From 7a37166a7a2c0c5dd8f8c5d8320f8f7505437f53 Mon Sep 17 00:00:00 2001
From: Jim Grosbach <grosbach@apple.com>
Date: Tue, 11 Mar 2014 00:44:14 +0000
Subject: [PATCH] X86: Enable ISel of 16-bit MOVBE instructions.

When the MOVBE instructions are available, use them for 16-bit endian
swapping as well as for 32 and 64 bit.

The patterns were already present on the instructions, but weren't being
matched because the operation was unconditionally marked to 'Expand.'
Change that to be conditional on whether the MOVBE instructions are
available. Use 'rolw' to implement the in-register version (32 and 64
bit have the dedicated 'bswap' instruction for that).

Patch by Louis Gerbarg <lgg@apple.com>.

rdar://15479984

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203524 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp |  4 ++-
 lib/Target/X86/X86InstrCompiler.td |  6 ++++
 test/CodeGen/X86/movbe.ll          | 45 ++++++++++++++++++++++--------
 3 files changed, 42 insertions(+), 13 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 951d3ded99d..07c83e8b294 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -505,7 +505,9 @@ void X86TargetLowering::resetOperationActions() {
   }
 
   setOperationAction(ISD::READCYCLECOUNTER , MVT::i64  , Custom);
-  setOperationAction(ISD::BSWAP            , MVT::i16  , Expand);
+
+  if (!Subtarget->hasMOVBE())
+    setOperationAction(ISD::BSWAP          , MVT::i16  , Expand);
 
   // These should be promoted to a larger select which is supported.
   setOperationAction(ISD::SELECT          , MVT::i1   , Promote);
diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 8446c2cfb90..a8aff44d06c 100644
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -1839,3 +1839,9 @@ def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
 def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
 def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
 def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
+
+// When HasMOVBE is enabled it is possible to get a non-legalized
+// register-register 16 bit bswap. This maps it to a ROL instruction.
+let Predicates = [HasMOVBE] in {
+ def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
+}
diff --git a/test/CodeGen/X86/movbe.ll b/test/CodeGen/X86/movbe.ll
index 3f459be70d2..e248410b202 100644
--- a/test/CodeGen/X86/movbe.ll
+++ b/test/CodeGen/X86/movbe.ll
@@ -1,45 +1,66 @@
 ; RUN: llc -mtriple=x86_64-linux -mcpu=atom < %s | FileCheck %s
 ; RUN: llc -mtriple=x86_64-linux -mcpu=slm < %s | FileCheck %s -check-prefix=SLM
 
+declare i16 @llvm.bswap.i16(i16) nounwind readnone
 declare i32 @llvm.bswap.i32(i32) nounwind readnone
 declare i64 @llvm.bswap.i64(i64) nounwind readnone
 
-define void @test1(i32* nocapture %x, i32 %y) nounwind {
+define void @test1(i16* nocapture %x, i16 %y) nounwind {
+  %bswap = call i16 @llvm.bswap.i16(i16 %y)
+  store i16 %bswap, i16* %x, align 2
+  ret void
+; CHECK-LABEL: test1:
+; CHECK: movbew %si, (%rdi)
+; SLM-LABEL: test1:
+; SLM: movbew   %si, (%rdi)
+}
+
+define i16 @test2(i16* %x) nounwind {
+  %load = load i16* %x, align 2
+  %bswap = call i16 @llvm.bswap.i16(i16 %load)
+  ret i16 %bswap
+; CHECK-LABEL: test2:
+; CHECK: movbew (%rdi), %ax
+; SLM-LABEL: test2:
+; SLM: movbew   (%rdi), %ax
+}
+
+define void @test3(i32* nocapture %x, i32 %y) nounwind {
   %bswap = call i32 @llvm.bswap.i32(i32 %y)
   store i32 %bswap, i32* %x, align 4
   ret void
-; CHECK-LABEL: test1:
+; CHECK-LABEL: test3:
 ; CHECK: movbel	%esi, (%rdi)
-; SLM-LABEL: test1:
+; SLM-LABEL: test3:
 ; SLM: movbel	%esi, (%rdi)
 }
 
-define i32 @test2(i32* %x) nounwind {
+define i32 @test4(i32* %x) nounwind {
   %load = load i32* %x, align 4
   %bswap = call i32 @llvm.bswap.i32(i32 %load)
   ret i32 %bswap
-; CHECK-LABEL: test2:
+; CHECK-LABEL: test4:
 ; CHECK: movbel	(%rdi), %eax
-; SLM-LABEL: test2:
+; SLM-LABEL: test4:
 ; SLM: movbel	(%rdi), %eax
 }
 
-define void @test3(i64* %x, i64 %y) nounwind {
+define void @test5(i64* %x, i64 %y) nounwind {
   %bswap = call i64 @llvm.bswap.i64(i64 %y)
   store i64 %bswap, i64* %x, align 8
   ret void
-; CHECK-LABEL: test3:
+; CHECK-LABEL: test5:
 ; CHECK: movbeq	%rsi, (%rdi)
-; SLM-LABEL: test3:
+; SLM-LABEL: test5:
 ; SLM: movbeq	%rsi, (%rdi)
 }
 
-define i64 @test4(i64* %x) nounwind {
+define i64 @test6(i64* %x) nounwind {
   %load = load i64* %x, align 8
   %bswap = call i64 @llvm.bswap.i64(i64 %load)
   ret i64 %bswap
-; CHECK-LABEL: test4:
+; CHECK-LABEL: test6:
 ; CHECK: movbeq	(%rdi), %rax
-; SLM-LABEL: test4:
+; SLM-LABEL: test6:
 ; SLM: movbeq	(%rdi), %rax
 }