Fix pr6111: Avoid using the LR register for the target address of an indirect

branch in ARM v4 code, since it gets clobbered by the return address before
it is used.  Instead of adding a new register class containing all the GPRs
except LR, just use the existing tGPR class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96360 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 6fe7c2c8c6a400ab9f4f61fa6a8a7e645b1c3f7c..7b35a06b945eaf399ae9f31b173a516d88b90af3 100644 (file)
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -643,6 +643,13 @@ ARMBaseInstrInfo::copyRegToReg(MachineBasicBlock &MBB,
   DebugLoc DL = DebugLoc::getUnknownLoc();
   if (I != MBB.end()) DL = I->getDebugLoc();
 
+  // tGPR is used sometimes in ARM instructions that need to avoid using
+  // certain registers.  Just treat it as GPR here.
+  if (DestRC == ARM::tGPRRegisterClass)
+    DestRC = ARM::GPRRegisterClass;
+  if (SrcRC == ARM::tGPRRegisterClass)
+    SrcRC = ARM::GPRRegisterClass;
+
   if (DestRC != SrcRC) {
     if (DestRC->getSize() != SrcRC->getSize())
       return false;

diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index 1c6f78ac8bfb30cb8a76e21d30398ecdcaddc088..a021388994f9f49cd1fc1eede4aa558b4dfb4c9a 100644 (file)
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -828,9 +828,10 @@ let isCall = 1,
   }
 
   // ARMv4T
-  def BX : ABXIx2<(outs), (ins GPR:$func, variable_ops),
+  // Note: Restrict $func to the tGPR regclass to prevent it being in LR.
+  def BX : ABXIx2<(outs), (ins tGPR:$func, variable_ops),
                   IIC_Br, "mov\tlr, pc\n\tbx\t$func",
-                  [(ARMcall_nolink GPR:$func)]>,
+                  [(ARMcall_nolink tGPR:$func)]>,
            Requires<[IsARM, IsNotDarwin]> {
     let Inst{7-4}   = 0b0001;
     let Inst{19-8}  = 0b111111111111;
@@ -865,9 +866,10 @@ let isCall = 1,
   }
 
   // ARMv4T
-  def BXr9 : ABXIx2<(outs), (ins GPR:$func, variable_ops),
+  // Note: Restrict $func to the tGPR regclass to prevent it being in LR.
+  def BXr9 : ABXIx2<(outs), (ins tGPR:$func, variable_ops),
                   IIC_Br, "mov\tlr, pc\n\tbx\t$func",
-                  [(ARMcall_nolink GPR:$func)]>, Requires<[IsARM, IsDarwin]> {
+                  [(ARMcall_nolink tGPR:$func)]>, Requires<[IsARM, IsDarwin]> {
     let Inst{7-4}   = 0b0001;
     let Inst{19-8}  = 0b111111111111;
     let Inst{27-20} = 0b00010010;

diff --git a/test/CodeGen/ARM/call.ll b/test/CodeGen/ARM/call.ll
index 3dd66ae71df86863acb9876afbec0111cd55bf0f..c60b75b574a26f33e2558e177b97379b40115177 100644 (file)
--- a/test/CodeGen/ARM/call.ll
+++ b/test/CodeGen/ARM/call.ll
@@ -20,3 +20,17 @@ define void @g.upgrd.1() {
         %tmp.upgrd.2 = tail call i32 %tmp( )            ; <i32> [#uses=0]
         ret void
 }
+
+define i32* @m_231b(i32, i32, i32*, i32*, i32*) nounwind {
+; CHECKV4: m_231b
+; CHECKV4: bx r{{.*}}
+BB0:
+  %5 = inttoptr i32 %0 to i32*                    ; <i32*> [#uses=1]
+  %t35 = volatile load i32* %5                    ; <i32> [#uses=1]
+  %6 = inttoptr i32 %t35 to i32**                 ; <i32**> [#uses=1]
+  %7 = getelementptr i32** %6, i32 86             ; <i32**> [#uses=1]
+  %8 = load i32** %7                              ; <i32*> [#uses=1]
+  %9 = bitcast i32* %8 to i32* (i32, i32*, i32, i32*, i32*, i32*)* ; <i32* (i32, i32*, i32, i32*, i32*, i32*)*> [#uses=1]
+  %10 = call i32* %9(i32 %0, i32* null, i32 %1, i32* %2, i32* %3, i32* %4) ; <i32*> [#uses=1]
+  ret i32* %10
+}