incorrect vector codegen

[llvmbot]

Bugzilla Link	42803
Resolution	FIXED
Resolved on	Feb 24, 2020 15:17
Version	9.0
OS	Linux
Blocks	#30613
Attachments	llvm-ir
Reporter	LLVM Bugzilla Contributor
CC	@andrewrk,@topperc,@efriedma-quic,@RKSimon
Fixed by commit(s)	a5fa778

Extended Description

Using llvm 8 (Ubuntu Disco, 8.0.0-3 (tags/RELEASE_800/final)) The following zig program:

pub export fn start() void {
var a: [3]bool = []bool{ false, true, false};
var x2: @​Vector(3, bool) = a;
if (x2[1] != true) unreachable;
}
const builtin = @​import("builtin");
pub fn panic(msg: []const u8, error_return_trace: ?*builtin.StackTrace) noreturn {
while (true) {}
}

compiles to this llvm-ir (full file included)

@​0 = internal unnamed_addr constant [3 x i1] [i1 false, i1 true, i1 false], align 1

; Function Attrs: nobuiltin nounwind
define void @​_start() #​2 !dbg !​136 {
Entry:
%a = alloca [3 x i1], align 1
%x2 = alloca <3 x i1>, align 4
%0 = bitcast [3 x i1]* %a to i8*, !dbg !​147
call void @​llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %0, i8* align 1 bitcast ([3 x i1]* @​0 to i8*), i64 3, i1 false), !dbg !​147
call void @​llvm.dbg.declare(metadata [3 x i1]* %a, metadata !​140, metadata !DIExpression()), !dbg !​147
%1 = load [3 x i1], [3 x i1]* %a, !dbg !​148
%vector_to_array = extractvalue [3 x i1] %1, 0, !dbg !​148
%2 = insertelement <3 x i1> undef, i1 %vector_to_array, i32 0, !dbg !​148
%vector_to_array1 = extractvalue [3 x i1] %1, 1, !dbg !​148
%3 = insertelement <3 x i1> %2, i1 %vector_to_array1, i32 1, !dbg !​148
%vector_to_array2 = extractvalue [3 x i1] %1, 2, !dbg !​148
%4 = insertelement <3 x i1> %3, i1 %vector_to_array2, i32 2, !dbg !​148
store <3 x i1> %4, <3 x i1>* %x2, align 4, !dbg !​149
call void @​llvm.dbg.declare(metadata <3 x i1>* %x2, metadata !​145, metadata !DIExpression()), !dbg !​149
%5 = load <3 x i1>, <3 x i1>* %x2, align 4, !dbg !​150
%6 = extractelement <3 x i1> %5, i32 1, !dbg !​150
%7 = icmp ne i1 %6, true, !dbg !​151
br i1 %7, label %Then, label %Else, !dbg !​152

Then: ; preds = %Entry
tail call fastcc void @​panic(%"[]u8"* @​2, %builtin.StackTrace* null), !dbg !​153
unreachable, !dbg !​153

Else: ; preds = %Entry
br label %EndIf, !dbg !​152

EndIf: ; preds = %Else
ret void, !dbg !​154
}

Which produces the following x86_64 (disassembled)

Dump of assembler code for function _start:
0x0000000000201010 <+0>: push %rbp
0x0000000000201011 <+1>: mov %rsp,%rbp
=> 0x0000000000201014 <+4>: sub $0x10,%rsp
0x0000000000201018 <+8>: mov -0xef8(%rip),%al # 0x200126 <__unnamed_1+2>
0x000000000020101e <+14>: mov %al,-0x2(%rbp)
0x0000000000201021 <+17>: mov -0xf04(%rip),%cx # 0x200124 <__unnamed_1>
0x0000000000201028 <+24>: mov %cx,-0x4(%rbp)
0x000000000020102c <+28>: mov -0x2(%rbp),%al
0x000000000020102f <+31>: mov -0x4(%rbp),%dl
0x0000000000201032 <+34>: mov -0x3(%rbp),%sil
0x0000000000201036 <+38>: add %sil,%sil
0x0000000000201039 <+41>: or %sil,%dl
0x000000000020103c <+44>: shl $0x2,%al
0x000000000020103f <+47>: or %al,%dl
0x0000000000201041 <+49>: and $0x7,%dl
0x0000000000201044 <+52>: mov %dl,-0xc(%rbp)
0x0000000000201047 <+55>: mov -0xc(%rbp),%al
0x000000000020104a <+58>: mov %al,-0x8(%rbp)
0x000000000020104d <+61>: mov -0x8(%rbp),%al
0x0000000000201050 <+64>: and $0x1,%al
0x0000000000201052 <+66>: neg %al
0x0000000000201054 <+68>: test $0x1,%al
0x0000000000201056 <+70>: jne 0x20106d <_start+93>
0x0000000000201058 <+72>: jmp 0x20105a <_start+74>
0x000000000020105a <+74>: xor %eax,%eax
0x000000000020105c <+76>: mov %eax,%esi
0x000000000020105e <+78>: movabs $0x200140,%rdi
0x0000000000201068 <+88>: callq 0x201000
0x000000000020106d <+93>: jmp 0x20106f <_start+95>
0x000000000020106f <+95>: add $0x10,%rsp
0x0000000000201073 <+99>: pop %rbp
0x0000000000201074 <+100>: retq

[topperc]

By "bad" do you mean incorrect or not optimal?

[llvmbot]

Incorrect. The middle element is false, and the program panics.

[llvmbot]

runnable executable

[llvmbot]

Here is a version that can be built with current zig (rather than my branch), and reproduces the problem, but its more difficult to follow, so I don't recommend looking at it. https://godbolt.org/z/o6lm3Z

[llvmbot]

I reproduced it, and reduced the test case on ARM:

pub export fn start() void {
var a: [3]bool = []bool{ false, true, false};
var x2: @​Vector(3, bool) = a;
if (x2[1] != true) unreachable;
}
const builtin = @​import("builtin");
pub fn panic(msg: []const u8, error_return_trace: ?*builtin.StackTrace) noreturn {
while (true) {}
}

=======================

arget datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64v8-unknown-linux-gnu"

%"[]u8" = type { i8*, i64 }
%"[]usize" = type { i64*, i64 }

@​0 = internal unnamed_addr constant [3 x i1] [i1 false, i1 true, i1 false], align 1

; Function Attrs: nobuiltin noreturn nounwind
define internal fastcc void @​panic() unnamed_addr #​0 {
Entry:
br label %WhileCond

WhileCond: ; preds = %WhileCond, %Entry
br label %WhileCond
}

; Function Attrs: nobuiltin nounwind
define void @​_start() #​2 {
Entry:
%a = alloca [3 x i1], align 1
%0 = bitcast [3 x i1]* %a to i8*
call void @​llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %0, i8* align 1 bitcast ([3 x i1]* @​0 to i8*), i64 3, i1 false)
%1 = load [3 x i1], [3 x i1]* %a
%vector_to_array = extractvalue [3 x i1] %1, 0
%2 = insertelement <3 x i1> undef, i1 %vector_to_array, i32 0
%vector_to_array1 = extractvalue [3 x i1] %1, 1
%3 = insertelement <3 x i1> %2, i1 %vector_to_array1, i32 1
%vector_to_array2 = extractvalue [3 x i1] %1, 2
%4 = insertelement <3 x i1> %3, i1 %vector_to_array2, i32 2
%v6 = extractelement <3 x i1> %4, i32 1
%v7 = icmp ne i1 %v6, true
br i1 %v7, label %Then, label %Else

Then: ; preds = %Entry
tail call fastcc void @​panic()
unreachable

Else: ; preds = %Entry
br label %EndIf

EndIf: ; preds = %Else
ret void
}

; Function Attrs: argmemonly nounwind
declare void @​llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1) #​3

attributes #​0 = { nobuiltin noreturn nounwind "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" }
attributes #​1 = { nounwind readnone speculatable }
attributes #​2 = { nobuiltin nounwind "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" }
attributes #​3 = { argmemonly nounwind }

=========

    .text
    .file   "min.ll"
    .p2align        2               // -- Begin function panic
    .type   panic,@function

panic: // @​panic
// %bb.0: // %Entry
.LBB0_1: // %WhileCond
// =>This Inner Loop Header: Depth=1
b .LBB0_1
.Lfunc_end0:
.size panic, .Lfunc_end0-panic
// -- End function
.globl _start // -- Begin function _start
.p2align 2
.type _start,@function
_start: // @​_start
// %bb.0: // %Entry
sub sp, sp, #​32 // =32
adrp x8, __unnamed_1
add x8, x8, :lo12:__unnamed_1
ldrh w9, [x8]
ldrb w8, [x8, #​2]
stp x29, x30, [sp, #​16] // 16-byte Folded Spill
add x29, sp, #​16 // =16
sturh w9, [x29, #-4]
ubfx w9, w9, #​8, #​8
cmp w9, #​1 // =1
sturb w8, [x29, #-2]
b.ne .LBB1_2
// %bb.1: // %EndIf
ldp x29, x30, [sp, #​16] // 16-byte Folded Reload
add sp, sp, #​32 // =32
ret
.LBB1_2: // %Then
bl panic
.Lfunc_end1:
.size _start, .Lfunc_end1-_start
// -- End function
.type __unnamed_1,@object // @​0
.section .rodata,"a",@progbits
__unnamed_1:
.byte 0 // 0x0
.byte 1 // 0x1
.byte 0 // 0x0
.size __unnamed_1, 3

    .section        ".note.GNU-stack","",@progbits

[topperc]

Do you know if this reproduces with trunk?

[llvmbot]

I reproduced it with 9.0.0-svn362869-1~exp1 in Debian experimental on arm64. The generated assembly appeared to be identical.

[topperc]

Are you compiling with sse disabled? That's the only way I could get the assembly you were showing.

The issue seems to be a bug in the handling of vector loads that don't fit exactly into some number of bytes. We get confused and end up treating all 3 bits as being in bit 0 of the same byte.

[llvmbot]

Are you compiling with sse disabled?

Zig uses -march=native and this laptop supports AVX-256. I also reproduced it on aarch64.

[llvmbot]

Ignore that. These are all with optimizations disabled, -O0. I could not reproduce with optimizations enabled.

[efriedma-quic]

Target-independent legalization for load and store operations on vector types with non-byte-sized element types has been a known problem in the past... it wasn't clear how we wanted to represent them. We've settled on tightly packing them (so <3 x i1> is one byte), and we fixed most of the legalization code to correctly honor that. But maybe there's something that still isn't fixed.

I can't reproduce the issue with the AArch64 testcase; can you give complete steps for building an executable that reproduces the issue?

[topperc]

Eli, on X86 at least with sse disabled. What I see happened in that the v3i1 store needs to be widened to v4i1 during type legalization. The call to FindMemType in the load widening code ends up returning i1 as the type because no other types evenly divide into WidenWidth which is 4. This causes the code to emit 3 identical i1 loads since the pointer doesn't increment. These of course get CSEd to the same node. This is then assembled together to make a v3i1 vector. Since the loads are just i1 the element offset within in the byte is now lost.

[topperc]

Oops that should have said "v3i1 load needs to be widened". v3i1 stores already scalarize, but not loads.

[efriedma-quic]

DAGTypeLegalizer::WidenVecOp_STORE has a special case for non-byte-sized stores. (See https://reviews.llvm.org/D42100 .) Looks like we never applied an equivalent fix to DAGTypeLegalizer::WidenVecRes_LOAD.

[andrewrk]

This is still an open downstream bug with LLVM 10 rc1. ziglang/zig#3246 (comment)