AMDGPU: Account for globals in AMDGPUPromoteAlloca pass

[arsenm]

lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

128–129

You probably want to use GV.users() instead

133

Missing space before {

test/CodeGen/AMDGPU/promote-alloca-globals.ll

2

You should also include an llc run line and check for the final LDS usage

4

This should be after a CHECK-LABEL for the function name and moved down to the function

10–12

Can you run opt -instnamer on the test

25

This GEP isn't used?

[bnieuwenhuizen]

test/CodeGen/AMDGPU/promote-alloca-globals.ll

2

How do you check for LDS usage? I can't find anything about LDS except the numerical representations of the RSRC registers.

25

It isn't as we just need a user of the global array to be used by an instruction inside the function to let the AMDGPUPromoteAlloca associate it with this function. I'll probably add a user though, to prevent llc from optimizing the LDS array away.

[• tstellarAMD]

test/CodeGen/AMDGPU/promote-alloca-globals.ll

2

It doesn't look like we emit this currently. You emit this information from the AMDGPUAsmPrinter.cpp with the rest of the Kernel info.

[arsenm]

I added the LDS size to the comments, so you should be able to check this easily now

[bnieuwenhuizen]

v2: - Use users.
    - Check LDS size in testcase.
    - Ran instnamer on testcase.
    - Changed unused GEP.
    - Enlarge array for 64 KiB LDS.

[arsenm]

LGTM

[arsenm]

r267791

lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

bool AMDGPUPromoteAlloca::runOnFunction(Function &F) {
   if (LocalMemAvailable == 0)
     return false;
 
   // Check how much local memory is being used by global objects
   for (GlobalVariable &GV : Mod->globals()) {
     if (GV.getType()->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS)
       continue;
 
-    for (Use &U : GV.uses()) {
+    for (User *U : GV.users()) {
       Instruction *Use = dyn_cast<Instruction>(U);
       if (!Use)
         continue;
 
-      if (Use->getParent()->getParent() == &F)
+      if (Use->getParent()->getParent() == &F) {
         LocalMemAvailable -=
           Mod->getDataLayout().getTypeAllocSize(GV.getValueType());
+        break;
+      }
     }
   }
 
   LocalMemAvailable = std::max(0, LocalMemAvailable);
   DEBUG(dbgs() << LocalMemAvailable << " bytes free in local memory.\n");
 
   BasicBlock &EntryBB = *F.begin();
   for (auto I = EntryBB.begin(), E = EntryBB.end(); I != E; ) {

test/CodeGen/AMDGPU/promote-alloca-globals.ll

+; RUN: opt -S -mtriple=amdgcn-unknown-unknown -amdgpu-promote-alloca < %s | FileCheck -check-prefix=IR %s
+; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=ASM %s
+
+
+@global_array= internal unnamed_addr addrspace(3) global [1500 x [10 x i32]] undef, align 4
+
+; IR-LABEL: define void @promote_alloca_size_256(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
+; IR: alloca [10 x i32]
+; ASM-LABEL: {{^}}promote_alloca_size_256:
+; ASM: ; LDSByteSize: 60000 bytes/workgroup (compile time only)
+
+define void @promote_alloca_size_256(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
+entry:
+  %stack = alloca [10 x i32], align 4
+  %tmp = load i32, i32 addrspace(1)* %in, align 4
+  %arrayidx1 = getelementptr inbounds [10 x i32], [10 x i32]* %stack, i32 0, i32 %tmp
+  store i32 4, i32* %arrayidx1, align 4
+  %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
+  %tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
+  %arrayidx3 = getelementptr inbounds [10 x i32], [10 x i32]* %stack, i32 0, i32 %tmp1
+  store i32 5, i32* %arrayidx3, align 4
+  %arrayidx10 = getelementptr inbounds [10 x i32], [10 x i32]* %stack, i32 0, i32 0
+  %tmp2 = load i32, i32* %arrayidx10, align 4
+  store i32 %tmp2, i32 addrspace(1)* %out, align 4
+  %arrayidx12 = getelementptr inbounds [10 x i32], [10 x i32]* %stack, i32 0, i32 1
+  %tmp3 = load i32, i32* %arrayidx12
+  %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
+  store i32 %tmp3, i32 addrspace(1)* %arrayidx13
+  %v = getelementptr inbounds [1500 x [10 x i32]], [1500 x [10 x i32]] addrspace(3)* @global_array, i32 0, i32 0, i32 0
+  store i32 %tmp3, i32 addrspace(3)* %v
+  ret void
+}