From 0628d38085b28a59a4b13d7e35760cce54f0af7a Mon Sep 17 00:00:00 2001
From: Eric Christopher <echristo@apple.com>
Date: Tue, 24 May 2011 22:10:34 +0000
Subject: [PATCH] Add support for the arm 'y' asm modifier.

Fixes part of rdar://9444657


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132011 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/ARM/ARMAsmPrinter.cpp | 12 ++++++++++++
 test/CodeGen/ARM/arm-modifier.ll | 15 +++++++++++++++
 2 files changed, 27 insertions(+)
 create mode 100644 test/CodeGen/ARM/arm-modifier.ll

diff --git a/lib/Target/ARM/ARMAsmPrinter.cpp b/lib/Target/ARM/ARMAsmPrinter.cpp
index ff8f8a8a8e0..680680b2406 100644
--- a/lib/Target/ARM/ARMAsmPrinter.cpp
+++ b/lib/Target/ARM/ARMAsmPrinter.cpp
@@ -404,6 +404,18 @@ bool ARMAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNum,
     case 'q': // Print a NEON quad precision register.
       printOperand(MI, OpNum, O);
       return false;
+    case 'y': // Print a VFP single precision register as indexed double.
+      // This uses the ordering of the alias table to get the first 'd' register
+      // that overlaps the 's' register. Also, s0 is an odd register, hence the
+      // odd modulus check below.
+      if (MI->getOperand(OpNum).isReg()) {
+        unsigned Reg = MI->getOperand(OpNum).getReg();
+        const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
+        O << ARMInstPrinter::getRegisterName(TRI->getAliasSet(Reg)[0]) <<
+        (((Reg % 2) == 1) ? "[0]" : "[1]");
+        return false;
+      }
+      // Fallthrough to unsupported.
     case 'Q':
     case 'R':
     case 'H':
diff --git a/test/CodeGen/ARM/arm-modifier.ll b/test/CodeGen/ARM/arm-modifier.ll
new file mode 100644
index 00000000000..051d7e7780a
--- /dev/null
+++ b/test/CodeGen/ARM/arm-modifier.ll
@@ -0,0 +1,15 @@
+; RUN: llc < %s -march=arm -mattr=+vfp2
+
+define i32 @foo(float %scale, float %scale2) nounwind ssp {
+entry:
+  %scale.addr = alloca float, align 4
+  %scale2.addr = alloca float, align 4
+  store float %scale, float* %scale.addr, align 4
+  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
+  ret i32 0
+}
+
+!0 = metadata !{i32 56, i32 89, i32 128, i32 168}