[PATCH] D35650: [PowerPC] Don't crash on larger splats that can be achieved through 1-byte splats

[Nemanja Ivanovic via Phabricator via llvm-commits]

nemanjai created this revision.

There's code in the back end to ensure that we leave nodes that are 1-byte splats as-is so that they'd match the XXSPLTISB instruction. However, the code doesn't handle a situation where the value being splat is a larger value that happens to have repeating bytes. This patch addresses those situation by adding `v16i8` constant splat and bitcasting it to the target type.

This fixes PR33808.


Repository:
  rL LLVM

https://reviews.llvm.org/D35650

Files:
  lib/Target/PowerPC/PPCISelLowering.cpp
  test/CodeGen/PowerPC/splat-larger-types-as-v16i8.ll


Index: test/CodeGen/PowerPC/splat-larger-types-as-v16i8.ll
===================================================================
--- test/CodeGen/PowerPC/splat-larger-types-as-v16i8.ll
+++ test/CodeGen/PowerPC/splat-larger-types-as-v16i8.ll
@@ -0,0 +1,20 @@
+; RUN: llc -mcpu=pwr9 -mtriple=powerpc64-unknown-unknown < %s | FileCheck %s
+; RUN: llc -mcpu=pwr9 -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s
+define <8 x i16> @test1() {
+entry:
+  ret <8 x i16> <i16 257, i16 257, i16 257, i16 257, i16 257, i16 257, i16 257, i16 257>
+; CHECK-LABEL: test1
+; CHECK: xxspltib 34, 1
+}
+define <8 x i16> @testAB() {
+entry:
+; CHECK-LABEL: testAB
+; CHECK: xxspltib 34, 171
+  ret <8 x i16> <i16 43947, i16 43947, i16 43947, i16 43947, i16 43947, i16 43947, i16 43947, i16 43947>
+}
+define <4 x i32> @testAB32() {
+entry:
+; CHECK-LABEL: testAB32
+; CHECK: xxspltib 34, 171
+  ret <4 x i32> <i32 2880154539, i32 2880154539, i32 2880154539, i32 2880154539>
+}
Index: lib/Target/PowerPC/PPCISelLowering.cpp
===================================================================
--- lib/Target/PowerPC/PPCISelLowering.cpp
+++ lib/Target/PowerPC/PPCISelLowering.cpp
@@ -7646,6 +7646,15 @@
         return DAG.getBitcast(Op.getValueType(), NewBV);
       return NewBV;
     }
+
+    // BuildVectorSDNode::isConstantSplat() is actually pretty smart. It'll
+    // detect that constant splats like v8i16: 0xABAB are really just splats
+    // of a 1-byte constant. In this case, we need to convert the node to a
+    // splat of v16i8 and a bitcast.
+    if (Op.getValueType() != MVT::v16i8)
+      return DAG.getBitcast(Op.getValueType(),
+                            DAG.getConstant(SplatBits, dl, MVT::v16i8));
+
     return Op;
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D35650.107407.patch
Type: text/x-patch
Size: 1735 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170719/ee07bd5b/attachment.bin>