[PATCH] D14840: [X86] Detect SAD patterns and emit psadbw instructions on X86.

[Ahmed Bougacha via llvm-commits]

ab added inline comments.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:28751-28763
@@ -28750,1 +28750,15 @@
 
+// Check if the given SDValue is a constant vector with all Val.
+static bool isConstVectorOf(SDValue V, int Val) {
+  BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
+  if (!BV || !BV->isConstant())
+    return false;
+  auto NumOperands = V.getNumOperands();
+  for (unsigned i = 0; i < NumOperands; i++) {
+    ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
+    if (!C || C->getSExtValue() != Val)
+      return false;
+  }
+  return true;
+}
+
----------------
Right, isBuildVectorAllOnes looks for -1 (not 1), as it checks that "all of the elements are ~0 or undef." 

================
Comment at: test/CodeGen/X86/sad.ll:2-7
@@ +1,8 @@
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+sse2 -force-target-max-vector-interleave=1 -unroll-count=1 | llc | FileCheck %s --check-prefix=SSE2
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx2 -force-target-max-vector-interleave=1 -unroll-count=1 | llc | FileCheck %s --check-prefix=AVX2
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx512bw -force-target-max-vector-interleave=1 -unroll-count=1 | llc | FileCheck %s --check-prefix=AVX512BW
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+sse2 -force-target-max-vector-interleave=1 | llc | FileCheck %s --check-prefix=SSE2-UNROLL
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx2 -force-target-max-vector-interleave=1 | llc | FileCheck %s --check-prefix=AVX2-UNROLL
+; RUN: opt < %s -O2 -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx512bw -force-target-max-vector-interleave=1 | llc | FileCheck %s --check-prefix=AVX512BW-UNROLL
+
+ at a = global [1024 x i8] zeroinitializer, align 16
----------------
Does it need to be huge? After all, this is only testing the PSAD ISel, and the pattern seems pretty minimal; would it work to test variants of your example in detectSADPattern?


http://reviews.llvm.org/D14840