205 lines · plain
1; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri -passes=amdgpu-promote-alloca < %s | FileCheck %s2 3 4; CHECK-LABEL: @branch_ptr_var_same_alloca(5; CHECK: [[GEP:%[0-9]+]] = getelementptr inbounds [256 x [64 x i32]], ptr addrspace(3) @branch_ptr_var_same_alloca.alloca, i32 0, i32 %{{[0-9]+}}6 7; CHECK: if:8; CHECK: %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(3) [[GEP]], i32 0, i32 %a9 10; CHECK: else:11; CHECK: %arrayidx1 = getelementptr inbounds [64 x i32], ptr addrspace(3) [[GEP]], i32 0, i32 %b12 13; CHECK: endif:14; CHECK: %phi.ptr = phi ptr addrspace(3) [ %arrayidx0, %if ], [ %arrayidx1, %else ]15; CHECK: store i32 0, ptr addrspace(3) %phi.ptr, align 416define amdgpu_kernel void @branch_ptr_var_same_alloca(i32 %a, i32 %b, i1 %c0) #0 {17entry:18 %alloca = alloca [64 x i32], align 4, addrspace(5)19 br i1 %c0, label %if, label %else20 21if:22 %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a23 br label %endif24 25else:26 %arrayidx1 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %b27 br label %endif28 29endif:30 %phi.ptr = phi ptr addrspace(5) [ %arrayidx0, %if ], [ %arrayidx1, %else ]31 store i32 0, ptr addrspace(5) %phi.ptr, align 432 ret void33}34 35; CHECK-LABEL: @branch_ptr_phi_alloca_null_0(36; CHECK: %phi.ptr = phi ptr addrspace(3) [ %arrayidx0, %if ], [ null, %entry ]37define amdgpu_kernel void @branch_ptr_phi_alloca_null_0(i32 %a, i32 %b, i1 %c0) #0 {38entry:39 %alloca = alloca [64 x i32], align 4, addrspace(5)40 br i1 %c0, label %if, label %endif41 42if:43 %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a44 br label %endif45 46endif:47 %phi.ptr = phi ptr addrspace(5) [ %arrayidx0, %if ], [ null, %entry ]48 store i32 0, ptr addrspace(5) %phi.ptr, align 449 ret void50}51 52; CHECK-LABEL: @branch_ptr_phi_alloca_null_1(53; CHECK: %phi.ptr = phi ptr addrspace(3) [ null, %entry ], [ %arrayidx0, %if ]54define amdgpu_kernel void @branch_ptr_phi_alloca_null_1(i32 %a, i32 %b, i1 %c0) #0 {55entry:56 %alloca = alloca [64 x i32], align 4, addrspace(5)57 br i1 %c0, label %if, label %endif58 59if:60 %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a61 br label %endif62 63endif:64 %phi.ptr = phi ptr addrspace(5) [ null, %entry ], [ %arrayidx0, %if ]65 store i32 0, ptr addrspace(5) %phi.ptr, align 466 ret void67}68 69; CHECK-LABEL: @one_phi_value(70; CHECK: [[GEP0:%[0-9]+]] = getelementptr inbounds [256 x [64 x i32]], ptr addrspace(3) @one_phi_value.alloca, i32 0, i32 %{{[0-9]+}}71; CHECK: %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(3) [[GEP0]], i32 0, i32 %a72 73; CHECK: br label %exit74; CHECK: %phi.ptr = phi ptr addrspace(3) [ %arrayidx0, %entry ]75; CHECK: store i32 0, ptr addrspace(3) %phi.ptr, align 476define amdgpu_kernel void @one_phi_value(i32 %a) #0 {77entry:78 %alloca = alloca [64 x i32], align 4, addrspace(5)79 %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a80 br label %exit81 82exit:83 %phi.ptr = phi ptr addrspace(5) [ %arrayidx0, %entry ]84 store i32 0, ptr addrspace(5) %phi.ptr, align 485 ret void86}87 88; CHECK-LABEL: @branch_ptr_alloca_unknown_obj(89; CHECK: %alloca = alloca [64 x i32], align 490 91; CHECK: if:92; CHECK: %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a93 94; CHECK: else:95; CHECK: %arrayidx1 = call ptr addrspace(5) @get_unknown_pointer()96 97; CHECK: endif:98; CHECK: %phi.ptr = phi ptr addrspace(5) [ %arrayidx0, %if ], [ %arrayidx1, %else ]99; CHECK: store i32 0, ptr addrspace(5) %phi.ptr, align 4100define amdgpu_kernel void @branch_ptr_alloca_unknown_obj(i32 %a, i32 %b, i1 %c0) #0 {101entry:102 %alloca = alloca [64 x i32], align 4, addrspace(5)103 br i1 %c0, label %if, label %else104 105if:106 %arrayidx0 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 %a107 br label %endif108 109else:110 %arrayidx1 = call ptr addrspace(5) @get_unknown_pointer()111 br label %endif112 113endif:114 %phi.ptr = phi ptr addrspace(5) [ %arrayidx0, %if ], [ %arrayidx1, %else ]115 store i32 0, ptr addrspace(5) %phi.ptr, align 4116 ret void117}118 119; kernel void ptr_induction_var_same_alloca(void)120; {121; int alloca[64];122; int i = 0;123 124; #pragma nounroll125; for (int* p = &alloca[2], *e = &alloca[48]; p != e; ++p, ++i)126; {127; *p = i;128; }129; }130 131; FIXME: This should be promotable. We need to use132; getUnderlyingObjects when looking at the icmp user.133 134; CHECK-LABEL: @ptr_induction_var_same_alloca(135; CHECK: %alloca = alloca [64 x i32], align 4136; CHECK: phi ptr addrspace(5) [ %arrayidx, %entry ], [ %incdec.ptr, %for.body ]137define amdgpu_kernel void @ptr_induction_var_same_alloca() #0 {138entry:139 %alloca = alloca [64 x i32], align 4, addrspace(5)140 %arrayidx = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 2141 %arrayidx1 = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 48142 br label %for.body143 144for.cond.cleanup: ; preds = %for.body145 ret void146 147for.body: ; preds = %for.body, %entry148 %i.09 = phi i32 [ 0, %entry ], [ %inc, %for.body ]149 %p.08 = phi ptr addrspace(5) [ %arrayidx, %entry ], [ %incdec.ptr, %for.body ]150 store i32 %i.09, ptr addrspace(5) %p.08, align 4151 %incdec.ptr = getelementptr inbounds i32, ptr addrspace(5) %p.08, i32 1152 %inc = add nuw nsw i32 %i.09, 1153 %cmp = icmp eq ptr addrspace(5) %incdec.ptr, %arrayidx1154 br i1 %cmp, label %for.cond.cleanup, label %for.body155}156 157 158; extern int* get_unknown_pointer(void);159 160; kernel void ptr_induction_var_alloca_unknown(void)161; {162; int alloca[64];163; int i = 0;164;165; for (int* p = &alloca[2], *e = get_unknown_pointer(); p != e; ++p, ++i)166; {167; *p = i;168; }169; }170 171; CHECK-LABEL: @ptr_induction_var_alloca_unknown(172; CHECK: %alloca = alloca [64 x i32], align 4173; CHECK: %p.08 = phi ptr addrspace(5) [ %incdec.ptr, %for.body ], [ %arrayidx, %for.body.preheader ]174; CHECK: %cmp = icmp eq ptr addrspace(5) %incdec.ptr, %call175define amdgpu_kernel void @ptr_induction_var_alloca_unknown() #0 {176entry:177 %alloca = alloca [64 x i32], align 4, addrspace(5)178 %arrayidx = getelementptr inbounds [64 x i32], ptr addrspace(5) %alloca, i32 0, i32 2179 %call = tail call ptr addrspace(5) @get_unknown_pointer() #2180 %cmp.7 = icmp eq ptr addrspace(5) %arrayidx, %call181 br i1 %cmp.7, label %for.cond.cleanup, label %for.body.preheader182 183for.body.preheader: ; preds = %entry184 br label %for.body185 186for.cond.cleanup.loopexit: ; preds = %for.body187 br label %for.cond.cleanup188 189for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry190 ret void191 192for.body: ; preds = %for.body, %for.body.preheader193 %i.09 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]194 %p.08 = phi ptr addrspace(5) [ %incdec.ptr, %for.body ], [ %arrayidx, %for.body.preheader ]195 store i32 %i.09, ptr addrspace(5) %p.08, align 4196 %incdec.ptr = getelementptr inbounds i32, ptr addrspace(5) %p.08, i32 1197 %inc = add nuw nsw i32 %i.09, 1198 %cmp = icmp eq ptr addrspace(5) %incdec.ptr, %call199 br i1 %cmp, label %for.cond.cleanup.loopexit, label %for.body200}201 202declare ptr addrspace(5) @get_unknown_pointer() #0203 204attributes #0 = { nounwind "amdgpu-waves-per-eu"="1,1" "amdgpu-flat-work-group-size"="1,256" }205