| Issue |
175909
|
| Summary |
Assertion "VPlan cost model and legacy cost model disagreed" when building Clang/LLVM for rv32gcv
|
| Labels |
vectorizers,
crash-on-valid
|
| Assignees |
|
| Reporter |
asb
|
Attempting to cross-compile Clang/LLVM targeting rv32gcv with recent HEAD we get some crashes from the "VPlan cost model and legacy cost model disagreed" assert in LoopVectorize.cpp.
Here is a reduced test case:
```
; ModuleID = 'reduced.bc'
target datalayout = "e-m:e-p:32:32-i64:64-n32-S128"
target triple = "riscv32-unknown-linux-gnu"
define ptr @_ZSt23__copy_move_backward_a2ILb1EPSt4pairIdyES2_ET1_T0_S4_S3_(ptr %__first, ptr %__last) {
entry:
%__last.addr1 = alloca ptr, align 4
%__result.addr = alloca ptr, align 4
store ptr %__first, ptr %__result.addr, align 4
br label %while.cond
while.cond: ; preds = %while.body, %entry
%0 = phi ptr [ %.pre, %while.body ], [ %__last, %entry ]
%cmp.not = icmp eq ptr %__first, %0
br i1 %cmp.not, label %while.end, label %while.body
while.body: ; preds = %while.cond
%incdec.ptr = getelementptr i8, ptr %0, i32 -16
store ptr %incdec.ptr, ptr %__last.addr1, align 4
%1 = load ptr, ptr %__result.addr, align 4
%incdec.ptr1 = getelementptr i8, ptr %1, i32 -16
store ptr %incdec.ptr1, ptr %__result.addr, align 4
%2 = call ptr @_ZSt12__assign_oneILb1EPSt4pairIdyES2_EvRT0_RT1_(ptr %__result.addr, ptr %__last.addr1)
%.pre = load ptr, ptr %__last.addr1, align 4
br label %while.cond
while.end: ; preds = %while.cond
ret ptr null
}
define ptr @_ZSt12__assign_oneILb1EPSt4pairIdyES2_EvRT0_RT1_(ptr %__out, ptr %__in) {
entry:
%0 = load ptr, ptr %__in, align 4
%1 = load double, ptr %0, align 8
%2 = load ptr, ptr %__out, align 4
store double %1, ptr %2, align 8
%second3.i = getelementptr i8, ptr %2, i32 8
store i64 0, ptr %second3.i, align 8
ret ptr null
}
```
Reproduce with `./build/rvrel/bin/clang++ --target=riscv32-linux-gnu -c -march=rv32gcv -O3 reduced.ll`
Does this checking of the VPlan vs legacy cost model still needed at this point, or is the VPlan based model mature enough to remove it?
CC @lukel97
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
