From 077774b8200b41d2a52196b4adcc1a67ba526ace Mon Sep 17 00:00:00 2001
From: Bruno Cardoso Lopes <bruno.cardoso@gmail.com>
Date: Tue, 3 Feb 2015 14:46:49 +0000
Subject: [PATCH] [X86][MMX] Improve transfer from mmx to i32

Improve EXTRACT_VECTOR_ELT DAG combine to catch conversion patterns
between x86mmx and i32 with more layers of indirection.

Before:
  movq2dq %mm0, %xmm0
  movd %xmm0, %eax
After:
  movd %mm0, %eax

llvm-svn: 227969
---
 llvm/lib/Target/X86/X86ISelLowering.cpp  | 31 +++++++++++++++++++++++--------
 llvm/test/CodeGen/X86/vec_extract-mmx.ll | 11 ++++-------
 2 files changed, 27 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index bd2f04f..df79845 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -22764,14 +22764,29 @@ static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
 
   SDValue InputVector = N->getOperand(0);
 
-  // Detect whether we are trying to convert from mmx to i32 and the bitcast
-  // from mmx to v2i32 has a single usage.
-  if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
-      InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
-      InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
-    return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
-                       N->getValueType(0),
-                       InputVector.getNode()->getOperand(0));
+  // Detect mmx to i32 conversion through a v2i32 elt extract.
+  if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
+      N->getValueType(0) == MVT::i32 &&
+      InputVector.getValueType() == MVT::v2i32) {
+
+    // The bitcast source is a direct mmx result.
+    SDValue MMXSrc = InputVector.getNode()->getOperand(0);
+    if (MMXSrc.getValueType() == MVT::x86mmx)
+      return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
+                         N->getValueType(0),
+                         InputVector.getNode()->getOperand(0));
+
+    // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
+    SDValue MMXSrcOp = MMXSrc.getOperand(0);
+    if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
+        MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
+        MMXSrcOp.getOpcode() == ISD::BITCAST &&
+        MMXSrcOp.getValueType() == MVT::v1i64 &&
+        MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
+      return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
+                         N->getValueType(0),
+                         MMXSrcOp.getOperand(0));
+  }
 
   // Only operate on vectors of 4 elements, where the alternative shuffling
   // gets to be more expensive.
diff --git a/llvm/test/CodeGen/X86/vec_extract-mmx.ll b/llvm/test/CodeGen/X86/vec_extract-mmx.ll
index 9b638b7..af65a96 100644
--- a/llvm/test/CodeGen/X86/vec_extract-mmx.ll
+++ b/llvm/test/CodeGen/X86/vec_extract-mmx.ll
@@ -4,8 +4,7 @@ define i32 @test0(<1 x i64>* %v4) {
 ; CHECK-LABEL: test0:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    pshufw $238, (%rdi), %mm0
-; CHECK-NEXT:    movq2dq %mm0, %xmm0
-; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    movd %mm0, %eax
 ; CHECK-NEXT:    addl $32, %eax
 ; CHECK-NEXT:    retq
   %v5 = load <1 x i64>* %v4, align 8
@@ -24,11 +23,10 @@ define i32 @test0(<1 x i64>* %v4) {
 define i32 @test1(i32* nocapture readonly %ptr) {
 ; CHECK-LABEL: test1:
 ; CHECK:       ## BB#0: ## %entry
-; CHECK-NEXT:    movd (%rdi), %xmm0
+; CHECK-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; CHECK-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    pshufw $232, -{{[0-9]+}}(%rsp), %mm0
-; CHECK-NEXT:    movq2dq %mm0, %xmm0
-; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    movd %mm0, %eax
 ; CHECK-NEXT:    emms
 ; CHECK-NEXT:    retq
 entry:
@@ -54,8 +52,7 @@ define i32 @test2(i32* nocapture readonly %ptr) {
 ; CHECK:       ## BB#0: ## %entry
 ; CHECK-NEXT:    movq (%rdi), %mm0
 ; CHECK-NEXT:    pshufw $232, %mm0, %mm0
-; CHECK-NEXT:    movq2dq %mm0, %xmm0
-; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    movd %mm0, %eax
 ; CHECK-NEXT:    emms
 ; CHECK-NEXT:    retq
 entry:
-- 
2.7.4