[NVPTX] Annotate call machine instructions as calls.

[jlebar]

Summary

Otherwise we'll try to do unsafe optimizations on these MIs, such as
sinking loads below calls.

(I suspect that this is not the only bug in the NVPTX instruction
tablegen files; I need to comb through them.)

[tra]

Looks good to my unqualified eye.

[jholewinski]

Looks reasonable to me.

lib/Target/NVPTX/NVPTXInstrInfo.td

 class StoreRetvalV4Inst<NVPTXRegClass regclass, string opstr> :
       NVPTXInst<(outs),
                 (ins regclass:$val, regclass:$val2, regclass:$val3,
                      regclass:$val4, i32imm:$a),
                 !strconcat(!strconcat("st.param.v4", opstr),
                 "\t[func_retval0+$a], {{$val, $val2, $val3, $val4}};"),
                 []>;
 
+let isCall = 1 in {
 def PrintCallRetInst1 : NVPTXInst<(outs), (ins),
 "call (retval0), ",
                                 [(PrintCall (i32 1))]>;
 def PrintCallRetInst2 : NVPTXInst<(outs), (ins),
 "call (retval0, retval1), ",
                                 [(PrintCall (i32 2))]>;
 def PrintCallRetInst3 : NVPTXInst<(outs), (ins),
 "call (retval0, retval1, retval2), ",
"call.uni (retval0, retval1, retval2, retval3, retval4, retval5, retval6), ",
                                 [(PrintCallUni (i32 7))]>;
 def PrintCallUniRetInst8 : NVPTXInst<(outs), (ins),
 !strconcat("call.uni (retval0, retval1, retval2, retval3, retval4",
            ", retval5, retval6, retval7), "),
                                 [(PrintCallUni (i32 8))]>;
 
 def PrintCallUniNoRetInst : NVPTXInst<(outs), (ins), "call.uni ",
                                 [(PrintCallUni (i32 0))]>;
+} // call instructions
 
 def LoadParamMemI64    : LoadParamMemInst<Int64Regs, ".b64">;
 def LoadParamMemI32    : LoadParamMemInst<Int32Regs, ".b32">;
 def LoadParamMemI16    : LoadParamMemInst<Int16Regs, ".b16">;
 def LoadParamMemI8     : LoadParamMemInst<Int16Regs, ".b8">;
 def LoadParamMemV2I64  : LoadParamV2MemInst<Int64Regs, ".b64">;
 def LoadParamMemV2I32  : LoadParamV2MemInst<Int32Regs, ".b32">;
 def LoadParamMemV2I16  : LoadParamV2MemInst<Int16Regs, ".b16">;

test/CodeGen/NVPTX/MachineSink-call.ll

+; RUN: llc < %s | FileCheck %s
+target triple = "nvptx64-nvidia-cuda"
+
+declare void @foo()
+
+; Load a value, then call a function.  Branch, and use the loaded value only on
+; one side of the branch.  The load shouldn't be sunk beneath the call, because
+; the call may modify memory.
+define i32 @f(i32 %x, i32* %ptr, i1 %cond) {
+Start:
+  ; CHECK: ld.u32
+  %ptr_val = load i32, i32* %ptr
+  ; CHECK: call.uni
+  call void @foo()
+  br i1 %cond, label %L1, label %L2
+L1:
+  %ptr_val2 = add i32 %ptr_val, 100
+  br label %L2
+L2:
+  %v4 = phi i32 [ %x, %Start ], [ %ptr_val2, %L1 ]
+  %v5 = add i32 %v4, 1000
+  ret i32 %v5
+}