[X86] Use (SUBREG_TO_REG (MOV32rm)) for extloadi64i8/extloadi64i16 when the load is 4 byte aligned or better and not volatile.

Summary:
Previously we would use MOVZXrm8/MOVZXrm16, but those are longer encodings.

This is similar to what we do in the loadi32 predicate.

Reviewers: RKSimon, spatel

Reviewed By: RKSimon

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D60341

llvm-svn: 357875

2 files changed, 17 insertions, 3 deletions

diff --git a/llvm/lib/Target/X86/X86InstrCompiler.td b/llvm/lib/Target/X86/X86InstrCompiler.td
index 1fcc33e8724..dbc2fbcc6e7 100644
--- a/llvm/lib/Target/X86/X86InstrCompiler.td
+++ b/llvm/lib/Target/X86/X86InstrCompiler.td
@@ -1279,14 +1279,16 @@ def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
 
 // For other extloads, use subregs, since the high contents of the register are
 // defined after an extload.
+// NOTE: The extloadi64i32 pattern needs to be first as it will try to form
+// 32-bit loads for 4 byte aligned i8/i16 loads.
+def : Pat<(extloadi64i32 addr:$src),
+          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
 def : Pat<(extloadi64i1 addr:$src),
           (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
 def : Pat<(extloadi64i8 addr:$src),
           (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
 def : Pat<(extloadi64i16 addr:$src),
           (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
-def : Pat<(extloadi64i32 addr:$src),
-          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
 
 // anyext. Define these to do an explicit zero-extend to
 // avoid partial-register updates.
diff --git a/llvm/lib/Target/X86/X86InstrInfo.td b/llvm/lib/Target/X86/X86InstrInfo.td
index dc5e04cdd67..0176c2d707a 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.td
+++ b/llvm/lib/Target/X86/X86InstrInfo.td
@@ -1121,7 +1121,19 @@ def extloadi32i16  : PatFrag<(ops node:$ptr), (i32 (extloadi16 node:$ptr))>;
 def extloadi64i1   : PatFrag<(ops node:$ptr), (i64 (extloadi1 node:$ptr))>;
 def extloadi64i8   : PatFrag<(ops node:$ptr), (i64 (extloadi8 node:$ptr))>;
 def extloadi64i16  : PatFrag<(ops node:$ptr), (i64 (extloadi16 node:$ptr))>;
-def extloadi64i32  : PatFrag<(ops node:$ptr), (i64 (extloadi32 node:$ptr))>;
+
+// We can treat an i8/i16 extending load to i64 as a 32 bit load if its known
+// to be 4 byte aligned or better.
+def extloadi64i32  : PatFrag<(ops node:$ptr), (i64 (unindexedload node:$ptr)), [{
+  LoadSDNode *LD = cast<LoadSDNode>(N);
+  ISD::LoadExtType ExtType = LD->getExtensionType();
+  if (ExtType != ISD::EXTLOAD)
+    return false;
+  if (LD->getMemoryVT() == MVT::i32)
+    return true;
+
+  return LD->getAlignment() >= 4 && !LD->isVolatile();
+}]>;
 
 
 // An 'and' node with a single use.