[ARM] Add support for long-call/short-call attribute

[weimingz]

Summary

GNU has long_call/short_call attribute support for ARM: https://gcc.gnu.org/onlinedocs/gcc-6.1.0/gcc/ARM-Function-Attributes.html#ARM-Function-Attributes.

LLVM/clang only supports the global switch/option (-mlong-call -mno-long-call). In some scenarios, we need to turn it on/off for individual functions.

Clang patch: http://reviews.llvm.org/D19610

[t.p.northover]

Hi Weiming,

I've got two issues with the current implementation:

First, I think this should be an attribute on the call-site rather than the callee. You could easily imagine (particularly in LTO situations targeting weird embedded memory layouts) mutually short-callable clusters of functions that need long-calls between those groups. If the attribute is fixed on the callee that's not going to work.

Second, I don't really think "target-features" is the place for this attribute. They're meant to affect the CodeGen of the function definition, not really what callees do with it as far as I can see.

Cheers.

Tim.

[t.p.northover]

Hmm, on the other hand annotating the call-site isn't going to work for indirect calls that become apparent after optimization. Annotating the callee's declaration looks like it happens to work in that case.

[weimingz]

In D19609#414292, @t.p.northover wrote:

Hi Weiming,

I've got two issues with the current implementation:

First, I think this should be an attribute on the call-site rather than the callee. You could easily imagine (particularly in LTO situations targeting weird embedded memory layouts) mutually short-callable clusters of functions that need long-calls between those groups. If the attribute is fixed on the callee that's not going to work.

Second, I don't really think "target-features" is the place for this attribute. They're meant to affect the CodeGen of the function definition, not really what callees do with it as far as I can see.

Cheers.

Tim.

Hi Tim,

Yes, we should let clang to carry the info on call site. (via metadata?)
If we do so, we don't need to worry about placing it into target-features.

[weimingz]

In D19609#414304, @t.p.northover wrote:

Hmm, on the other hand annotating the call-site isn't going to work for indirect calls that become apparent after optimization. Annotating the callee's declaration looks like it happens to work in that case.

Should we still use taget-features to hold the flag?

[weimingz]

For LTO scenario, if we annotate callee declaration, when they are merged from different modules, we can merge them conservatively.

lib/Target/ARM/ARMISelLowering.cpp

ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
   // node so that legalize doesn't hack it.
   bool isDirect = false;
   bool isARMFunc = false;
   bool isLocalARMFunc = false;
   ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
   auto PtrVt = getPointerTy(DAG.getDataLayout());
 
-  if (Subtarget->genLongCalls()) {
+  bool isCalleeLongCall = false;
+  bool isCalleeShortCall = false;
+  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
+    const Function *CalleeF = dyn_cast<Function>(G->getGlobal());
+    if (CalleeF) {
+      StringRef Features =
+          CalleeF->getFnAttribute("target-features").getValueAsString();
+      SmallVector<StringRef, 3> Tmp;
+      Features.split(Tmp, ',');
+      isCalleeLongCall =
+          std::find(Tmp.begin(), Tmp.end(), "+long-calls") != Tmp.end();
+      isCalleeShortCall =
+          std::find(Tmp.begin(), Tmp.end(), "-long-calls") != Tmp.end();
+    }
+  }
+
+  // Generate long call if (1) callee is not short call and (2) either callee
+  // or caller has long-call attr.
+  if (!isCalleeShortCall && (isCalleeLongCall || Subtarget->genLongCalls())) {
     assert((Subtarget->isTargetWindows() ||
             getTargetMachine().getRelocationModel() == Reloc::Static) &&
            "long-calls with non-static relocation model!");
     // Handle a global address or an external symbol. If it's not one of
     // those, the target's already in a register, so we don't need to do
     // anything extra.
     if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
       const GlobalValue *GV = G->getGlobal();

test/CodeGen/ARM/long-calls.ll

+; RUN: llc -mtriple=thumbv7-linux-gnueabi -relocation-model=static %s -O0 -o - | FileCheck %s
+target triple = "armv4t--linux-gnueabi"
+
+define i32 @caller_long() #0 {
+; CHECK-LABEL: caller_long
+entry:
+  call void @callee0()
+; CHECK: ldr [[R0:r[0-9]+]], [[L0:.*]]
+; CHECK: blx [[R0]]
+  call void @callee_long()
+; CHECK: ldr [[R1:r[0-9]+]], [[L1:.*]]
+; CHECK: blx [[R1]]
+  tail call void @callee_short()
+; CHECK: bl callee_short
+  ret i32 0
+; CHECK: [[L0]]:
+; CHECK: .long {{_?}}callee0
+; CHECK: [[L1]]:
+; CHECK: .long {{_?}}callee_long
+}
+
+define i32 @caller1() {
+; CHECK-LABEL: caller1
+entry:
+  call void @callee0()
+; CHECK: bl callee0
+  call void @callee_long()
+; CHECK: ldr{{.*}} [[R2:.+]], [[L2:.*]]
+; CHECK: blx [[R2]]
+  tail call void @callee_short()
+; CHECK: bl callee_short
+  ret i32 0
+; CHECK: [[L2]]:
+; CHECK: .long callee_long
+}
+
+declare void @callee0()
+
+declare void @callee_long() #0
+
+declare void @callee_short() #1
+
+attributes #0 = { "target-features"="+long-calls" }
+attributes #1 = { "target-features"="-long-calls" }