58 lines · plain
1; RUN: llc -mtriple=amdgcn-- -o - %s | FileCheck %s2 3declare float @llvm.fma.f32(float, float, float)4 5; This checks that rematerialization support of the coalescer does not6; unnecessarily widen the register class. Without those fixes > 20 VGprs7; are used here8; Also check that some rematerialization of the 0 constant happened.9; CHECK-LABEL: foobar10; CHECK: v_mov_b32_e32 v{{[0-9]+}}, 011; CHECK: v_mov_b32_e32 v{{[0-9]+}}, 012; CHECK: v_mov_b32_e32 v{{[0-9]+}}, 013; CHECK: v_mov_b32_e32 v{{[0-9]+}}, 014; It's probably OK if this is slightly higher:15; CHECK: ; NumVgprs: 816define amdgpu_kernel void @foobar(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %flag) {17entry:18 %cmpflag = icmp eq i32 %flag, 119 br i1 %cmpflag, label %loop, label %exit20 21loop:22 %c = phi i32 [0, %entry], [%cnext, %loop]23 %v0 = phi float [0.0, %entry], [%fma.0, %loop]24 %v1 = phi float [0.0, %entry], [%fma.1, %loop]25 %v2 = phi float [0.0, %entry], [%fma.2, %loop]26 %v3 = phi float [0.0, %entry], [%fma.3, %loop]27 28 ; Try to get the 0 constant to get coalesced into a wide register29 %blup = insertelement <4 x float> poison, float %v0, i32 030 store <4 x float> %blup, ptr addrspace(1) %out31 32 %load = load <4 x float>, ptr addrspace(1) %in33 %load.0 = extractelement <4 x float> %load, i32 034 %load.1 = extractelement <4 x float> %load, i32 135 %load.2 = extractelement <4 x float> %load, i32 236 %load.3 = extractelement <4 x float> %load, i32 337 %fma.0 = call float @llvm.fma.f32(float %v0, float %load.0, float %v0)38 %fma.1 = call float @llvm.fma.f32(float %v1, float %load.1, float %v1)39 %fma.2 = call float @llvm.fma.f32(float %v2, float %load.2, float %v2)40 %fma.3 = call float @llvm.fma.f32(float %v3, float %load.3, float %v3)41 42 %cnext = add nsw i32 %c, 143 %cmp = icmp eq i32 %cnext, 4244 br i1 %cmp, label %exit, label %loop45 46exit:47 %ev0 = phi float [0.0, %entry], [%fma.0, %loop]48 %ev1 = phi float [0.0, %entry], [%fma.1, %loop]49 %ev2 = phi float [0.0, %entry], [%fma.2, %loop]50 %ev3 = phi float [0.0, %entry], [%fma.3, %loop]51 %dst.0 = insertelement <4 x float> poison, float %ev0, i32 052 %dst.1 = insertelement <4 x float> %dst.0, float %ev1, i32 153 %dst.2 = insertelement <4 x float> %dst.1, float %ev2, i32 254 %dst.3 = insertelement <4 x float> %dst.2, float %ev3, i32 355 store <4 x float> %dst.3, ptr addrspace(1) %out56 ret void57}58