Implement the 'M' output modifier for arm inline asm. This is fairly

register allocation dependent and will occasionally break. WIP in the
register allocator to model paired/etc registers.

rdar://9119939


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132242 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 54 insertions, 12 deletions

diff --git a/lib/Target/ARM/ARMAsmPrinter.cpp b/lib/Target/ARM/ARMAsmPrinter.cpp
index 709d8c6e62..eb7390236d 100644
--- a/lib/Target/ARM/ARMAsmPrinter.cpp
+++ b/lib/Target/ARM/ARMAsmPrinter.cpp
@@ -387,16 +387,41 @@ bool ARMAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNum,
         return true;
       O << (MI->getOperand(OpNum).getImm() & 0xffff);
       return false;
+    case 'M': { // A register range suitable for LDM/STM.
+      if (!MI->getOperand(OpNum).isReg())
+        return true;
+      const MachineOperand &MO = MI->getOperand(OpNum);
+      unsigned RegBegin = MO.getReg();
+      // This takes advantage of the 2 operand-ness of ldm/stm and that we've
+      // already got the operands in registers that are operands to the
+      // inline asm statement.
+      
+      O << "{" << ARMInstPrinter::getRegisterName(RegBegin);
+      
+      // FIXME: The register allocator not only may not have given us the
+      // registers in sequence, but may not be in ascending registers. This
+      // will require changes in the register allocator that'll need to be
+      // propagated down here if the operands change.
+      unsigned RegOps = OpNum + 1;
+      while (MI->getOperand(RegOps).isReg()) {
+        O << ", " 
+          << ARMInstPrinter::getRegisterName(MI->getOperand(RegOps).getReg());
+        RegOps++;
+      }
+
+      O << "}";
+
+      return false;
+    }
+    // These modifiers are not yet supported.
     case 'p': // The high single-precision register of a VFP double-precision
               // register.
     case 'e': // The low doubleword register of a NEON quad register.
     case 'f': // The high doubleword register of a NEON quad register.
     case 'h': // A range of VFP/NEON registers suitable for VLD1/VST1.
-    case 'M': // A register range suitable for LDM/STM.
     case 'Q': // The least significant register of a pair.
     case 'R': // The most significant register of a pair.
     case 'H': // The highest-numbered register of a pair.
-      // These modifiers are not yet supported.
       return true;
     }
   }
diff --git a/test/CodeGen/ARM/arm-modifier.ll b/test/CodeGen/ARM/arm-modifier.ll
index 6dd3989728..0a7bb6cd21 100644
--- a/test/CodeGen/ARM/arm-modifier.ll
+++ b/test/CodeGen/ARM/arm-modifier.ll
@@ -1,6 +1,6 @@
 ; RUN: llc < %s -march=arm -mattr=+vfp2 | FileCheck %s
 
-define i32 @foo(float %scale, float %scale2) nounwind ssp {
+define i32 @foo(float %scale, float %scale2) nounwind {
 entry:
   %scale.addr = alloca float, align 4
   %scale2.addr = alloca float, align 4
@@ -8,35 +8,52 @@ entry:
   store float %scale2, float* %scale2.addr, align 4
   %tmp = load float* %scale.addr, align 4
   %tmp1 = load float* %scale2.addr, align 4
-  call void asm sideeffect "vmul.f32    q0, q0, ${0:y} \0A\09vmul.f32    q1, q1, ${0:y} \0A\09vmul.f32    q1, q0, ${1:y} \0A\09", "w,w,~{q0},~{q1}"(float %tmp, float %tmp1) nounwind, !srcloc !0
+  call void asm sideeffect "vmul.f32    q0, q0, ${0:y} \0A\09vmul.f32    q1, q1, ${0:y} \0A\09vmul.f32    q1, q0, ${1:y} \0A\09", "w,w,~{q0},~{q1}"(float %tmp, float %tmp1) nounwind
   ret i32 0
 }
 
-!0 = metadata !{i32 56, i32 89, i32 128, i32 168}
-
-define void @f0() nounwind ssp {
+define void @f0() nounwind {
 entry:
 ; CHECK: f0
 ; CHECK: .word -1
-call void asm sideeffect ".word ${0:B} \0A\09", "i"(i32 0) nounwind, !srcloc !0
+call void asm sideeffect ".word ${0:B} \0A\09", "i"(i32 0) nounwind
 ret void
 }
 
-define void @f1() nounwind ssp {
+define void @f1() nounwind {
 entry:
 ; CHECK: f1
 ; CHECK: .word 65535
-call void asm sideeffect ".word ${0:L} \0A\09", "i"(i32 -1) nounwind, !srcloc !0
+call void asm sideeffect ".word ${0:L} \0A\09", "i"(i32 -1) nounwind
 ret void
 }
 
 @f2_ptr = internal global i32* @f2_var, align 4
 @f2_var = external global i32
 
-define void @f2() nounwind ssp {
+define void @f2() nounwind {
 entry:
 ; CHECK: f2
 ; CHECK: ldr r0, [r{{[0-9]+}}]
-call void asm sideeffect "ldr r0, [${0:m}]\0A\09", "*m,~{r0}"(i32** @f2_ptr) nounwind, !srcloc !0
+call void asm sideeffect "ldr r0, [${0:m}]\0A\09", "*m,~{r0}"(i32** @f2_ptr) nounwind
+ret void
+}
+
+@f3_ptr = internal global i64* @f3_var, align 4
+@f3_var = external global i64
+@f3_var2 = external global i64
+
+define void @f3() nounwind {
+entry:
+; CHECK: f3
+; CHECK: stm r{{[0-9]+}}, {[[REG1:(r[0-9]+)]], r{{[0-9]+}}}
+; CHECK: adds lr, [[REG1]]
+; CHECK: ldm r{{[0-9]+}}, {r{{[0-9]+}}, r{{[0-9]+}}}
+%tmp = load i64* @f3_var, align 4
+%tmp1 = load i64* @f3_var2, align 4
+%0 = call i64 asm sideeffect "stm ${0:m}, ${1:M}\0A\09adds $3, $1\0A\09", "=*m,=r,1,r"(i64** @f3_ptr, i64 %tmp, i64 %tmp1) nounwind
+store i64 %0, i64* @f3_var, align 4
+%1 = call i64 asm sideeffect "ldm ${1:m}, ${0:M}\0A\09", "=r,*m"(i64** @f3_ptr) nounwind
+store i64 %1, i64* @f3_var, align 4
 ret void
 }