[PATCH] D105587: [Instcombine]Transform reduction+(sext/zext(<n x i1>) to <n x im>) to [-]zext/trunc(ctpop(bitcast <n x i1> to in)) to im.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev created this revision.
ABataev added reviewers: RKSimon, spatel, lebedev.ri.
Herald added a subscriber: hiraditya.
ABataev requested review of this revision.
Herald added a project: LLVM.

Some of the SPEC tests end up with reduction+(sext/zext(<n x i1>) to <n x im>) pattern, which can be transformed to [-]zext/trunc(ctpop(bitcast <n x i1> to in)) to im.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D105587

Files:
  llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
  llvm/test/Transforms/InstCombine/reduction-add-sext-zext-i1.ll


Index: llvm/test/Transforms/InstCombine/reduction-add-sext-zext-i1.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/InstCombine/reduction-add-sext-zext-i1.ll
@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+define i32 @reduce_add_sext(<4 x i1> %x) {
+; CHECK-LABEL: @reduce_add_sext(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <4 x i1> [[X:%.*]] to i4
+; CHECK-NEXT:    [[TMP2:%.*]] = call i4 @llvm.ctpop.i4(i4 [[TMP1]]), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i4 [[TMP2]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = sub nsw i32 0, [[TMP3]]
+; CHECK-NEXT:    ret i32 [[TMP4]]
+;
+  %sext = sext <4 x i1> %x to <4 x i32>
+  %res = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %sext)
+  ret i32 %res
+}
+
+define i64 @reduce_add_zext(<8 x i1> %x) {
+; CHECK-LABEL: @reduce_add_zext(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <8 x i1> [[X:%.*]] to i8
+; CHECK-NEXT:    [[TMP2:%.*]] = call i8 @llvm.ctpop.i8(i8 [[TMP1]]), !range [[RNG1:![0-9]+]]
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP2]] to i64
+; CHECK-NEXT:    ret i64 [[TMP3]]
+;
+  %zext = zext <8 x i1> %x to <8 x i64>
+  %res = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> %zext)
+  ret i64 %res
+}
+
+define i16 @reduce_add_sext_same(<16 x i1> %x) {
+; CHECK-LABEL: @reduce_add_sext_same(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <16 x i1> [[X:%.*]] to i16
+; CHECK-NEXT:    [[TMP2:%.*]] = call i16 @llvm.ctpop.i16(i16 [[TMP1]]), !range [[RNG2:![0-9]+]]
+; CHECK-NEXT:    [[TMP3:%.*]] = sub nsw i16 0, [[TMP2]]
+; CHECK-NEXT:    ret i16 [[TMP3]]
+;
+  %sext = sext <16 x i1> %x to <16 x i16>
+  %res = call i16 @llvm.vector.reduce.add.v16i16(<16 x i16> %sext)
+  ret i16 %res
+}
+
+define i8 @reduce_add_zext_long(<128 x i1> %x) {
+; CHECK-LABEL: @reduce_add_zext_long(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <128 x i1> [[X:%.*]] to i128
+; CHECK-NEXT:    [[TMP2:%.*]] = call i128 @llvm.ctpop.i128(i128 [[TMP1]]), !range [[RNG3:![0-9]+]]
+; CHECK-NEXT:    [[TMP3:%.*]] = trunc i128 [[TMP2]] to i8
+; CHECK-NEXT:    [[TMP4:%.*]] = sub i8 0, [[TMP3]]
+; CHECK-NEXT:    ret i8 [[TMP4]]
+;
+  %zext = sext <128 x i1> %x to <128 x i8>
+  %res = call i8 @llvm.vector.reduce.add.v128i8(<128 x i8> %zext)
+  ret i8 %res
+}
+
+declare i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)
+declare i64 @llvm.vector.reduce.add.v8i64(<8 x i64> %a)
+declare i16 @llvm.vector.reduce.add.v16i16(<16 x i16> %a)
+declare i8 @llvm.vector.reduce.add.v128i8(<128 x i8> %a)
Index: llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -1995,7 +1995,31 @@
       }
     LLVM_FALLTHROUGH;
   }
-  case Intrinsic::vector_reduce_add:
+  case Intrinsic::vector_reduce_add: {
+    if (IID == Intrinsic::vector_reduce_add) {
+      // Convert vector_reduce_add(ZExt(<n x i1>)) to
+      // ZExt(ctpop(bitcast <n x i1> to in)).
+      // Convert vector_reduce_add(SExt(<n x i1>)) to
+      // -ZExt(ctpop(bitcast <n x i1> to in)).
+      Value *Arg = II->getArgOperand(0);
+      Value *Vect;
+      if (match(Arg, m_ZExtOrSExt(m_Value(Vect)))) {
+        if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
+          if (FTy->getElementType() == Builder.getInt1Ty()) {
+            Value *V = Builder.CreateBitCast(
+                Vect, Builder.getIntNTy(FTy->getNumElements()));
+            Value *Res = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, V);
+            if (Res->getType() != II->getType())
+              Res = Builder.CreateZExtOrTrunc(Res, II->getType());
+            if (cast<Instruction>(Arg)->getOpcode() == Instruction::SExt)
+              Res = Builder.CreateNeg(Res);
+            replaceInstUsesWith(CI, Res);
+            return eraseInstFromFunction(CI);
+          }
+      }
+    }
+    LLVM_FALLTHROUGH;
+  }
   case Intrinsic::vector_reduce_mul:
   case Intrinsic::vector_reduce_xor:
   case Intrinsic::vector_reduce_umax:


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D105587.357063.patch
Type: text/x-patch
Size: 4161 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210707/a4519060/attachment.bin>