[LSV] Don't move stores across may-load instrs, and loosen restrictions on moving loads.

[jlebar]

Summary

Previously we wouldn't move loads/stores across instructions that had
side-effects, where that was defined as may-write may-throw. But this
is both overly- and underly-restrictive:

• Stores can't safely be moved across instructions that may load.
• Loads can be moved across may-throw instructions; it's only may-write instructions that must be a barrier.

This patch also adds a DEBUG check that all instructions in our chain
are either loads or stores.

[jlebar]

Per dannyb's comment, don't move loads above throwing instructions. See rL270828 and PR 27858.

[asbirlea]

test/Transforms/LoadStoreVectorizer/NVPTX/merge-across-side-effects.ll

2 ↗

(On Diff #64608)

Could you double check the validation for these tests is done in getVectorizablePrefix, and there isn't an earlier exit on the data layout (just before the calls to getVectorizablePrefix)? Possibly add a data layout to avoid that, in case the defaults change?

[jlebar]

test/Transforms/LoadStoreVectorizer/NVPTX/merge-across-side-effects.ll

2 ↗

(On Diff #64608)

Could you double check the validation for these tests is done in getVectorizablePrefix

Sorry, not sure what you mean.

there isn't an earlier exit on the data layout

I think we should be good here because if the pass doesn't modify any of the functions, the test will also fail.

[asbirlea]

test/Transforms/LoadStoreVectorizer/NVPTX/merge-across-side-effects.ll

2 ↗

(On Diff #64608)

Ok, you're right, on early exit it wouldn't vectorize any of these.

llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp

Vectorizer::splitOddVectorElts(ArrayRef<Value *> Chain,
   return std::make_pair(Chain.slice(0, NumLeft), Chain.slice(NumLeft));
 }
 
 ArrayRef<Value *> Vectorizer::getVectorizablePrefix(ArrayRef<Value *> Chain) {
   // These are in BB order, unlike Chain, which is in address order.
   SmallVector<std::pair<Value *, unsigned>, 16> MemoryInstrs;
   SmallVector<std::pair<Value *, unsigned>, 16> ChainInstrs;
 
+  bool IsLoadChain = isa<LoadInst>(Chain[0]);
+  DEBUG({
+    for (Value *V : Chain) {
+      if (IsLoadChain)
+        assert(isa<LoadInst>(V) &&
+               "All elements of Chain must be loads, or all must be stores.");
+      else
+        assert(isa<StoreInst>(V) &&
+               "All elements of Chain must be loads, or all must be stores.");
+    }
+  });
+
   unsigned InstrIdx = 0;
   for (Instruction &I : make_range(getBoundaryInstrs(Chain))) {
     ++InstrIdx;
     if (isa<LoadInst>(I) || isa<StoreInst>(I)) {
       if (!is_contained(Chain, &I))
         MemoryInstrs.push_back({&I, InstrIdx});
       else
         ChainInstrs.push_back({&I, InstrIdx});
-    } else if (I.mayHaveSideEffects()) {
-      DEBUG(dbgs() << "LSV: Found side-effecting operation: " << I << '\n');
+    } else if (IsLoadChain && (I.mayWriteToMemory() || I.mayThrow())) {
+      DEBUG(dbgs() << "LSV: Found may-write/throw operation: " << I << '\n');
+      break;
+    } else if (!IsLoadChain && (I.mayReadOrWriteMemory() || I.mayThrow())) {
+      DEBUG(dbgs() << "LSV: Found may-read/write/throw operation: " << I
+                   << '\n');
       break;
     }
   }
 
   // Loop until we find an instruction in ChainInstrs that we can't vectorize.
   unsigned ChainInstrIdx, ChainInstrsLen;
   for (ChainInstrIdx = 0, ChainInstrsLen = ChainInstrs.size();
        ChainInstrIdx < ChainInstrsLen; ++ChainInstrIdx) {

llvm/trunk/test/Transforms/LoadStoreVectorizer/NVPTX/merge-across-side-effects.ll

 ; RUN: opt -mtriple=nvptx64-nvidia-cuda -load-store-vectorizer -S -o - %s | FileCheck %s
 
-; If we have a chain of loads or stores with a side-effecting operation in the
-; middle, we should still be able to merge the loads/stores that appear
-; before/after the side-effecting op.  We just can't merge *across* the
-; side-effecting op.
+; Check that the load/store vectorizer is willing to move loads/stores across
+; intervening instructions only if it's safe.
+;
+;  - Loads can be moved across instructions that don't write or throw.
+;  - Stores can only be moved across instructions which don't read, write, or
+;    throw.
+
+declare void @fn()
+declare void @fn_nounwind() #0
+declare void @fn_nounwind_writeonly() #1
+declare void @fn_nounwind_readonly() #2
+declare void @fn_writeonly() #3
+declare void @fn_readonly() #4
+declare void @fn_readnone() #5
 
-declare void @fn() #0
-
-; CHECK-LABEL: @merge_stores
-; CHECK: store <2 x i32> <i32 100, i32 101>
+; CHECK-LABEL: @load_fn
+; CHECK: load
 ; CHECK: call void @fn()
-; CHECK: store <2 x i32> <i32 102, i32 103>
-define void @merge_stores(i32* %out) #0 {
-  %out.gep.1 = getelementptr i32, i32* %out, i32 1
-  %out.gep.2 = getelementptr i32, i32* %out, i32 2
-  %out.gep.3 = getelementptr i32, i32* %out, i32 3
+; CHECK: load
+define void @load_fn(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
 
-  store i32 101, i32* %out.gep.1
-  store i32 100, i32* %out
+  %v0 = load i32, i32* %p
   call void @fn()
-  store i32 102, i32* %out.gep.2
-  store i32 103, i32* %out.gep.3
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @load_fn_nounwind
+; CHECK: load
+; CHECK: call void @fn_nounwind()
+; CHECK: load
+define void @load_fn_nounwind(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_nounwind() #0
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @load_fn_nounwind_writeonly
+; CHECK: load
+; CHECK: call void @fn_nounwind_writeonly()
+; CHECK: load
+define void @load_fn_nounwind_writeonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_nounwind_writeonly() #1
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @load_fn_nounwind_readonly
+; CHECK-DAG: load <2 x i32>
+; CHECK-DAG: call void @fn_nounwind_readonly()
+define void @load_fn_nounwind_readonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_nounwind_readonly() #2
+  %v1 = load i32, i32* %p.1
   ret void
 }
 
-; CHECK-LABEL: @merge_loads
-; CHECK: load <2 x i32>
+; CHECK-LABEL: @load_fn_readonly
+; CHECK: load
+; CHECK: call void @fn_readonly
+; CHECK: load
+define void @load_fn_readonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_readonly() #4
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @load_fn_writeonly
+; CHECK: load
+; CHECK: call void @fn_writeonly()
+; CHECK: load
+define void @load_fn_writeonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_writeonly() #3
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @load_fn_readnone
+; CHECK-DAG: load <2 x i32>
+; CHECK-DAG: call void @fn_readnone()
+define void @load_fn_readnone(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  %v0 = load i32, i32* %p
+  call void @fn_readnone() #5
+  %v1 = load i32, i32* %p.1
+  ret void
+}
+
+; ------------------------------------------------
+; Same tests, but now for stores instead of loads.
+; ------------------------------------------------
+
+; CHECK-LABEL: @store_fn
+; CHECK: store
 ; CHECK: call void @fn()
-; CHECK: load <2 x i32>
-define i32 @merge_loads(i32* %in) #0 {
-  %in.gep.1 = getelementptr i32, i32* %in, i32 1
-  %in.gep.2 = getelementptr i32, i32* %in, i32 2
-  %in.gep.3 = getelementptr i32, i32* %in, i32 3
+; CHECK: store
+define void @store_fn(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
 
-  %v1 = load i32, i32* %in
-  %v2 = load i32, i32* %in.gep.1
+  store i32 0, i32* %p
   call void @fn()
-  %v3 = load i32, i32* %in.gep.2
-  %v4 = load i32, i32* %in.gep.3
+  store i32 0, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @store_fn_nounwind
+; CHECK: store
+; CHECK: call void @fn_nounwind()
+; CHECK: store
+define void @store_fn_nounwind(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_nounwind() #0
+  store i32 0, i32* %p.1
+  ret void
+}
 
-  %sum1 = add i32 %v1, %v2
-  %sum2 = add i32 %sum1, %v3
-  %sum3 = add i32 %sum2, %v4
-  ret i32 %v4
+; CHECK-LABEL: @store_fn_nounwind_writeonly
+; CHECK: store
+; CHECK: call void @fn_nounwind_writeonly()
+; CHECK: store
+define void @store_fn_nounwind_writeonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_nounwind_writeonly() #1
+  store i32 0, i32* %p.1
+  ret void
 }
 
+; CHECK-LABEL: @store_fn_nounwind_readonly
+; CHECK: store
+; CHECK: call void @fn_nounwind_readonly()
+; CHECK: store
+define void @store_fn_nounwind_readonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_nounwind_readonly() #2
+  store i32 0, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @store_fn_readonly
+; CHECK: store
+; CHECK: call void @fn_readonly
+; CHECK: store
+define void @store_fn_readonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_readonly() #4
+  store i32 0, i32* %p.1
+  ret void
+}
+
+; CHECK-LABEL: @store_fn_writeonly
+; CHECK: store
+; CHECK: call void @fn_writeonly()
+; CHECK: store
+define void @store_fn_writeonly(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_writeonly() #3
+  store i32 0, i32* %p.1
+  ret void
+}
+
+; This is the only store idiom we can vectorize.
+; CHECK-LABEL: @store_fn_readnone
+; CHECK-DAG: store <2 x i32>
+; CHECK-DAG: call void @fn_readnone()
+define void @store_fn_readnone(i32* %p) #0 {
+  %p.1 = getelementptr i32, i32* %p, i32 1
+
+  store i32 0, i32* %p
+  call void @fn_readnone() #5
+  store i32 0, i32* %p.1
+  ret void
+}
+
+
 attributes #0 = { nounwind }
+attributes #1 = { nounwind writeonly }
+attributes #2 = { nounwind readonly }
+attributes #3 = { writeonly }
+attributes #4 = { readonly }
+; readnone implies nounwind, so no need to test separately
+attributes #5 = { nounwind readnone }