xref: /llvm-project/llvm/test/CodeGen/NVPTX/dynamic_stackalloc.ll (revision 0068078dca60b41ad1c7bdd4448e7de718b82a5d)
1b279f6b0SFangrui Song; RUN: not llc < %s -mtriple=nvptx -mattr=+ptx72 -mcpu=sm_52 2>&1 | FileCheck %s --check-prefixes=CHECK-FAILS
2b279f6b0SFangrui Song; RUN: not llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_50 2>&1 | FileCheck %s --check-prefixes=CHECK-FAILS
3f9304974SYoungsuk Kim
4b279f6b0SFangrui Song; RUN: llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_52 | FileCheck %s --check-prefixes=CHECK,CHECK-32
5b279f6b0SFangrui Song; RUN: llc < %s -mtriple=nvptx64 -mattr=+ptx73 -mcpu=sm_52 | FileCheck %s --check-prefixes=CHECK,CHECK-64
6b279f6b0SFangrui Song; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_52 | %ptxas-verify %}
7b279f6b0SFangrui Song; RUN: %if ptxas %{ llc < %s -mtriple=nvptx64 -mattr=+ptx73 -mcpu=sm_52 | %ptxas-verify %}
8f9304974SYoungsuk Kim
989b7b3b9SAlex MacLean; CHECK-FAILS: in function test_dynamic_stackalloc{{.*}}: Support for dynamic alloca introduced in PTX ISA version 7.3 and requires target sm_52.
1089b7b3b9SAlex MacLean
1189b7b3b9SAlex MacLean; CHECK-LABEL: .visible .func  (.param .b32 func_retval0) test_dynamic_stackalloc(
1289b7b3b9SAlex MacLean; CHECK-NOT: __local_depot
1389b7b3b9SAlex MacLean
1489b7b3b9SAlex MacLean; CHECK-32:       ld.param.u32  %r[[SIZE:[0-9]]], [test_dynamic_stackalloc_param_0];
15*0068078dSpeterbell10; CHECK-32-NEXT:  add.s32 %r[[SIZE2:[0-9]]], %r[[SIZE]], 7;
1689b7b3b9SAlex MacLean; CHECK-32-NEXT:  and.b32         %r[[SIZE3:[0-9]]], %r[[SIZE2]], -8;
1789b7b3b9SAlex MacLean; CHECK-32-NEXT:  alloca.u32  %r[[ALLOCA:[0-9]]], %r[[SIZE3]], 16;
1889b7b3b9SAlex MacLean; CHECK-32-NEXT:  cvta.local.u32  %r[[ALLOCA]], %r[[ALLOCA]];
1989b7b3b9SAlex MacLean; CHECK-32-NEXT:  { // callseq 0, 0
2089b7b3b9SAlex MacLean; CHECK-32-NEXT:  .param .b32 param0;
210f0a96b8SYoungsuk Kim; CHECK-32-NEXT:  st.param.b32  [param0], %r[[ALLOCA]];
2289b7b3b9SAlex MacLean
2389b7b3b9SAlex MacLean; CHECK-64:       ld.param.u64  %rd[[SIZE:[0-9]]], [test_dynamic_stackalloc_param_0];
2489b7b3b9SAlex MacLean; CHECK-64-NEXT:  add.s64 %rd[[SIZE2:[0-9]]], %rd[[SIZE]], 7;
2589b7b3b9SAlex MacLean; CHECK-64-NEXT:  and.b64 %rd[[SIZE3:[0-9]]], %rd[[SIZE2]], -8;
2689b7b3b9SAlex MacLean; CHECK-64-NEXT:  alloca.u64  %rd[[ALLOCA:[0-9]]], %rd[[SIZE3]], 16;
2789b7b3b9SAlex MacLean; CHECK-64-NEXT:  cvta.local.u64  %rd[[ALLOCA]], %rd[[ALLOCA]];
2889b7b3b9SAlex MacLean; CHECK-64-NEXT:  { // callseq 0, 0
2989b7b3b9SAlex MacLean; CHECK-64-NEXT:  .param .b64 param0;
300f0a96b8SYoungsuk Kim; CHECK-64-NEXT:  st.param.b64  [param0], %rd[[ALLOCA]];
3189b7b3b9SAlex MacLean
3289b7b3b9SAlex MacLean; CHECK-NEXT:     .param .b32 retval0;
3389b7b3b9SAlex MacLean; CHECK-NEXT:     call.uni (retval0),
3489b7b3b9SAlex MacLean; CHECK-NEXT:     bar,
3589b7b3b9SAlex MacLean
3689b7b3b9SAlex MacLeandefine i32 @test_dynamic_stackalloc(i64 %n) {
3789b7b3b9SAlex MacLean  %alloca = alloca i8, i64 %n, align 16
3889b7b3b9SAlex MacLean  %call = call i32 @bar(ptr %alloca)
3989b7b3b9SAlex MacLean  ret i32 %call
40f9304974SYoungsuk Kim}
4189b7b3b9SAlex MacLean
4289b7b3b9SAlex MacLeandeclare i32 @bar(ptr)
43