ARM: Negative offset support problem

[oota-llvm.git] / test / CodeGen / ARM / longMAC.ll

diff --git a/test/CodeGen/ARM/longMAC.ll b/test/CodeGen/ARM/longMAC.ll
index e4a00e9ac303d5225ad56c530315e00e600331ce..fed6ec02f32d9018d0e51cc7807b9dc813b4c114 100644 (file)
--- a/test/CodeGen/ARM/longMAC.ll
+++ b/test/CodeGen/ARM/longMAC.ll
@@ -1,8 +1,11 @@
-; RUN: llc < %s -march=arm | FileCheck %s
+; RUN: llc -mtriple=arm-eabi %s -o - | FileCheck %s -check-prefix=CHECK --check-prefix=CHECK-LE
+; RUN: llc -mtriple=armv7-eabi %s -o - | FileCheck %s --check-prefix=CHECK-V7-LE
+; RUN: llc -mtriple=armeb-eabi %s -o - | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-BE
+; RUN: llc -mtriple=armebv7-eabi %s -o - | FileCheck %s -check-prefix=CHECK-V7-BE
 ; Check generated signed and unsigned multiply accumulate long.
 
 define i64 @MACLongTest1(i32 %a, i32 %b, i64 %c) {
-;CHECK: MACLongTest1:
+;CHECK-LABEL: MACLongTest1:
 ;CHECK: umlal
   %conv = zext i32 %a to i64
   %conv1 = zext i32 %b to i64
@@ -12,7 +15,7 @@ define i64 @MACLongTest1(i32 %a, i32 %b, i64 %c) {
 }
 
 define i64 @MACLongTest2(i32 %a, i32 %b, i64 %c)  {
-;CHECK: MACLongTest2:
+;CHECK-LABEL: MACLongTest2:
 ;CHECK: smlal
   %conv = sext i32 %a to i64
   %conv1 = sext i32 %b to i64
@@ -22,7 +25,7 @@ define i64 @MACLongTest2(i32 %a, i32 %b, i64 %c)  {
 }
 
 define i64 @MACLongTest3(i32 %a, i32 %b, i32 %c) {
-;CHECK: MACLongTest3:
+;CHECK-LABEL: MACLongTest3:
 ;CHECK: umlal
   %conv = zext i32 %b to i64
   %conv1 = zext i32 %a to i64
@@ -33,7 +36,7 @@ define i64 @MACLongTest3(i32 %a, i32 %b, i32 %c) {
 }
 
 define i64 @MACLongTest4(i32 %a, i32 %b, i32 %c) {
-;CHECK: MACLongTest4:
+;CHECK-LABEL: MACLongTest4:
 ;CHECK: smlal
   %conv = sext i32 %b to i64
   %conv1 = sext i32 %a to i64
@@ -42,3 +45,33 @@ define i64 @MACLongTest4(i32 %a, i32 %b, i32 %c) {
   %add = add nsw i64 %mul, %conv2
   ret i64 %add
 }
+
+; Two things to check here: the @earlyclobber constraint (on <= v5) and the "$Rd = $R" ones.
+;    + Without @earlyclobber the v7 code is natural. With it, the first two
+;      registers must be distinct from the third.
+;    + Without "$Rd = $R", this can be satisfied without a mov before the umlal
+;      by trying to use 6 different registers in the MachineInstr. The natural
+;      evolution of this attempt currently leaves only two movs in the final
+;      function, both after the umlal. With it, *some* move has to happen
+;      before the umlal.
+define i64 @MACLongTest5(i64 %c, i32 %a, i32 %b) {
+; CHECK-V7-LE-LABEL: MACLongTest5:
+; CHECK-V7-LE-LABEL: umlal r0, r1, r0, r0
+; CHECK-V7-BE-LABEL: MACLongTest5:
+; CHECK-V7-BE-LABEL: umlal r1, r0, r1, r1
+
+; CHECK-LABEL: MACLongTest5:
+; CHECK-LE: mov [[RDLO:r[0-9]+]], r0
+; CHECK-LE: umlal [[RDLO]], r1, r0, r0
+; CHECK-LE: mov r0, [[RDLO]]
+; CHECK-BE: mov [[RDLO:r[0-9]+]], r1
+; CHECK-BE: umlal [[RDLO]], r0, r1, r1
+; CHECK-BE: mov r1, [[RDLO]]
+
+  %conv.trunc = trunc i64 %c to i32
+  %conv = zext i32 %conv.trunc to i64
+  %conv1 = zext i32 %b to i64
+  %mul = mul i64 %conv, %conv
+  %add = add i64 %mul, %c
+  ret i64 %add
+}