Compiling hyper 0.12 on armv7-linux-androideabi with target-features=+neon fails with LLVM ERROR: ran out of registers during register allocation

[Eijebong]

root@7f26157a3837:~/hyper# cargo rustc --release -v --target "armv7-linux-androideabi" -- -C target-feature=+neon                                             
     Running `rustc --crate-name hyper src/lib.rs --crate-type lib --emit=dep-info,link -C opt-level=3 -C codegen-units=1 -C target-feature=+neon --cfg 'feature="__internal_flaky_tests"' --cfg 'feature="default"' --cfg 'feature="futures-cpupool"' --cfg 'feature="net2"' --cfg 'feature="runtime"' --cfg 'feature="tokio"' --cfg 'feature="tokio-executor"' --cfg 'feature="tokio-reactor"' --cfg 'feature="tokio-tcp"' --cfg 'feature="tokio-timer"' -C metadata=5c7c44dab8eed49e -C extra-filename=-5c7c44dab8eed49e --out-dir /root/hyper/target/armv7-linux-androideabi/release/deps --target armv7-linux-androideabi -L dependency=/root/hyper/target/armv7-linux-androideabi/release/deps -L dependency=/root/hyper/target/release/deps --extern bytes=/root/hyper/target/armv7-linux-androideabi/release/deps/libbytes-ddc5925e1332c4e2.rlib --extern futures=/root/hyper/target/armv7-linux-androideabi/release/deps/libfutures-169e26de4e2e0883.rlib --extern futures_cpupool=/root/hyper/target/armv7-linux-androideabi/release/deps/libfutures_cpupool-a7ff7f77e82e2fd1.rlib --extern h2=/root/hyper/target/armv7-linux-androideabi/release/deps/libh2-a6ab8093d4aaef1e.rlib --extern http=/root/hyper/target/armv7-linux-androideabi/release/deps/libhttp-a998474d6086d0cc.rlib --extern httparse=/root/hyper/target/armv7-linux-androideabi/release/deps/libhttparse-fe59ea708b984d2c.rlib --extern iovec=/root/hyper/target/armv7-linux-androideabi/release/deps/libiovec-0fbee7688c5d5997.rlib --extern itoa=/root/hyper/target/armv7-linux-androideabi/release/deps/libitoa-6968692ba2ead4d2.rlib --extern log=/root/hyper/target/armv7-linux-androideabi/release/deps/liblog-982ca16cb4a51ba3.rlib --extern net2=/root/hyper/target/armv7-linux-androideabi/release/deps/libnet2-eca709a2543aac40.rlib --extern time=/root/hyper/target/armv7-linux-androideabi/release/deps/libtime-d2ab1c7a701d0c5c.rlib --extern tokio=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio-f318b1cbbd06c200.rlib --extern tokio_executor=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio_executor-a34c5d48733eb612.rlib --extern tokio_io=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio_io-615c1ba46c31b299.rlib --extern tokio_reactor=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio_reactor-5f33b571b5aa2433.rlib --extern tokio_tcp=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio_tcp-dcd0b854ed32b20c.rlib --extern tokio_timer=/root/hyper/target/armv7-linux-androideabi/release/deps/libtokio_timer-b8925bce2a2b5c77.rlib --extern want=/root/hyper/target/armv7-linux-androideabi/release/deps/libwant-c71f954f09dd620c.rlib`                                                                         
LLVM ERROR: ran out of registers during register allocation
error: Could not compile `hyper`.

[est31]

Can reproduce this locally.

$ git clone https://github.com/hyperium/hyper
$ cd hyper
$ git checkout v0.12.10
$ cargo +nightly rustc --release -v --target "armv7-linux-androideabi" -- -C target-feature=+neon
[...]
LLVM ERROR: ran out of registers during register allocation                                                      
error: Could not compile `hyper`

Rustc version:

$ rustc +nightly -vV
rustc 1.31.0-nightly (4699283c5 2018-10-13)
binary: rustc
commit-hash: 4699283c5b549d1559c198123a67fef498aa6a44
commit-date: 2018-10-13
host: x86_64-unknown-linux-gnu
release: 1.31.0-nightly
LLVM version: 8.0

[SimonSapin]

(This is blocking the Hyper upgrade in Servo.)

[est31]

I've traced down the bug to originate from the record_header_indices function in src/proto/h1/role.rs, here. If I remove all of its contents, compilation succeeds.

[est31]

This gives us this self-contained example:

pub struct Header<'a> {
    pub name: &'a [u8],
    pub value: &'a [u8],
}

pub struct HeaderIndices {
    name: (usize, usize),
    value: (usize, usize),
}


pub fn record_header_indices(bytes: &[u8], headers: &[Header], indices: &mut [HeaderIndices]) {
    let bytes_ptr = bytes.as_ptr() as usize;
    for (header, indices) in headers.iter().zip(indices.iter_mut()) {
        let name_start = header.name.as_ptr() as usize - bytes_ptr;
        let name_end = name_start + header.name.len();
        indices.name = (name_start, name_end);
        let value_start = header.value.as_ptr() as usize - bytes_ptr;
        let value_end = value_start + header.value.len();
        indices.value = (value_start, value_end);
    }
}

Compiled with the same invocation as above. cargo +nightly rustc --release -v --target "armv7-linux-androideabi" -- -C target-feature=+neon

For a workaround, splitting up the loop seems to have fixed the bug:

pub fn record_header_indices(bytes: &[u8], headers: &[Header], indices: &mut [HeaderIndices]) {
    let bytes_ptr = bytes.as_ptr() as usize;
    for (header, indices) in headers.iter().zip(indices.iter_mut()) {
        let name_start = header.name.as_ptr() as usize - bytes_ptr;
        let name_end = name_start + header.name.len();
        indices.name = (name_start, name_end);
    }
    for (header, indices) in headers.iter().zip(indices.iter_mut()) {
        let value_start = header.value.as_ptr() as usize - bytes_ptr;
        let value_end = value_start + header.value.len();
        indices.value = (value_start, value_end);
    }
}

[est31]

I've tried to make a C version that reproduces this, but failed. This is a Rust version very close to C that reproduces the bug:

pub struct Header {
    pub name: * const u8,
    pub name_len: usize,
    pub value: * const u8,
    pub value_len: usize,
}

pub struct HeaderIndices {
    name_a :usize,
    name_b: usize,
    value_a: usize,
    value_b: usize,
}


pub fn record_header_indices(bytes_ptr: usize, headers: * const Header, indices: *mut HeaderIndices, len: isize) {
    for i in 0 .. len {
        let mut indices = unsafe { &mut *indices.offset(i) };
        let header = unsafe { &*headers.offset(i) };
        let name_start = header.name as usize - bytes_ptr;
        let name_end = name_start + header.name_len;
        indices.name_a = name_start;
        indices.name_b = name_end;
        let value_start = header.value as usize - bytes_ptr;
        let value_end = value_start + header.value_len;
        indices.value_a = value_start;
        indices.value_b = value_end;
    }
}

But this C version doesn't (clang invocation was clang -cc1 test.c -triple armv7-linux-androideabi -O3 -target-feature +neon):

typedef unsigned __INT32_TYPE__ size_t;

typedef struct header {
    char *name;
    size_t name_len;
    char *val;
    size_t val_len;
} header;

typedef struct header_indices {
    size_t name_a;
    size_t name_b;
    size_t val_a;
    size_t val_b;
} header_indices;

void record_header_indices(size_t bytes_ptr, const header *headers, header_indices *indices, size_t len) {
    for (size_t i = 0; i < len; i++) {
        header_indices *idxs = &indices[i];
        const header *hdr = &headers[i];
        size_t name_start = ((size_t)hdr->name) - bytes_ptr;
        size_t name_end = name_start + hdr->name_len;
        idxs->name_a = name_start;
        idxs->name_b = name_end;
        size_t val_start = ((size_t)hdr->val) - bytes_ptr;
        size_t val_end = name_start + hdr->val_len;
        idxs->val_a = name_start;
        idxs->val_b = name_end;
    }
}

[Eijebong]

Here's the output of emit=llvm-ir

; ModuleID = 'foo.3vww23kh-cgu.0'
source_filename = "foo.3vww23kh-cgu.0"
target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
target triple = "armv7-none-linux-android"

%Header = type { [0 x i32], i8*, [0 x i32], i32, [0 x i32], i8*, [0 x i32], i32, [0 x i32] }
%HeaderIndices = type { [0 x i32], i32, [0 x i32], i32, [0 x i32], i32, [0 x i32], i32, [0 x i32] }
%"unwind::libunwind::_Unwind_Exception" = type { [0 x i64], i64, [0 x i32], void (i32, %"unwind::libunwind::_Unwind_Exception"*)*, [0 x i32], [20 x i32], [1 x i32] }
%"unwind::libunwind::_Unwind_Context" = type { [0 x i8] }

; foo::record_header_indices
; Function Attrs: nounwind nonlazybind uwtable
define void @_ZN3foo21record_header_indices17h1de73b969897f0d0E(i32 %bytes_ptr, %Header* nocapture readonly %headers, %HeaderIndices* nocapture %indices, i32 %len) unnamed_addr #0 personality i32 (i32, %"unwind::libunwind::_Unwind_Exception"*, %"unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
start:
  %0 = icmp sgt i32 %len, 0
  br i1 %0, label %bb6.preheader, label %bb4

bb6.preheader:                                    ; preds = %start
  %min.iters.check = icmp ult i32 %len, 4
  br i1 %min.iters.check, label %bb6.preheader21, label %vector.memcheck

bb6.preheader21:                                  ; preds = %middle.block, %vector.memcheck, %bb6.preheader
  %iter.sroa.0.010.ph = phi i32 [ 0, %vector.memcheck ], [ 0, %bb6.preheader ], [ %n.vec, %middle.block ]
  br label %bb6

vector.memcheck:                                  ; preds = %bb6.preheader
  %scevgep = getelementptr %HeaderIndices, %HeaderIndices* %indices, i32 %len
  %scevgep14 = getelementptr %Header, %Header* %headers, i32 %len
  %1 = bitcast %Header* %scevgep14 to %HeaderIndices*
  %bound0 = icmp ugt %HeaderIndices* %1, %indices
  %2 = bitcast %HeaderIndices* %scevgep to %Header*
  %bound1 = icmp ugt %Header* %2, %headers
  %found.conflict = and i1 %bound0, %bound1
  br i1 %found.conflict, label %bb6.preheader21, label %vector.ph

vector.ph:                                        ; preds = %vector.memcheck
  %n.vec = and i32 %len, -4
  %broadcast.splatinsert19 = insertelement <4 x i32> undef, i32 %bytes_ptr, i32 0
  %broadcast.splat20 = shufflevector <4 x i32> %broadcast.splatinsert19, <4 x i32> undef, <4 x i32> zeroinitializer
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %vector.ph
  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
  %3 = getelementptr inbounds %Header, %Header* %headers, i32 %index, i32 0, i32 0
  %4 = bitcast i32* %3 to <16 x i32>*
  %wide.vec = load <16 x i32>, <16 x i32>* %4, align 4
  %strided.vec = shufflevector <16 x i32> %wide.vec, <16 x i32> undef, <4 x i32> <i32 0, i32 4, i32 8, i32 12>
  %strided.vec16 = shufflevector <16 x i32> %wide.vec, <16 x i32> undef, <4 x i32> <i32 1, i32 5, i32 9, i32 13>
  %strided.vec17 = shufflevector <16 x i32> %wide.vec, <16 x i32> undef, <4 x i32> <i32 2, i32 6, i32 10, i32 14>
  %strided.vec18 = shufflevector <16 x i32> %wide.vec, <16 x i32> undef, <4 x i32> <i32 3, i32 7, i32 11, i32 15>
  %5 = sub <4 x i32> %strided.vec, %broadcast.splat20
  %6 = add <4 x i32> %5, %strided.vec16
  %7 = sub <4 x i32> %strided.vec17, %broadcast.splat20
  %8 = add <4 x i32> %7, %strided.vec18
  %9 = getelementptr inbounds %HeaderIndices, %HeaderIndices* %indices, i32 %index, i32 7
  %10 = getelementptr inbounds i32, i32* %9, i32 -3
  %11 = bitcast i32* %10 to <16 x i32>*
  %12 = shufflevector <4 x i32> %5, <4 x i32> %6, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %13 = shufflevector <4 x i32> %7, <4 x i32> %8, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
  %interleaved.vec = shufflevector <8 x i32> %12, <8 x i32> %13, <16 x i32> <i32 0, i32 4, i32 8, i32 12, i32 1, i32 5, i32 9, i32 13, i32 2, i32 6, i32 10, i32 14, i32 3, i32 7, i32 11, i32 15>
  store <16 x i32> %interleaved.vec, <16 x i32>* %11, align 4
  %index.next = add i32 %index, 4
  %14 = icmp eq i32 %index.next, %n.vec
  br i1 %14, label %middle.block, label %vector.body, !llvm.loop !1

middle.block:                                     ; preds = %vector.body
  %cmp.n = icmp eq i32 %n.vec, %len
  br i1 %cmp.n, label %bb4, label %bb6.preheader21

bb4:                                              ; preds = %bb6, %middle.block, %start
  ret void

bb6:                                              ; preds = %bb6.preheader21, %bb6
  %iter.sroa.0.010 = phi i32 [ %15, %bb6 ], [ %iter.sroa.0.010.ph, %bb6.preheader21 ]
  %15 = add nuw nsw i32 %iter.sroa.0.010, 1
  %16 = getelementptr inbounds %Header, %Header* %headers, i32 %iter.sroa.0.010, i32 0, i32 0
  %17 = load i32, i32* %16, align 4
  %18 = sub i32 %17, %bytes_ptr
  %19 = getelementptr inbounds %Header, %Header* %headers, i32 %iter.sroa.0.010, i32 3
  %20 = load i32, i32* %19, align 4
  %21 = add i32 %18, %20
  %22 = getelementptr inbounds %HeaderIndices, %HeaderIndices* %indices, i32 %iter.sroa.0.010, i32 0, i32 0
  store i32 %18, i32* %22, align 4
  %23 = getelementptr inbounds %HeaderIndices, %HeaderIndices* %indices, i32 %iter.sroa.0.010, i32 3
  store i32 %21, i32* %23, align 4
  %24 = getelementptr inbounds %Header, %Header* %headers, i32 %iter.sroa.0.010, i32 5
  %25 = bitcast i8** %24 to i32*
  %26 = load i32, i32* %25, align 4
  %27 = sub i32 %26, %bytes_ptr
  %28 = getelementptr inbounds %Header, %Header* %headers, i32 %iter.sroa.0.010, i32 7
  %29 = load i32, i32* %28, align 4
  %30 = add i32 %27, %29
  %31 = getelementptr inbounds %HeaderIndices, %HeaderIndices* %indices, i32 %iter.sroa.0.010, i32 5
  store i32 %27, i32* %31, align 4
  %32 = getelementptr inbounds %HeaderIndices, %HeaderIndices* %indices, i32 %iter.sroa.0.010, i32 7
  store i32 %30, i32* %32, align 4
  %exitcond = icmp eq i32 %15, %len
  br i1 %exitcond, label %bb4, label %bb6, !llvm.loop !3
}

; Function Attrs: nonlazybind uwtable
declare i32 @rust_eh_personality(i32, %"unwind::libunwind::_Unwind_Exception"*, %"unwind::libunwind::_Unwind_Context"*) unnamed_addr #1

attributes #0 = { nounwind nonlazybind uwtable "target-features"="+v7,+thumb-mode,+thumb2,+vfp3,+d16,-neon,+neon" }
attributes #1 = { nonlazybind uwtable "target-cpu"="generic" "target-features"="+v7,+thumb-mode,+thumb2,+vfp3,+d16,-neon,+neon" }

!llvm.module.flags = !{!0}

!0 = !{i32 2, !"RtLibUseGOT", i32 1}
!1 = distinct !{!1, !2}
!2 = !{!"llvm.loop.isvectorized", i32 1}
!3 = distinct !{!3, !2}`

[SimonSapin]

Thanks @est31 for reducing and finding a work-around! I’ve filed it at hyperium/hyper#1671