From 10f786ca0ffdabfacd9ea60168488ee16d95c7e4 Mon Sep 17 00:00:00 2001 From: Stanislav Mekhanoshin Date: Tue, 16 Jul 2019 20:06:00 +0000 Subject: [PATCH] [AMDGPU] Change register type for v32 vectors 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. Differential Revision: https://reviews.llvm.org/D64815 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@366252 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/AMDGPU/SIISelLowering.cpp | 4 ++-- test/CodeGen/AMDGPU/v1024.ll | 29 ++++++++++++++++++++++++++++ 2 files changed, 31 insertions(+), 2 deletions(-) create mode 100644 test/CodeGen/AMDGPU/v1024.ll diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp index a3226577cd0..db0782e2bf3 100644 --- a/lib/Target/AMDGPU/SIISelLowering.cpp +++ b/lib/Target/AMDGPU/SIISelLowering.cpp @@ -152,8 +152,8 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM, } 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()); diff --git a/test/CodeGen/AMDGPU/v1024.ll b/test/CodeGen/AMDGPU/v1024.ll new file mode 100644 index 00000000000..a5e0454a363 --- /dev/null +++ b/test/CodeGen/AMDGPU/v1024.ll @@ -0,0 +1,29 @@ +; 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) -- 2.40.0