On some ARM cpus, flags setting movs with shifter operand, i.e. lsl, lsr, asr,

are more expensive than the non-flag setting variant. Teach thumb2 size
reduction pass to avoid generating them unless we are optimizing for size.

rdar://12892707

llvm-svn: 170728

1 files changed, 69 insertions, 13 deletions

diff --git a/llvm/test/CodeGen/Thumb2/thumb2-shifter.ll b/llvm/test/CodeGen/Thumb2/thumb2-shifter.ll
index 98854a1205f..05dd90cfbfe 100644
--- a/llvm/test/CodeGen/Thumb2/thumb2-shifter.ll
+++ b/llvm/test/CodeGen/Thumb2/thumb2-shifter.ll
@@ -1,24 +1,27 @@
-; RUN: llc < %s -march=thumb -mattr=+thumb2,+t2xtpk | FileCheck %s
+; RUN: llc < %s -march=thumb -mcpu=cortex-a8 | FileCheck %s --check-prefix=A8
+; RUN: llc < %s -march=thumb -mcpu=swift | FileCheck %s --check-prefix=SWIFT
+
+; rdar://12892707
 
 define i32 @t2ADDrs_lsl(i32 %X, i32 %Y) {
-; CHECK: t2ADDrs_lsl
-; CHECK: add.w  r0, r0, r1, lsl #16
+; A8: t2ADDrs_lsl
+; A8: add.w  r0, r0, r1, lsl #16
         %A = shl i32 %Y, 16
         %B = add i32 %X, %A
         ret i32 %B
 }
 
 define i32 @t2ADDrs_lsr(i32 %X, i32 %Y) {
-; CHECK: t2ADDrs_lsr
-; CHECK: add.w  r0, r0, r1, lsr #16
+; A8: t2ADDrs_lsr
+; A8: add.w  r0, r0, r1, lsr #16
         %A = lshr i32 %Y, 16
         %B = add i32 %X, %A
         ret i32 %B
 }
 
 define i32 @t2ADDrs_asr(i32 %X, i32 %Y) {
-; CHECK: t2ADDrs_asr
-; CHECK: add.w  r0, r0, r1, asr #16
+; A8: t2ADDrs_asr
+; A8: add.w  r0, r0, r1, asr #16
         %A = ashr i32 %Y, 16
         %B = add i32 %X, %A
         ret i32 %B
@@ -26,8 +29,8 @@ define i32 @t2ADDrs_asr(i32 %X, i32 %Y) {
 
 ; i32 ror(n) = (x >> n) | (x << (32 - n))
 define i32 @t2ADDrs_ror(i32 %X, i32 %Y) {
-; CHECK: t2ADDrs_ror
-; CHECK: add.w  r0, r0, r1, ror #16
+; A8: t2ADDrs_ror
+; A8: add.w  r0, r0, r1, ror #16
         %A = lshr i32 %Y, 16
         %B = shl  i32 %Y, 16
         %C = or   i32 %B, %A
@@ -36,13 +39,66 @@ define i32 @t2ADDrs_ror(i32 %X, i32 %Y) {
 }
 
 define i32 @t2ADDrs_noRegShift(i32 %X, i32 %Y, i8 %sh) {
-; CHECK: t2ADDrs_noRegShift
-; CHECK: uxtb r2, r2
-; CHECK: lsls r1, r2
-; CHECK: add  r0, r1
+; A8: t2ADDrs_noRegShift
+; A8: uxtb r2, r2
+; A8: lsls r1, r2
+; A8: add  r0, r1
+
+; SWIFT: t2ADDrs_noRegShift
+; SWIFT-NOT: lsls
+; SWIFT: lsl.w
+        %shift.upgrd.1 = zext i8 %sh to i32
+        %A = shl i32 %Y, %shift.upgrd.1
+        %B = add i32 %X, %A
+        ret i32 %B
+}
+
+define i32 @t2ADDrs_noRegShift2(i32 %X, i32 %Y, i8 %sh) {
+; A8: t2ADDrs_noRegShift2
+; A8: uxtb r2, r2
+; A8: lsrs r1, r2
+; A8: add  r0, r1
+
+; SWIFT: t2ADDrs_noRegShift2
+; SWIFT-NOT: lsrs
+; SWIFT: lsr.w
+        %shift.upgrd.1 = zext i8 %sh to i32
+        %A = lshr i32 %Y, %shift.upgrd.1
+        %B = add i32 %X, %A
+        ret i32 %B
+}
+
+define i32 @t2ADDrs_noRegShift3(i32 %X, i32 %Y, i8 %sh) {
+; A8: t2ADDrs_noRegShift3
+; A8: uxtb r2, r2
+; A8: asrs r1, r2
+; A8: add  r0, r1
+
+; SWIFT: t2ADDrs_noRegShift3
+; SWIFT-NOT: asrs
+; SWIFT: asr.w
+        %shift.upgrd.1 = zext i8 %sh to i32
+        %A = ashr i32 %Y, %shift.upgrd.1
+        %B = add i32 %X, %A
+        ret i32 %B
+}
+
+define i32 @t2ADDrs_optsize(i32 %X, i32 %Y, i8 %sh) optsize {
+; SWIFT: t2ADDrs_optsize
+; SWIFT-NOT: lsl.w
+; SWIFT: lsls
         %shift.upgrd.1 = zext i8 %sh to i32
         %A = shl i32 %Y, %shift.upgrd.1
         %B = add i32 %X, %A
         ret i32 %B
 }
 
+define i32 @t2ADDrs_minsize(i32 %X, i32 %Y, i8 %sh) minsize {
+; SWIFT: t2ADDrs_minsize
+; SWIFT-NOT: lsr.w
+; SWIFT: lsrs
+        %shift.upgrd.1 = zext i8 %sh to i32
+        %A = lshr i32 %Y, %shift.upgrd.1
+        %B = add i32 %X, %A
+        ret i32 %B
+}