[PowerPC] Infinite loop in PeepholeCROps

[nikic]

Bugzilla Link	49509
Resolution	FIXED
Resolved on	Mar 29, 2021 16:55
Version	trunk
OS	Linux
Blocks	#48246
CC	@jrmuizel,@nemanjai,@nagisa
Fixed by commit(s)	2489cba 8ca5690

Extended Description

; RUN: llc -O2 < %s

target datalayout = "E-m:e-p:32:32-i64:64-n32"
target triple = "powerpc-unknown-linux-gnu"

define void @​test() {
bb:
br i1 undef, label %bb2, label %bb1

bb2: ; preds = %bb
%i = select i1 undef, i64 0, i64 72057594037927936
store i64 %i, i64* undef, align 8
ret void

bb1: ; preds = %bb
%i50 = load i8, i8* undef, align 8
%i52 = load i128, i128* null, align 8
%i62 = icmp eq i8 %i50, 0
br i1 undef, label %bb66, label %bb64

bb64: ; preds = %bb63
ret void

bb66: ; preds = %bb63
%i67 = lshr i128 -1, 0
%i68 = xor i128 %i52, -1
%i69 = add i128 0, %i68
%i70 = and i128 %i67, %i69
%i71 = icmp eq i128 %i70, 0
%i74 = select i1 %i62, i64 0, i64 72057594037927936
%i75 = select i1 %i71, i64 144115188075855872, i64 %i74
store i64 %i75, i64* undef, align 8
ret void
}

This enters an infinite combine loop inside PPCDAGToDAGISel::PeepholeCROps().

[nikic]

DAG:

Optimized legalized selection DAG: %bb.4 'test:bb66'
SelectionDAG has 30 nodes:
t0: ch = EntryToken
t4: i32,ch = CopyFromReg t0, Register:i32 %0
t20: i1,ch = CopyFromReg t0, Register:i1 %4
t6: i32,ch = CopyFromReg t4:1, Register:i32 %1
t8: i32,ch = CopyFromReg t6:1, Register:i32 %2
t10: i32,ch = CopyFromReg t8:1, Register:i32 %3
t37: i32 = and t10, t6
t38: i32 = and t8, t4
t39: i32 = and t37, t38
t41: i32 = select t20, Constant:i32<0>, OpaqueConstant:i32<16777216>
t48: i32 = select_cc t39, Constant:i32<-1>, Constant:i32<33554432>, t41, seteq:ch
t44: ch = store<(store 4 into i64* undef, align 8)> t0, t48, undef:i32, undef:i32
t29: i1 = setcc t39, Constant:i32<-1>, seteq:ch
t49: i1 = or t29, t20
t50: i32 = select t49, Constant:i32<0>, OpaqueConstant:i32<0>
t46: ch = store<(store 4 into i64* undef + 4, basealign 8)> t0, t50, undef:i32, undef:i32
t47: ch = TokenFactor t44, t46
t28: ch = PPCISD::RET_FLAG t47

Start of the peephole combine loop:

===== Instruction selection ends:
Creating new machine node: t64: i1 = CRNOR t29, t20
CR Peephole replacing:
Old: t50: i32 = SELECT_I4 t49, t34, t34

New: t50: i32 = SELECT_I4 t49, t34, t34

CR Peephole replacing:
Old: t49: i1 = CROR t29, t20

New: t64: i1 = CRNOR t29, t20

CR Peephole replacing:
Old: t50: i32 = SELECT_I4 t64, t34, t34

New: t50: i32 = SELECT_I4 t64, t34, t34

CR Peephole replacing:
Old: t64: i1 = CRNOR t29, t20

New: t49: i1 = CROR t29, t20

Creating new machine node: t65: i1 = CRNOR t29, t20
CR Peephole replacing:
Old: t50: i32 = SELECT_I4 t49, t34, t34

New: t50: i32 = SELECT_I4 t49, t34, t34

CR Peephole replacing:
Old: t49: i1 = CROR t29, t20

New: t65: i1 = CRNOR t29, t20

CR Peephole replacing:
Old: t50: i32 = SELECT_I4 t65, t34, t34

New: t50: i32 = SELECT_I4 t65, t34, t34

CR Peephole replacing:
Old: t65: i1 = CRNOR t29, t20

New: t49: i1 = CROR t29, t20

[nikic]

Here's the DAG when entering PeepholeCROps:

SelectionDAG has 38 nodes:
t0: ch = EntryToken
t34: i32 = LI TargetConstant:i32<0>
t4: i32,ch = CopyFromReg t0, Register:i32 %0
t20: i1,ch = CopyFromReg t0, Register:i1 %4
t6: i32,ch = CopyFromReg t4:1, Register:i32 %1
t8: i32,ch = CopyFromReg t6:1, Register:i32 %2
t10: i32,ch = CopyFromReg t8:1, Register:i32 %3
t37: i32 = AND t10, t6
t38: i32 = AND t8, t4
t39: i32 = AND t37, t38
t60: i32 = TargetConstant<16777216>
t62: i32 = TargetConstant<33554432>
t58: i32 = CMPWI t39, TargetConstant:i32<65535>
t31: i32 = LIS TargetConstant:i32<512>
t35: i32 = LIS TargetConstant:i32<256>
t41: i32 = SELECT_I4 t20, t34, t35
t48: i32 = SELECT_CC_I4 t58, t31, t41, TargetConstant:i32<76>
t44: ch = STW<Mem:(store 4 into i64* undef, align 8)> t48, TargetConstant:i32<0>, IMPLICIT_DEF:i32, t0
t55: i32 = CMPWI t39, TargetConstant:i32<-1>
t29: i1 = EXTRACT_SUBREG t55, TargetConstant:i32<3>
t49: i1 = CROR t29, t20
t50: i32 = SELECT_I4 t49, t34, t34
t46: ch = STW<Mem:(store 4 into i64* undef + 4, basealign 8)> t50, TargetConstant:i32<0>, IMPLICIT_DEF:i32, t0
t47: ch = TokenFactor t44, t46
t28: ch = BLR t47

With the problematic part being:

t34: i32 = LI TargetConstant:i32<0>
t49: i1 = CROR t29, t20
t50: i32 = SELECT_I4 t49, t34, t34

The transform converts select x, y, y into select !x, y, y swapping the select operands, but of course that doesn't actually do anything if they're equal. And then we can convert that back to select x, y, y etc.

[nikic]

Candidate patch: https://reviews.llvm.org/D98340

[nikic]

Fixed by 2489cba on main.

[tstellar]

Merged: 8ca5690