From bdceb9fb14595d10f7d94e1dd950cf2d94d2f2d3 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Fri, 23 Aug 2019 05:33:27 +0000
Subject: [PATCH] [X86] Improve lowering of v2i32 SAD handling in
 combineLoopSADPattern.

For v2i32 we only feed 2 i8 elements into the psadbw instructions
with 0s in the other 14 bytes. The resulting psadbw instruction
will produce zeros in bits [127:16] of the output. We need to take
the result and feed it to a v2i32 add where the first element
includes bits [15:0] of the sad result. The other element should
be zero.

Prior to this patch we were using a truncate to take 0 from
bits 95:64 of the psadbw. This results in a pshufd to move those
bits to 63:32. But since we also have zeroes in bits 63:32 of
the psadbw output, we should just take those bits.

The previous code probably worked better with promoting legalization,
but now we use widening legalization. I've preserved the old
behavior if -x86-experimental-vector-widening-legalization=false
until we get that option removed.

llvm-svn: 369733
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 13 ++++++++++---
 llvm/test/CodeGen/X86/sad.ll            |  2 --
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 8bc83f150389..a29b6a8283cb 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -43647,10 +43647,13 @@ static SDValue combineLoopSADPattern(SDNode *N, SelectionDAG &DAG,
   // The output of PSADBW is a vector of i64.
   // We need to turn the vector of i64 into a vector of i32.
   // If the reduction vector is at least as wide as the psadbw result, just
-  // bitcast. If it's narrower, truncate - the high i32 of each i64 is zero
-  // anyway.
+  // bitcast. If it's narrower which can only occur for v2i32, bits 127:16 of
+  // the PSADBW will be zero. If we promote/ narrow vectors, truncate the v2i64
+  // result to v2i32 which will be removed by type legalization. If we/ widen
+  // narrow vectors then we bitcast to v4i32 and extract v2i32.
   MVT ResVT = MVT::getVectorVT(MVT::i32, Sad.getValueSizeInBits() / 32);
-  if (VT.getSizeInBits() >= ResVT.getSizeInBits())
+  if (ExperimentalVectorWideningLegalization ||
+      VT.getSizeInBits() >= ResVT.getSizeInBits())
     Sad = DAG.getNode(ISD::BITCAST, DL, ResVT, Sad);
   else
     Sad = DAG.getNode(ISD::TRUNCATE, DL, VT, Sad);
@@ -43660,6 +43663,10 @@ static SDValue combineLoopSADPattern(SDNode *N, SelectionDAG &DAG,
     SDValue Zero = DAG.getConstant(0, DL, VT);
     Sad = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Zero, Sad,
                       DAG.getIntPtrConstant(0, DL));
+  } else if (ExperimentalVectorWideningLegalization &&
+             VT.getSizeInBits() < ResVT.getSizeInBits()) {
+    Sad = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Sad,
+                      DAG.getIntPtrConstant(0, DL));
   }
 
   // Preserve the reduction flag on the ADD. We may need to revisit for the
diff --git a/llvm/test/CodeGen/X86/sad.ll b/llvm/test/CodeGen/X86/sad.ll
index 648f9b44cab1..918841a1864b 100644
--- a/llvm/test/CodeGen/X86/sad.ll
+++ b/llvm/test/CodeGen/X86/sad.ll
@@ -1074,7 +1074,6 @@ define i32 @sad_2i8() nounwind {
 ; SSE2-NEXT:    pand %xmm1, %xmm3
 ; SSE2-NEXT:    pand %xmm1, %xmm2
 ; SSE2-NEXT:    psadbw %xmm3, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
 ; SSE2-NEXT:    paddd %xmm2, %xmm0
 ; SSE2-NEXT:    addq $4, %rax
 ; SSE2-NEXT:    jne .LBB3_1
@@ -1097,7 +1096,6 @@ define i32 @sad_2i8() nounwind {
 ; AVX-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm0[1,2,3,4,5,6,7]
 ; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm0[1,2,3,4,5,6,7]
 ; AVX-NEXT:    vpsadbw %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
 ; AVX-NEXT:    vpaddd %xmm1, %xmm2, %xmm1
 ; AVX-NEXT:    addq $4, %rax
 ; AVX-NEXT:    jne .LBB3_1