| Issue |
244773
|
| Summary |
[ARM][Thumb-2][PBQP] Crash due to ARMLoadStoreOpt::findFreeReg clobbering reserved BasePtr (r6)
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
alinas
|
Root cause: #222450 (`9fcdae44a646`).
#226523 (`bd51cc5c5556`) fixed 3 of the 4 reg allocs, RegAllocPBQP was not updated.
Reproduces with `llc -mtriple=thumbv7-unknown-linux-gnueabihf -O2 -regalloc=pbqp repro.ll -o -` on
```
target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
target triple = "thumbv7-unknown-linux-gnueabihf"
declare void @escape(ptr, ptr)
declare i32 @bar(i32, i32, i32, i32, i32, i32, i32, i32, i32, i32)
define i32 @foo(i32 %n, i32 %cond, ptr readonly %p, i32 %c0) nounwind "frame-pointer"="all" {
entry:
%buf = alloca [80 x i32], align 4
%sz = shl i32 %n, 2
%dyn = alloca i8, i32 %sz, align 8
store i32 %c0, ptr %buf, align 4
call void @escape(ptr nonnull %buf, ptr nonnull %dyn)
%p10 = getelementptr inbounds i8, ptr %p, i32 40
%c1 = load i32, ptr %p10, align 4
%p11 = getelementptr inbounds i8, ptr %p, i32 44
%c2 = load i32, ptr %p11, align 4
%p12 = getelementptr inbounds i8, ptr %p, i32 48
%c3 = load i32, ptr %p12, align 4
%p13 = getelementptr inbounds i8, ptr %p, i32 52
%c4 = load i32, ptr %p13, align 4
%cmp = icmp eq i32 %cond, 0
br i1 %cmp, label %exit, label %then
then:
%x0 = load i32, ptr %p, align 4
%p1 = getelementptr inbounds i8, ptr %p, i32 4
%x1 = load i32, ptr %p1, align 4
%p2 = getelementptr inbounds i8, ptr %p, i32 8
%x2 = load i32, ptr %p2, align 4
%p3 = getelementptr inbounds i8, ptr %p, i32 12
%x3 = load i32, ptr %p3, align 4
%p4 = getelementptr inbounds i8, ptr %p, i32 16
%y0 = load i32, ptr %p4, align 4
%p5 = getelementptr inbounds i8, ptr %p, i32 20
%y1 = load i32, ptr %p5, align 4
%call = call i32 @bar(i32 %c0, i32 %c1, i32 %y0, i32 %y1, i32 %c2, i32 %c3, i32 %x2, i32 %x1, i32 %x0, i32 %x3)
br label %exit
exit:
%ret = phi i32 [ %call, %then ], [ 0, %entry ]
%b0 = load i32, ptr %buf, align 4
%p14 = getelementptr inbounds i8, ptr %p, i32 56
%c5 = load i32, ptr %p14, align 4
%s0 = add i32 %c1, %c0
%s1 = add i32 %s0, %c2
%s2 = add i32 %s1, %c3
%s3 = add i32 %s2, %c4
%s4 = add i32 %s3, %ret
%s5 = add i32 %s4, %b0
%s6 = add i32 %s5, %c5
ret i32 %s6
}
```
Additional info: When targeting 32-bit ARM Thumb-2 with `-mllvm -regalloc=pbqp` (`-O2`), `ARMLoadStoreOpt::findFreeReg` can allocate the reserved Thumb-2 base pointer `r6` (`BasePtr`) as a scratch base register for `stm.w` in functions with both a dynamic `alloca` and a local frame `>= 128` bytes. Overwriting `r6` corrupts subsequent `[r6, #imm]` stack/spill accesses.
AI suggested fixes to resolve:
1. Update `RegAllocPBQP` (`llvm/lib/CodeGen/RegAllocPBQP.cpp`) to require `MachineRegisterClassInfoWrapperPass` and call `RCI.updateReservedRegs(MF.getRegInfo().getReservedRegs())` immediately after `MF.getRegInfo().freezeReservedRegs()`, matching `RegAllocBase::init` and `RegAllocFast` (#226523).
2. Add `assert(RCI->hasCurrentReservedRegs() && "RegisterClassInfo misses a change to the reserved registers");` in `ARMLoadStoreOpt::runOnMachineFunction`, matching `AMDGPURewriteAGPRCopyMFMA` (#237446).
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs