================
@@ -446,6 +512,96 @@ bool AMDGPUPromoteAllocaImpl::run(Function &F, bool
PromoteToLDS) {
return Changed;
}
+// Give an alloca in the VGPR ("as memory") address space a place there,
+// recorded as !amdgpu.allocated.vgprs, plus lifetime markers the backend
+// understands. AMDGPUPrivateObjectVGPRs turns those into defs and uses of the
+// physical registers, which keeps register allocation off them.
+void AMDGPUPromoteAllocaImpl::allocateVgprs(AllocaAnalysis &AA) {
+ LLVMContext &Ctx = Mod.getContext();
+ const unsigned AllocaSize =
+ DL.getTypeAllocSize(AA.Alloca->getAllocatedType()).getFixedValue();
+
+ // The generic lifetime intrinsics do not survive into the backend, so use
+ // the address-space specific ones. An object with no start is live from its
+ // definition.
+ //
+ // An already-converted marker counts as a start: this runs over IR it has
+ // converted before, and a second start would move the live range back to
+ // the alloca and keep the object live from there.
+ bool HaveLifetimeStart = false;
+ for (Use &U : AA.Alloca->uses()) {
+ auto *II = dyn_cast<IntrinsicInst>(U.getUser());
+ if (!II)
+ continue;
+ switch (II->getIntrinsicID()) {
+ case Intrinsic::lifetime_start:
+ II->setCalledFunction(Intrinsic::getOrInsertDeclaration(
+ &Mod, Intrinsic::amdgcn_vgpr_lifetime_start, AA.Alloca->getType()));
+ HaveLifetimeStart = true;
+ break;
+ case Intrinsic::lifetime_end:
+ II->setCalledFunction(Intrinsic::getOrInsertDeclaration(
+ &Mod, Intrinsic::amdgcn_vgpr_lifetime_end, AA.Alloca->getType()));
+ break;
+ case Intrinsic::amdgcn_vgpr_lifetime_start:
+ HaveLifetimeStart = true;
+ break;
+ default:
+ break;
+ }
+ }
+
+ if (!HaveLifetimeStart) {
+ Instruction *IP = AA.Alloca->getNextNode();
+ while (isa<AllocaInst>(IP))
+ IP = IP->getNextNode();
+ IRBuilder<> B(IP);
+ B.SetCurrentDebugLocation(AA.Alloca->getDebugLoc());
+ B.CreateIntrinsic(B.getVoidTy(), Intrinsic::amdgcn_vgpr_lifetime_start,
+ AA.Alloca);
+ }
+
+ // An object that already has its place keeps it, or a second allocation
+ // would land on top of it. The markers above are still rewritten, since IR
+ // carrying the metadata need not carry them.
+ unsigned Address = AllocVGPROffset;
+ if (MDNode *MD = AA.Alloca->getMetadata("amdgpu.allocated.vgprs")) {
+ Address = cast<AMDGPU::AllocatedVGPRsMetadata>(MD)->getAddress();
+ AllocVGPROffset =
+ std::max(AllocVGPROffset, Address + alignTo(AllocaSize, 4));
+ } else {
+ // Find space for the alloca. Objects are packed in order at dword
+ // granularity, since a register cannot be shared between two of them.
+ Type *I32 = Type::getInt32Ty(Ctx);
+ AA.Alloca->setMetadata(
+ "amdgpu.allocated.vgprs",
+ MDNode::get(
+ Ctx, {ConstantAsMetadata::get(ConstantInt::get(I32, Address)),
+ ConstantAsMetadata::get(ConstantInt::get(I32,
AllocaSize))}));
+ AllocVGPROffset += alignTo(AllocaSize, 4);
+ }
+
+ // Nothing may move or spill the object, so it has to fit the register
+ // budget this function is entitled to - for a non-entry function, the 32
----------------
perlfu wrote:
32 seems oddly specific.
https://github.com/llvm/llvm-project/pull/214248
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits