[AArch64] Mark a move of system register TPIDR_EL0 (thread pointer) as not having side effects.

[mcrosier]

Summary

This allows Machine LICM to hoist the invariant 'mrs xX, TPIDR_EL0' instruction from within a loop.

Chad

[gberry]

'TP' is a little terse for a name. How about 'MOVbaseTLS' to match with 'MOVaddrTLS'?
Could you also add a test where MachineCSE commons multiple 'mrs' from multiple thread_local loads into a single one?

[mcrosier]

-Address Geoff's comments.

[mcrosier]

Ping. It would be great to get some feedback from other sub-target owners. FWIW, this is the current behavior of gcc, AFAICT.

[t.p.northover]

The idea looks fine to me, though I'd suggest including Utils/AArch64BaseInfo.h and using the TPIDR_EL0 enum instead of 0xde82 directly (from AArch64GenSystemOperands.inc).

[mcrosier]

-Address Tim's feedback.

[mcrosier]

Approving, per Tim's comment.

llvm/trunk/lib/Target/AArch64/AArch64ExpandPseudoInsts.cpp

 // pass should be run after register allocation but before the post-regalloc
 // scheduling pass.
 //
 //===----------------------------------------------------------------------===//
 
 #include "MCTargetDesc/AArch64AddressingModes.h"
 #include "AArch64InstrInfo.h"
 #include "AArch64Subtarget.h"
+#include "Utils/AArch64BaseInfo.h"
 #include "llvm/CodeGen/LivePhysRegs.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/Support/MathExtras.h"
 using namespace llvm;
 
 #define AARCH64_EXPAND_PSEUDO_NAME "AArch64 pseudo instruction expansion pass"
 
MachineInstrBuilder MIB2 =
             .addReg(DstReg)
             .add(MI.getOperand(2))
             .addImm(0);
 
     transferImpOps(MI, MIB1, MIB2);
     MI.eraseFromParent();
     return true;
   }
+  case AArch64::MOVbaseTLS: {
+    unsigned DstReg = MI.getOperand(0).getReg();
+    BuildMI(MBB, MBBI, MI.getDebugLoc(), TII->get(AArch64::MRS), DstReg)
+        .addImm(AArch64SysReg::TPIDR_EL0);
+    MI.eraseFromParent();
+    return true;
+  }
 
   case AArch64::MOVi32imm:
     return expandMOVImm(MBB, MBBI, 32);
   case AArch64::MOVi64imm:
     return expandMOVImm(MBB, MBBI, 64);
   case AArch64::RET_ReallyLR: {
     // Hiding the LR use with RET_ReallyLR may lead to extra kills in the
     // function and missing live-ins. We are fine in practice because callee

llvm/trunk/lib/Target/AArch64/AArch64InstrInfo.td

 def : InstAlias<"isb", (ISB 0xf)>;
 
 def MRS    : MRSI;
 def MSR    : MSRI;
 def MSRpstateImm1 : MSRpstateImm0_1;
 def MSRpstateImm4 : MSRpstateImm0_15;
 
 // The thread pointer (on Linux, at least, where this has been implemented) is
-// TPIDR_EL0.
-def : Pat<(AArch64threadpointer), (MRS 0xde82)>;
+// TPIDR_EL0.  Add pseudo op so we can mark it as not having any side effects.
+let hasSideEffects = 0 in
+def MOVbaseTLS : Pseudo<(outs GPR64:$dst), (ins),
+                       [(set GPR64:$dst, AArch64threadpointer)]>, Sched<[]>;
 
 // The cycle counter PMC register is PMCCNTR_EL0.
 let Predicates = [HasPerfMon] in
 def : Pat<(readcyclecounter), (MRS 0xdce8)>;
 
 // Generic system instructions
 def SYSxt  : SystemXtI<0, "sys">;
 def SYSLxt : SystemLXtI<1, "sysl">;

llvm/trunk/test/CodeGen/AArch64/thread-pointer.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -verify-machineinstrs -o - %s | FileCheck %s
+
+@x = thread_local local_unnamed_addr global i32 0, align 4
+@y = thread_local local_unnamed_addr global i32 0, align 4
+
+; Machine LICM should hoist the mrs into the loop preheader.
+; CHECK-LABEL: @test1
+; CHECK: BB#1:
+; CHECK:   mrs x[[BASE:[0-9]+]], TPIDR_EL0
+; CHECK:   add x[[REG1:[0-9]+]], x[[BASE]], :tprel_hi12:x
+; CHECK:   add x[[REG2:[0-9]+]], x[[REG1]], :tprel_lo12_nc:x
+;
+; CHECK: .LBB0_2:
+; CHECK:   ldr w0, [x[[REG2]]]
+; CHECK:   bl bar
+; CHECK:   sub w[[REG3:[0-9]+]], w{{[0-9]+}}, #1
+; CHECK:   cbnz w[[REG3]], .LBB0_2
+
+define void @test1(i32 %n) local_unnamed_addr {
+entry:
+  %cmp3 = icmp sgt i32 %n, 0
+  br i1 %cmp3, label %bb1, label %bb2
+
+bb1:
+  br label %for.body
+
+for.body:
+  %i.04 = phi i32 [ %inc, %for.body ], [ 0, %bb1 ]
+  %0 = load i32, i32* @x, align 4
+  tail call void @bar(i32 %0) #2
+  %inc = add nuw nsw i32 %i.04, 1
+  %exitcond = icmp eq i32 %inc, %n
+  br i1 %exitcond, label %bb2, label %for.body
+
+bb2:
+  ret void
+}
+
+; Machine CSE should combine the the mrs between the load of %x and %y.
+; CHECK-LABEL: @test2
+; CHECK: mrs x{{[0-9]+}}, TPIDR_EL0
+; CHECK-NOT: mrs x{{[0-9]+}}, TPIDR_EL0
+; CHECK: ret
+define void @test2(i32 %c) local_unnamed_addr #0 {
+entry:
+  %0 = load i32, i32* @x, align 4
+  tail call void @bar(i32 %0) #2
+  %cmp = icmp eq i32 %c, 0
+  br i1 %cmp, label %if.end, label %if.then
+
+if.then:
+  %1 = load i32, i32* @y, align 4
+  tail call void @bar(i32 %1) #2
+  br label %if.end
+
+if.end:
+  ret void
+}
+
+declare void @bar(i32) local_unnamed_addr