[PATCH] D154717: [LV] Check if ops can safely be truncated in computeMinimumValueSizes.

[Florian Hahn via Phabricator via llvm-commits]

fhahn created this revision.
fhahn added reviewers: Ayal, gilr, rengolin, reames.
Herald added subscribers: artagnon, StephenFan, hiraditya.
Herald added a project: All.
fhahn requested review of this revision.
Herald added a subscriber: wangpc.
Herald added a project: LLVM.

Update computeMinimumValueSizes to check if an instruction's operands
can safely be truncated.

If more than MinBW bits are demanded by for the operand or if the
operand is a constant and cannot be safely truncated, it is not safe to
evaluate the instruction in the narrower MinBW. Skip those cases.

Fixes https://github.com/llvm/llvm-project/issues/47927


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D154717

Files:
  llvm/lib/Analysis/VectorUtils.cpp
  llvm/test/Transforms/LoopVectorize/trunc-shifts.ll


Index: llvm/test/Transforms/LoopVectorize/trunc-shifts.ll
===================================================================
--- llvm/test/Transforms/LoopVectorize/trunc-shifts.ll
+++ llvm/test/Transforms/LoopVectorize/trunc-shifts.ll
@@ -17,17 +17,15 @@
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i8
 ; CHECK-NEXT:    [[TMP0:%.*]] = add i8 [[OFFSET_IDX]], 0
-; CHECK-NEXT:    [[TMP1:%.*]] = trunc <4 x i32> [[BROADCAST_SPLAT]] to <4 x i8>
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr <4 x i8> [[TMP1]], <i8 18, i8 18, i8 18, i8 18>
-; CHECK-NEXT:    [[TMP3:%.*]] = zext <4 x i8> [[TMP2]] to <4 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = trunc <4 x i32> [[TMP3]] to <4 x i8>
-; CHECK-NEXT:    [[TMP5:%.*]] = zext i8 [[TMP0]] to i64
-; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 [[TMP5]]
-; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i8, ptr [[TMP6]], i32 0
-; CHECK-NEXT:    store <4 x i8> [[TMP4]], ptr [[TMP7]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr <4 x i32> [[BROADCAST_SPLAT]], <i32 18, i32 18, i32 18, i32 18>
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc <4 x i32> [[TMP1]] to <4 x i8>
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP0]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 [[TMP3]]
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 0
+; CHECK-NEXT:    store <4 x i8> [[TMP2]], ptr [[TMP5]], align 8
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 100
-; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], 100
+; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 100, 100
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
Index: llvm/lib/Analysis/VectorUtils.cpp
===================================================================
--- llvm/lib/Analysis/VectorUtils.cpp
+++ llvm/lib/Analysis/VectorUtils.cpp
@@ -663,8 +663,25 @@
       Type *Ty = M->getType();
       if (Roots.count(M))
         Ty = cast<Instruction>(M)->getOperand(0)->getType();
-      if (MinBW < Ty->getScalarSizeInBits())
-        MinBWs[cast<Instruction>(M)] = MinBW;
+
+      if (MinBW >= Ty->getScalarSizeInBits())
+        continue;
+
+      // Check if any of M's operands demand more bits than MinBW or if it is a
+      // constant cannot be safely truncated to MinBW. In any of those cases, M
+      // cannot be performed safely in MinBW.
+      if (any_of(cast<Instruction>(M)->operands(), [&DB, MinBW](Use &U) {
+            if (auto *CI = dyn_cast<ConstantInt>(U)) {
+              APInt I = CI->getValue();
+              if (I.trunc(MinBW).zext(I.getBitWidth()) == I)
+                return false;
+            }
+            uint64_t BW = bit_width(DB.getDemandedBits(&U).getZExtValue());
+            return bit_ceil(BW) > MinBW;
+          }))
+        continue;
+
+      MinBWs[cast<Instruction>(M)] = MinBW;
     }
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D154717.538126.patch
Type: text/x-patch
Size: 3308 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230707/9b41ec0f/attachment.bin>