19.i.i = getelementptr inbounds %struct.CvNode1D, ptr undef, i32 %i.1424.i.i, i32 1
   %load_p = load ptr, ptr %next19.i.i, align 4
+  %dst.ptr = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i
+  %dst.ptr.1 = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i, i32 1
+  store double %load_d, ptr %dst.ptr, align 4
+  store ptr %load_p, ptr %dst.ptr.1, align 4
   %inc21.i.i = add nuw nsw i32 %i.1424.i.i, 1
   %exitcond438.i.i = icmp eq i32 %inc21.i.i, 0
   br i1 %exitcond438.i.i, label %for.end22.i.i, label %for.body14.i.i
@@ -87,7 +91,7 @@ for.end22.i.i:                                    ; preds = %for.body14.i.i
 ; CHECK-LABEL: @cvCalcEMD3_2
 ; CHECK: vector.body
 ; CHECK: ptrtoint <{{[0-9]+}} x ptr>
-define void @cvCalcEMD3_2() local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
+define void @cvCalcEMD3_2(ptr %dst) local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
 entry:
   br label %for.body14.i.i
 
@@ -98,6 +102,10 @@ for.body14.i.i:                                   ; preds = %for.body14.i.i, %en
   %arrayidx15.i.i1427 = getelementptr inbounds %struct.CvNode1D2, ptr undef, i32 %i.1424.i.i
   %val.i.i = getelementptr inbounds %struct.CvNode1D2, ptr %arrayidx15.i.i1427, i32 0, i32 1
   %load_d = load double, ptr %val.i.i, align 4
+  %dst.ptr = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i
+  %dst.ptr.1 = getelementptr inbounds %struct.CvNode1D, ptr %dst, i32 %i.1424.i.i, i32 1
+  store double %load_d, ptr %dst.ptr, align 4
+  store ptr %load_p, ptr %dst.ptr.1, align 4
   %inc21.i.i = add nuw nsw i32 %i.1424.i.i, 1
   %exitcond438.i.i = icmp eq i32 %inc21.i.i, 0
   br i1 %exitcond438.i.i, label %for.end22.i.i, label %for.body14.i.i

[arcbbb]

ping

[fhahn]

Thanks for the patch!

We should also have a dedicated test with an unused interleave group, to test the patch. Could you add a target-independent one? Both the new test and the test changes in this PR should probably be submitted via a separate PR first, so the current PR then only shows the difference caused. by the patch (i.e. removed the dead interleave group from the new test)

[fhahn]

While you are here, could you also replace the undef pointer operands here and elsewhere to make the test defined properly? The new stores are needed to avoid the loads being optimized away, which would be good to do separately.

[arcbbb]

Sure!

[fhahn]

LGTM on top of the test changes in #75026, thanks!