1; RUN: llc < %s -mtriple=nvptx64-unknown-unknown | FileCheck %s 2; RUN: %if ptxas %{ llc < %s -mtriple=nvptx64-unknown-unknown | %ptxas-verify %} 3; 4; NVPTXTargetLowering::getFunctionParamOptimizedAlign, which was introduces in 5; D120129, contained a poorly designed assertion checking that a function with 6; internal or private linkage is not a kernel. It relied on invariants that 7; were not actually guaranteed, and that resulted in compiler crash with some 8; CUDA versions (see discussion with @jdoerfert in D120129). This test contains 9; metadata that caused compiler crash and a function with internal linkage 10; which purpose is to let compiler run on path where the crash happened. 11; Metadata was obtained from libdevice.10.bc shipped with cuda-11-0. 12 13 14define internal i32 @foo() { 15 ; CHECK-LABEL: .func (.param .b32 func_retval0) foo() 16 ret i32 42 17} 18 19define i32 @bar() { 20 ; CHECK-LABEL: .visible .func (.param .b32 func_retval0) bar() 21 %x = call i32 @foo() 22 ret i32 %x 23} 24 25!nvvmir.version = !{!0} 26!nvvm.annotations = !{!1, !2, !1, !3, !3, !3, !3, !4, !4, !3} 27 28!0 = !{i32 1, i32 4} 29!1 = !{null, !"align", i32 8} 30!2 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080} 31!3 = !{null, !"align", i32 16} 32!4 = !{null, !"align", i32 16, !"align", i32 65552, !"align", i32 131088} 33!5 = distinct !{!5, !6} 34!6 = !{!"llvm.loop.unroll.count", i32 1} 35!7 = distinct !{!7, !6} 36