[AMDGPU] Change register type for v32 vectors

[rampitec]

Summary

When it is AReg_1024 this results in unnecessary copying into
AGPRs of a 32 element vectors even though they are not intended
for an mfma instruction.

[arsenm]

test/CodeGen/AMDGPU/v1024.ll

8 ↗

(On Diff #210152)

You can uses GCN-COUNT-<number>

[arsenm]

test/CodeGen/AMDGPU/v1024.ll

2 ↗

(On Diff #210152)

Can you add a comment explaining what this tests

[rampitec]

Updated test.

llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

if (Subtarget->has16BitInsts()) {
     // Unless there are also VOP3P operations, not operations are really legal.
     addRegisterClass(MVT::v2i16, &AMDGPU::SReg_32_XM0RegClass);
     addRegisterClass(MVT::v2f16, &AMDGPU::SReg_32_XM0RegClass);
     addRegisterClass(MVT::v4i16, &AMDGPU::SReg_64RegClass);
     addRegisterClass(MVT::v4f16, &AMDGPU::SReg_64RegClass);
   }
 
   if (Subtarget->hasMAIInsts()) {
-    addRegisterClass(MVT::v32i32, &AMDGPU::AReg_1024RegClass);
-    addRegisterClass(MVT::v32f32, &AMDGPU::AReg_1024RegClass);
+    addRegisterClass(MVT::v32i32, &AMDGPU::VReg_1024RegClass);
+    addRegisterClass(MVT::v32f32, &AMDGPU::VReg_1024RegClass);
   }
 
   computeRegisterProperties(Subtarget->getRegisterInfo());
 
   // We need to custom lower vector stores from local memory
   setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v3i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v4i32, Custom);

llvm/trunk/test/CodeGen/AMDGPU/v1024.ll

+; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; Check that we do not use AGPRs for v32i32 type
+
+; GCN-LABEL: {{^}}test_v1024:
+; GCN-NOT: v_accvgpr
+; GCN-COUNT-32: v_mov_b32_e32
+; GCN-NOT: v_accvgpr
+define amdgpu_kernel void @test_v1024() {
+entry:
+  %alloca = alloca <32 x i32>, align 16, addrspace(5)
+  %cast = bitcast <32 x i32> addrspace(5)* %alloca to i8 addrspace(5)*
+  br i1 undef, label %if.then.i.i, label %if.else.i
+
+if.then.i.i:                                      ; preds = %entry
+  call void @llvm.memcpy.p5i8.p5i8.i64(i8 addrspace(5)* align 16 %cast, i8 addrspace(5)* align 4 undef, i64 128, i1 false)
+  br label %if.then.i62.i
+
+if.else.i:                                        ; preds = %entry
+  br label %if.then.i62.i
+
+if.then.i62.i:                                    ; preds = %if.else.i, %if.then.i.i
+  call void @llvm.memcpy.p1i8.p5i8.i64(i8 addrspace(1)* align 4 undef, i8 addrspace(5)* align 16 %cast, i64 128, i1 false)
+  ret void
+}
+
+declare void @llvm.memcpy.p5i8.p5i8.i64(i8 addrspace(5)* nocapture writeonly, i8 addrspace(5)* nocapture readonly, i64, i1 immarg)
+
+declare void @llvm.memcpy.p1i8.p5i8.i64(i8 addrspace(1)* nocapture writeonly, i8 addrspace(5)* nocapture readonly, i64, i1 immarg)