X86: Shrink certain forms of movsx.

In particular:
movsbw %al, %ax   --> cbtw
movswl %ax, %eax  --> cwtl
movslq %eax, %rax --> cltq

According to Intel's manual those have the same performance characteristics but
come with a smaller encoding.

llvm-svn: 186174

4 files changed, 27 insertions, 6 deletions

diff --git a/llvm/test/CodeGen/X86/fast-isel-ret-ext.ll b/llvm/test/CodeGen/X86/fast-isel-ret-ext.ll
index fd768cb9b33..395ad1e616a 100644
--- a/llvm/test/CodeGen/X86/fast-isel-ret-ext.ll
+++ b/llvm/test/CodeGen/X86/fast-isel-ret-ext.ll
@@ -26,7 +26,7 @@ define signext i16 @test4(i32 %y) nounwind {
   %conv = trunc i32 %y to i16
   ret i16 %conv
   ; CHECK: test4:
-  ; CHECK: movswl {{.*}}, %eax
+  ; CHECK: {{(movswl.%.x, %eax|cwtl)}}
 }
 
 define zeroext i1 @test5(i32 %y) nounwind {
diff --git a/llvm/test/CodeGen/X86/mcinst-lowering.ll b/llvm/test/CodeGen/X86/mcinst-lowering.ll
index 1ef5a971bab..391f9398f0b 100644
--- a/llvm/test/CodeGen/X86/mcinst-lowering.ll
+++ b/llvm/test/CodeGen/X86/mcinst-lowering.ll
@@ -24,3 +24,21 @@ if.end:                                           ; preds = %entry
 return:                                           ; preds = %entry
   ret i32 0
 }
+
+define i32 @f1() nounwind {
+  %ax = tail call i16 asm sideeffect "", "={ax},~{dirflag},~{fpsr},~{flags}"()
+  %conv = sext i16 %ax to i32
+  ret i32 %conv
+
+; CHECK: f1:
+; CHECK: cwtl ## encoding: [0x98]
+}
+
+define i64 @f2() nounwind {
+  %eax = tail call i32 asm sideeffect "", "={ax},~{dirflag},~{fpsr},~{flags}"()
+  %conv = sext i32 %eax to i64
+  ret i64 %conv
+
+; CHECK: f2:
+; CHECK: cltq ## encoding: [0x48,0x98]
+}
diff --git a/llvm/test/CodeGen/X86/shl_elim.ll b/llvm/test/CodeGen/X86/shl_elim.ll
index 83e1eb5c39e..e99ecac8723 100644
--- a/llvm/test/CodeGen/X86/shl_elim.ll
+++ b/llvm/test/CodeGen/X86/shl_elim.ll
@@ -1,6 +1,4 @@
-; RUN: llc < %s -march=x86 | grep "movl	8(.esp), %eax"
-; RUN: llc < %s -march=x86 | grep "shrl	.eax"
-; RUN: llc < %s -march=x86 | grep "movswl	.ax, .eax"
+; RUN: llc < %s -march=x86 | FileCheck %s
 
 define i32 @test1(i64 %a) nounwind {
         %tmp29 = lshr i64 %a, 24                ; <i64> [#uses=1]
@@ -9,5 +7,10 @@ define i32 @test1(i64 %a) nounwind {
         %tmp45 = trunc i32 %tmp410 to i16               ; <i16> [#uses=1]
         %tmp456 = sext i16 %tmp45 to i32                ; <i32> [#uses=1]
         ret i32 %tmp456
+
+; CHECK: test1:
+; CHECK: movl 8(%esp), %eax
+; CHECK: shrl %eax
+; CHECK: cwtl
 }
 
diff --git a/llvm/test/CodeGen/X86/widen_conv-2.ll b/llvm/test/CodeGen/X86/widen_conv-2.ll
index 969cb512beb..db8fa93dc79 100644
--- a/llvm/test/CodeGen/X86/widen_conv-2.ll
+++ b/llvm/test/CodeGen/X86/widen_conv-2.ll
@@ -1,6 +1,6 @@
 ; RUN: llc < %s -march=x86 -mattr=+sse42 | FileCheck %s
-; CHECK: movswl
-; CHECK: movswl
+; CHECK: cwtl
+; CHECK: cwtl
 
 ; sign extension v2i32 to v2i16