95 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: not llc < %s -mtriple=nvptx -mattr=+ptx72 -mcpu=sm_52 2>&1 | FileCheck %s --check-prefixes=CHECK-FAILS3; RUN: not llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_50 2>&1 | FileCheck %s --check-prefixes=CHECK-FAILS4 5; RUN: llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_52 | FileCheck %s --check-prefixes=CHECK-326; RUN: llc < %s -mtriple=nvptx64 -mattr=+ptx73 -mcpu=sm_52 | FileCheck %s --check-prefixes=CHECK-647; RUN: %if ptxas-isa-7.3 && ptxas-ptr32 %{ llc < %s -mtriple=nvptx -mattr=+ptx73 -mcpu=sm_52 | %ptxas-verify %}8; RUN: %if ptxas-isa-7.3 %{ llc < %s -mtriple=nvptx64 -mattr=+ptx73 -mcpu=sm_52 | %ptxas-verify %}9 10; CHECK-FAILS: in function test_dynamic_stackalloc{{.*}}: Support for dynamic alloca introduced in PTX ISA version 7.3 and requires target sm_52.11 12define i32 @test_dynamic_stackalloc(i64 %n) {13; CHECK-32-LABEL: test_dynamic_stackalloc(14; CHECK-32: {15; CHECK-32-NEXT: .reg .b32 %r<7>;16; CHECK-32-EMPTY:17; CHECK-32-NEXT: // %bb.0:18; CHECK-32-NEXT: ld.param.b32 %r1, [test_dynamic_stackalloc_param_0];19; CHECK-32-NEXT: add.s32 %r2, %r1, 7;20; CHECK-32-NEXT: and.b32 %r3, %r2, -8;21; CHECK-32-NEXT: alloca.u32 %r4, %r3, 16;22; CHECK-32-NEXT: cvta.local.u32 %r5, %r4;23; CHECK-32-NEXT: { // callseq 0, 024; CHECK-32-NEXT: .param .b32 param0;25; CHECK-32-NEXT: .param .b32 retval0;26; CHECK-32-NEXT: st.param.b32 [param0], %r5;27; CHECK-32-NEXT: call.uni (retval0), bar, (param0);28; CHECK-32-NEXT: ld.param.b32 %r6, [retval0];29; CHECK-32-NEXT: } // callseq 030; CHECK-32-NEXT: st.param.b32 [func_retval0], %r6;31; CHECK-32-NEXT: ret;32;33; CHECK-64-LABEL: test_dynamic_stackalloc(34; CHECK-64: {35; CHECK-64-NEXT: .reg .b32 %r<2>;36; CHECK-64-NEXT: .reg .b64 %rd<6>;37; CHECK-64-EMPTY:38; CHECK-64-NEXT: // %bb.0:39; CHECK-64-NEXT: ld.param.b64 %rd1, [test_dynamic_stackalloc_param_0];40; CHECK-64-NEXT: add.s64 %rd2, %rd1, 7;41; CHECK-64-NEXT: and.b64 %rd3, %rd2, -8;42; CHECK-64-NEXT: alloca.u64 %rd4, %rd3, 16;43; CHECK-64-NEXT: cvta.local.u64 %rd5, %rd4;44; CHECK-64-NEXT: { // callseq 0, 045; CHECK-64-NEXT: .param .b64 param0;46; CHECK-64-NEXT: .param .b32 retval0;47; CHECK-64-NEXT: st.param.b64 [param0], %rd5;48; CHECK-64-NEXT: call.uni (retval0), bar, (param0);49; CHECK-64-NEXT: ld.param.b32 %r1, [retval0];50; CHECK-64-NEXT: } // callseq 051; CHECK-64-NEXT: st.param.b32 [func_retval0], %r1;52; CHECK-64-NEXT: ret;53 %alloca = alloca i8, i64 %n, align 1654 %call = call i32 @bar(ptr %alloca)55 ret i32 %call56}57 58define float @test_dynamic_stackalloc_unaligned(i64 %0) {59; CHECK-32-LABEL: test_dynamic_stackalloc_unaligned(60; CHECK-32: {61; CHECK-32-NEXT: .reg .b32 %r<7>;62; CHECK-32-EMPTY:63; CHECK-32-NEXT: // %bb.0:64; CHECK-32-NEXT: ld.param.b32 %r1, [test_dynamic_stackalloc_unaligned_param_0];65; CHECK-32-NEXT: shl.b32 %r2, %r1, 2;66; CHECK-32-NEXT: add.s32 %r3, %r2, 7;67; CHECK-32-NEXT: and.b32 %r4, %r3, -8;68; CHECK-32-NEXT: alloca.u32 %r5, %r4, 8;69; CHECK-32-NEXT: ld.local.b32 %r6, [%r5];70; CHECK-32-NEXT: st.param.b32 [func_retval0], %r6;71; CHECK-32-NEXT: ret;72;73; CHECK-64-LABEL: test_dynamic_stackalloc_unaligned(74; CHECK-64: {75; CHECK-64-NEXT: .reg .b32 %r<2>;76; CHECK-64-NEXT: .reg .b64 %rd<6>;77; CHECK-64-EMPTY:78; CHECK-64-NEXT: // %bb.0:79; CHECK-64-NEXT: ld.param.b64 %rd1, [test_dynamic_stackalloc_unaligned_param_0];80; CHECK-64-NEXT: shl.b64 %rd2, %rd1, 2;81; CHECK-64-NEXT: add.s64 %rd3, %rd2, 7;82; CHECK-64-NEXT: and.b64 %rd4, %rd3, -8;83; CHECK-64-NEXT: alloca.u64 %rd5, %rd4, 8;84; CHECK-64-NEXT: ld.local.b32 %r1, [%rd5];85; CHECK-64-NEXT: st.param.b32 [func_retval0], %r1;86; CHECK-64-NEXT: ret;87 %4 = alloca float, i64 %0, align 488 %5 = getelementptr float, ptr %4, i64 089 %6 = load float, ptr %5, align 490 ret float %691}92 93declare i32 @bar(ptr)94 95