[Bug d/94496] [D] Use aggressive optimizations in release mode
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94496 Witold Baryluk changed: What|Removed |Added CC||witold.baryluk+gcc at gmail dot co ||m --- Comment #1 from Witold Baryluk --- We are close to making 'in' mean 'scope const', it is already available as a preview in dmd 2.092: https://dlang.org/changelog/2.092.0.html#preview-in
[Bug d/95120] New: [D] Incorrectly allows fqdn access to imported symbols when doing selective imports.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95120
Bug ID: 95120
Summary: [D] Incorrectly allows fqdn access to imported symbols
when doing selective imports.
Product: gcc
Version: 10.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
Created attachment 48529
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48529&action=edit
Example of incorrectly accepted d source by gdc-10
gdc does violatate D language spec:
https://dlang.org/spec/module.html#selective_imports
4.7 Selective Imports
Specific symbols can be exclusively imported from a module and bound into the
current namespace:
import std.stdio : writeln, foo = write;
void main()
{
std.stdio.writeln("hello!"); // error, std is undefined
writeln("hello!"); // ok, writeln bound into current namespace
write("world"); // error, write is undefined
foo("world");// ok, calls std.stdio.write()
fwritefln(stdout, "abc");// error, fwritefln undefined
}
=
I found that in some weird situations the gdc-10 does behave differently
than dmd and ldc2.
Here are the versions I used:
$ dmd --version
DMD64 D Compiler v2.092.0
Copyright (C) 1999-2020 by The D Language Foundation, All Rights Reserved
written by Walter Bright
$ ldc2 --version
LDC - the LLVM D compiler (1.20.1):
based on DMD v2.090.1 and LLVM 9.0.1
built with LDC - the LLVM D compiler (1.20.1)
Default target: x86_64-pc-linux-gnu
Host CPU: znver1
$ gdc-10 --version
gdc-10 (Debian 10.1.0-1) 10.1.0
$
All on Debian testing/unstable, amd64.
badimport.d =
void main() {
import std.stdio;
import std.algorithm.comparison : min;
static struct S {
int min_;
// int min() { return min_; }
void opOpAssign(string op)(const S other) if (op == "+") {
min_ = std.algorithm.comparison.min(min_, other.min_);
}
}
S x = {3};
x += x;
}
=
(the intention was to use fqdn here, to not reference struct member function
min; using `.min(min_, other.min_)`, is another option, but it actually
shouldn't work either, due to other reasons).
Anyway:
$ gdc-10 badimport.d # Compiles.
$
$ ldc2 badimport.d # Correct error.
badimport.d(11): Error: undefined identifier algorithm in package std, perhaps
add static import std.algorithm;
badimport.d(16): Error: template instance badimport.main.S.opOpAssign!"+" error
instantiating
$
$ dmd badimport.d # Correct error.
badimport.d(11): Error: undefined identifier algorithm in package std, perhaps
add static import std.algorithm;
badimport.d(16): Error: template instance badimport.main.S.opOpAssign!"+" error
instantiating
$
Produced code by gdc-10 does work correctly. However, it shouldn't compile at
all.
>From what I can see, it is some kind of interaction with preceding imports,
that is the `import std.stdio;`. Removing `import std.stdio;` makes gdc-10
correctly report the error and stop compilation.
The test case can be further minimizes, and it attached to the bug.
Same behaviour:
==
import std.stdio;
import std.algorithm.comparison : min;
struct S {
int min_;
void add(const S other) {
min_ = std.algorithm.comparison.min(min_, other.min_);
}
}
void main() {
S x = {3};
x.add(x);
}
==
[Bug d/95120] [D] Incorrectly allows fqdn access to imported symbols when doing selective imports.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95120
--- Comment #1 from Witold Baryluk ---
Further minimized:
==
import std.stdio;
import std.algorithm.comparison : min;
int main() {
return std.algorithm.comparison.min(3, 2);
}
==
Removing `import std.stdio;`, results in the same error messages in gdc-10, dmd
and ldc2.
$ gdc badimport.d
badimport.d:5:10: error: undefined identifier ‘std’
5 | return std.algorithm.comparison.min(3, 2);
| ^
$
$ ldc2 badimport.d
badimport.d(5): Error: undefined identifier std
$ dmd badimport.d
badimport.d(5): Error: undefined identifier std
$
it complains about `unknown std`.
When I use `import std.stdio;` at the start, dmd and ldc complain about
`unknown algorithm in package std`.
Not sure if this is something in `std.stdio` package maybe.
[Bug d/95120] [D] Incorrectly allows fqdn access to imported symbols when doing selective imports.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95120 --- Comment #2 from Witold Baryluk --- Created attachment 48530 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48530&action=edit Minimized example
[Bug d/94496] [D] Use aggressive optimizations in release mode
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94496
--- Comment #3 from Witold Baryluk ---
Also about 'nothrow' and Errors.
I would really welcome a flag to compiler that simply terminates all threads
immidetly any Error is thrown at throw location. They aren't really
recoverable. The only option is either catch them really high in the stack or
terminate the program (in D this will I think unwind all the stack and destroy
scoped structs, also call full GC collection, optionally call all class
destrustors, and module destructors). But in many cases terminating the program
at the spot (_exit(2) or _Exit(2), from glibc (not kernel) to terminate all
threads via exit_group).
As of the 'nothrow' itself. I belive it doesn't mean there is 'no thrown
exceptions in the call tree'. I think it means there is no 'uncought exceptions
possibly throw by call to this function'.
```d
extern int g(int x); // not nothrow
int f(int x) nothrow {
try {
return g(x);
throw new MyException("ble");
} catch (Exception e) {
return 1;
}
return 0;
}
```
https://gcc.godbolt.org/z/Y3vNQr
As of the asm pure, considering there is asm volatile, wouldn't it make sense
to not allow 'asm pure volatilve' in the first place in the source?
strict aliasing should be enabled for dynamic arrays, static arrays and normal
pointer to other types. I.e.
```d
void f(int* x, float[] y); // x, y, y.ptr should not alias.
```
```d
void f(int* x, int[] y); // x, y and y.ptr can alias.
```
Also how about using `restrict` automatically for transitively const types?
I.e.
```d
void f(const scope int[] a, int *b); // can't alias. if b aliases a, then it
is UB.
```
[Bug d/95173] New: [D] ICE on some architecture targets when trying to use unknown attribute
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95173
Bug ID: 95173
Summary: [D] ICE on some architecture targets when trying to
use unknown attribute
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
https://explore.dgnu.org/z/bseyKQ
```
import gcc.attribute;
@attribute("foo")
void f() {}
```
This code crashes the compiler when compiling for alpha-linux-gnu, sparc64-elf,
hppa-linux-gnu, mmix-knuth-mmixware, pdp11-aout, lm32-elf and possibly more
with variations. (But sparc64-sun-solaris2.11 for examples works fine).
The code compiles correctly for known attributes.
Other targets do compile without a crash, just with a warning about unknown
attribute. That is correct behaviour.
[Bug d/95174] New: [D] Incorrect compiled functions involving const fixed size arrays
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95174
Bug ID: 95174
Summary: [D] Incorrect compiled functions involving const fixed
size arrays
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
https://explore.dgnu.org/z/LppySp
```
void f(immutable(float[64]) x, float[64] o) {
o[] = x[] * 2.0f;
}
```
and
```
void f(immutable(float[64]) x, float[64] o) {
foreach (i; 0 .. 64) {
o[i] = x[i] * 2.0f;
}
}
```
and
```
void f(immutable(float[64]) x, float[64] o) {
o[1] = x[5] + x[7];
}
```
Is incorrectly compiled to 'nop; ret'
It appears DMD (v2.092) also essentially do the same, and do not perform any
computations in the function.
LDC2 (1.20.1, based on DMD v2.090.1) does generate correct code in some cases
(fully unrolled and fully vectorized in this specific case), but in some other
also do nothing and simply does 'ret' in the function.
As a bonus:
```
void f(immutable(float[4]) x, float[4] o) {
o[2] = x[1] + x[3];
}
import std.stdio : writeln;
void main() {
immutable(float[4]) k = [7.0f, 5.3f, 1.2f, 3.2f];
float[4] o;
f(k, o);
writeln(o);
}
```
prints '[nan, nan, nan, nan]', but it should: '[nan, nan, 8.5, nan]'.
I got the same results using my local gdc version 10.1.0 on amd64.
[Bug d/95174] [D] Incorrect compiled functions involving const fixed size arrays
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95174 --- Comment #2 from Witold Baryluk --- Doh. Of course. My bad. Sorry. static arrays are value type, dynamic arrays are reference type. Changing signature to: ``` void f(immutable(float[64]) x, float[] o); ``` solves the problem.
[Bug d/95198] New: [D] extern(C) private final functions should use 'local' linker attribute
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95198
Bug ID: 95198
Summary: [D] extern(C) private final functions should use
'local' linker attribute
Product: gcc
Version: 10.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
```
module t1;
extern(C)
private final int f() {
return 5;
}
pragma(msg, f.mangleof);
```
`gdc -c t1.d -o t1.o` results in object with this symbols:
D _D2t111__moduleRefZ
D _D2t112__ModuleInfoZ
U _d_dso_registry
T f
W gdc.dso_ctor
W gdc.dso_dtor
u gdc.dso_initialized
u gdc.dso_slot
0016 t _GLOBAL__D_2t1
000b t _GLOBAL__I_2t1
U _GLOBAL_OFFSET_TABLE_
U __start_minfo
U __stop_minfo
Symbol, ' T f' should instead be ' t f'
Additional when using optimizations, I would expect the f to not be emitted at
all, but it is still there (unless compiler decides not to inline it or its
address is not taken and passed around), even with `gdc -O3`.
gcc for C does use LOCAL for static functions and variables in translation
unit. Similarly probably for C++ symbols in anonymous namespaces.
Example of linking issues:
t1.d:
```
module t1;
extern(C)
private final int f() {
return 5;
}
```
t2.d:
```
module t2;
extern(C)
private final int f() {
return 10;
}
```
tm.d:
```
module tm;
void main() {
}
```
$ gdc -O0 -c t1.d -o t1.o
$ gdc -O0 -c t2.d -o t2.o
$ gdc t1.o t2.o tm.d -o t12
/usr/bin/ld: t2.o: in function `f':
t2.d:(.text+0x0): multiple definition of `f'; t1.o:t1.d:(.text+0x0): first
defined here
collect2: error: ld returned 1 exit status
$
This code should link, similar to equivalent code in C. The use case is local
function that is passed in some other module function or method (or static
module constructor for example), to C libraries or other modules as a callback
or for variables a return value.
[Bug d/95198] [D] extern(C) private final functions should use 'local' linker attribute
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95198
--- Comment #1 from Witold Baryluk ---
BTW.
Using:
```
extern(C) private final static int f() { ... }
```
doesn't change anything.
[Bug d/95198] [D] extern(C) private final functions should use 'local' linker attribute
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95198
--- Comment #3 from Witold Baryluk ---
> The main example to demonstrate the current behaviour is correct would be the
> following:
```
extern(C)
private final int f() {
return 5;
}
auto pubf()() {
return f();
}
```
I see, I guess you are right. I don't know how would one go to fix this to work
correctly with existing linkers and not break other code.
Thanks for clarifications.
[Bug d/95250] New: [D] ICE instead of error when trying to use bad template type inside template
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95250
Bug ID: 95250
Summary: [D] ICE instead of error when trying to use bad
template type inside template
Product: gcc
Version: 10.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
https://godbolt.org/z/xWrXP5
Minimized version
```
module m;
import std.traits : Unsigned;
void* f(T)(T a, T b) {
alias UnsignedVoid = Unsigned!(T);
return cast(T)(cast(T)(cast(UnsignedVoid)(a-b) / 2));
}
//static assert(is(typeof(f(null, null)) == void*)); // ICE
static assert(is(typeof(f!(void*)(null, null)) == void*)); // ICE
```
The code is not correct, but on DMD v2.092.0 and LDC 1.20.1 (LLVM 9.0.1) it
does say static assert is false (which is also incorrect), and doesn't crash.
Instead it should say, something like this:
/usr/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/traits.d:7163:13: error:
static assert "Type void* does not have an Unsigned counterpart"
7163 | static assert(false, "Type " ~ T.stringof ~
| ^
Here is a local run, on Linux, amd64.
$ gdc gdc_ice.d
d21: internal compiler error: Segmentation fault
0xbd63ef crash_signal
../../src/gcc/toplev.c:328
0x7f31b746c7ff ???
./signal/../sysdeps/unix/sysv/linux/x86_64/sigaction.c:0
0x71ed1e isAggregate(Type*)
../../src/gcc/d/dmd/opover.c:161
0x71ed1e visit
../../src/gcc/d/dmd/opover.c:586
0x71e935 op_overload(Expression*, Scope*)
../../src/gcc/d/dmd/opover.c:1385
0x6d0b88 Expression::op_overload(Scope*)
../../src/gcc/d/dmd/expression.h:213
0x6d0b88 ExpressionSemanticVisitor::visit(DivExp*)
../../src/gcc/d/dmd/expressionsem.c:6891
0x6cd7b4 semantic(Expression*, Scope*)
../../src/gcc/d/dmd/expressionsem.c:8214
0x6cd7b4 unaSemantic(UnaExp*, Scope*)
../../src/gcc/d/dmd/expressionsem.c:8164
0x6cd7b4 ExpressionSemanticVisitor::visit(CastExp*)
../../src/gcc/d/dmd/expressionsem.c:4203
0x6cd7b4 semantic(Expression*, Scope*)
../../src/gcc/d/dmd/expressionsem.c:8214
0x6cd7b4 unaSemantic(UnaExp*, Scope*)
../../src/gcc/d/dmd/expressionsem.c:8164
0x6cd7b4 ExpressionSemanticVisitor::visit(CastExp*)
../../src/gcc/d/dmd/expressionsem.c:4203
0x6c5a45 semantic(Expression*, Scope*)
../../src/gcc/d/dmd/expressionsem.c:8214
0x74795f StatementSemanticVisitor::visit(ReturnStatement*)
../../src/gcc/d/dmd/statementsem.c:2757
0x74a949 semantic(Statement*, Scope*)
../../src/gcc/d/dmd/statementsem.c:3782
0x74a949 StatementSemanticVisitor::visit(CompoundStatement*)
../../src/gcc/d/dmd/statementsem.c:142
0x743755 semantic(Statement*, Scope*)
../../src/gcc/d/dmd/statementsem.c:3782
0x6e8ba9 FuncDeclaration::semantic3(Scope*)
../../src/gcc/d/dmd/func.c:1711
0x6e8ba9 FuncDeclaration::semantic3(Scope*)
../../src/gcc/d/dmd/func.c:1354
$
$ gdc -v
Using built-in specs.
COLLECT_GCC=gdc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/10/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none:amdgcn-amdhsa:hsa
OFFLOAD_TARGET_DEFAULT=1
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Debian 10.1.0-1'
--with-bugurl=file:///usr/share/doc/gcc-10/README.Bugs
--enable-languages=c,ada,c++,go,brig,d,fortran,objc,obj-c++,m2 --prefix=/usr
--with-gcc-major-version-only --program-suffix=-10
--program-prefix=x86_64-linux-gnu- --enable-shared --enable-linker-build-id
--libexecdir=/usr/lib --without-included-gettext --enable-threads=posix
--libdir=/usr/lib --enable-nls --enable-clocale=gnu --enable-libstdcxx-debug
--enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new
--enable-gnu-unique-object --disable-vtable-verify --enable-plugin
--enable-default-pie --with-system-zlib --enable-libphobos-checking=release
--with-target-system-zlib=auto --enable-objc-gc=auto --enable-multiarch
--disable-werror --with-arch-32=i686 --with-abi=m64
--with-multilib-list=m32,m64,mx32 --enable-multilib --with-tune=generic
--enable-offload-targets=nvptx-none,amdgcn-amdhsa,hsa --without-cuda-driver
--enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu
--target=x86_64-linux-gnu
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 10.1.0 (Debian 10.1.0-1)
$
[Bug c/96275] New: Vectorizer doesn't take into account bitmask condition from branch conditions.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96275
Bug ID: 96275
Summary: Vectorizer doesn't take into account bitmask condition
from branch conditions.
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
https://godbolt.org/z/Gfebjd
With gcc trunk 20200720
If the loop to be vectorized is inside a if condition that check for loop
counter, or there is preceding assert / function return on such condition, the
gcc seems to forgot about it and not take into account in the optimizer /
vectorizer, and still emits the backup scalar code to take care of stragglers
despite it being a dead code.
#include "assert.h"
void fillArray(const unsigned int N, float * restrict a, const float* restrict
b, const float* restrict c) {
//assert(N >= 1024);
for (int i = 0; i < (N & ~31u); i++) {
a[i] = b[0] * c[i];
}
}
produces:
fillArray:
and edi, -32
je .L8
shr edi, 3
vbroadcastssymm1, DWORD PTR [rdx]
xor eax, eax
mov edx, edi
sal rdx, 5
.L3:
vmulps ymm0, ymm1, YMMWORD PTR [rcx+rax]
vmovups YMMWORD PTR [rsi+rax], ymm0
add rax, 32
cmp rax, rdx
jne .L3
vzeroupper
.L8:
ret
but:
#include "assert.h"
void fillArray(const unsigned int N, float * restrict a, const float* restrict
b, const float* restrict c) {
//assert(N >= 1024);
if ((N & 31u) == 0) {
for (int i = 0; i < N; i++) {
a[i] = b[0] * c[i];
}
}
}
produces this sub-optimal code:
fillArray:
mov eax, edi
and eax, 31
jne .L14
testedi, edi
je .L14
lea r8d, [rdi-1]
vmovss xmm1, DWORD PTR [rdx]
cmp r8d, 6
jbe .L8
mov edx, edi
vbroadcastssymm2, xmm1
xor eax, eax
shr edx, 3
sal rdx, 5
.L4:
vmulps ymm0, ymm2, YMMWORD PTR [rcx+rax]
vmovups YMMWORD PTR [rsi+rax], ymm0
add rax, 32
cmp rdx, rax
jne .L4
mov eax, edi
and eax, -8
mov edx, eax
cmp edi, eax
je .L16
vzeroupper
.L3:
mov r9d, edi
sub r8d, eax
sub r9d, eax
cmp r8d, 2
jbe .L6
mov eax, eax
vshufps xmm0, xmm1, xmm1, 0
vmulps xmm0, xmm0, XMMWORD PTR [rcx+rax*4]
vmovups XMMWORD PTR [rsi+rax*4], xmm0
mov eax, r9d
and eax, -4
add edx, eax
cmp r9d, eax
je .L14
.L6:
movsx rax, edx
vmulss xmm0, xmm1, DWORD PTR [rcx+rax*4]
vmovss DWORD PTR [rsi+rax*4], xmm0
lea eax, [rdx+1]
cmp edi, eax
jbe .L14
cdqe
add edx, 2
vmulss xmm0, xmm1, DWORD PTR [rcx+rax*4]
vmovss DWORD PTR [rsi+rax*4], xmm0
cmp edi, edx
jbe .L14
movsx rdx, edx
vmulss xmm1, xmm1, DWORD PTR [rcx+rdx*4]
vmovss DWORD PTR [rsi+rdx*4], xmm1
.L14:
ret
.L16:
vzeroupper
ret
.L8:
xor edx, edx
jmp .L3
Adding `assert(N == (N & ~31u));` doesn't help.
[Bug c/96275] Vectorizer doesn't take into account bitmask condition from branch conditions.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96275
--- Comment #1 from Witold Baryluk ---
FYI.
clang trunk 12 / 76a0c0ee6ffa9c38485776921948d8f930109674, doesn't do that
either:
fillArray: # @fillArray
testdil, 31
jne .LBB0_8
testedi, edi
je .LBB0_8
vmovss xmm0, dword ptr [rdx] # xmm0 = mem[0],zero,zero,zero
mov eax, edi
cmp edi, 32
jae .LBB0_4
xor edx, edx
jmp .LBB0_7
.LBB0_4:
vbroadcastssymm1, xmm0
mov edx, eax
xor edi, edi
and edx, -32
.LBB0_5:# =>This Inner Loop Header: Depth=1
vmulps ymm2, ymm1, ymmword ptr [rcx + 4*rdi]
vmulps ymm3, ymm1, ymmword ptr [rcx + 4*rdi + 32]
vmulps ymm4, ymm1, ymmword ptr [rcx + 4*rdi + 64]
vmulps ymm5, ymm1, ymmword ptr [rcx + 4*rdi + 96]
vmovups ymmword ptr [rsi + 4*rdi], ymm2
vmovups ymmword ptr [rsi + 4*rdi + 32], ymm3
vmovups ymmword ptr [rsi + 4*rdi + 64], ymm4
vmovups ymmword ptr [rsi + 4*rdi + 96], ymm5
add rdi, 32
cmp rdx, rdi
jne .LBB0_5
cmp rdx, rax
je .LBB0_8
.LBB0_7:# =>This Inner Loop Header: Depth=1
vmulss xmm1, xmm0, dword ptr [rcx + 4*rdx]
vmovss dword ptr [rsi + 4*rdx], xmm1
inc rdx
cmp rax, rdx
jne .LBB0_7
.LBB0_8:
vzeroupper
ret
the main inner loop is unrolled / pipelined more aggressively, and the fallback
code is simpler (just handle scalars scalarly), which is unrelated. But the
fallback code is still there.
Changing to different variations of the condition, like `if ((N/32)*32 == N)
{`, `if ((N % 32) == 0) {`, `if ((N & ~31u) == N) {`, `if ((N >> 5) << 5 == N)
{`, doesn't make any difference.
I tried with signed int, and unsigned int. Same effect.
Reassigning to N (after removing constness), i.e. `N = N & ~31u`, or `N = (N >>
5) << 5`, does appear to do something, but if it is inside the condition it is
already too late.
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130 --- Comment #9 from Witold Baryluk --- Indeed, passing -fno-tree-pre in the first example does make it be vectorized. In the mesh_simple.c this corresponds to ONTHEFLY_CONSTANTS being defined, but USE_LOOP_CONSTANTS being not. The SIMPLIFIED can be defined or not, it vectorizes now in both cases. Targeting -march=knm. This is with #define OCTAVES 12, a compile time constant, so compiler fully unrolls the most inner loop. Without -fno-tree-pre: 1230 : 1230: 41 57 push %r15 1232: 62 a1 7d 40 ef c0 vpxord %zmm16,%zmm16,%zmm16 1238: 49 ba 53 ec 85 1a femovabs $0xc4ceb9fe1a85ec53,%r10 123f: b9 ce c4 1242: 41 56 push %r14 1244: c5 7a 10 0d f8 0d 00vmovss 0xdf8(%rip),%xmm9# 2044 <_IO_stdin_used+0x44> 124b: 00 124c: 62 31 7c 48 28 d0 vmovaps %zmm16,%zmm10 1252: 41 55 push %r13 1254: c5 7a 10 3d ec 0d 00vmovss 0xdec(%rip),%xmm15# 2048 <_IO_stdin_used+0x48> 125b: 00 125c: 62 a1 7c 48 28 d0 vmovaps %zmm16,%zmm18 1262: 41 54 push %r12 1264: c5 7a 10 35 e0 0d 00vmovss 0xde0(%rip),%xmm14# 204c <_IO_stdin_used+0x4c> 126b: 00 126c: 49 b9 cd 8c 55 ed d7movabs $0xff51afd7ed558ccd,%r9 1273: af 51 ff 1276: 55 push %rbp 1277: c5 7a 10 2d d1 0d 00vmovss 0xdd1(%rip),%xmm13# 2050 <_IO_stdin_used+0x50> 127e: 00 127f: 49 be 68 66 ac 6a bfmovabs $0xfa8d7ebf6aac6668,%r14 1286: 7e 8d fa 1289: 53 push %rbx 128a: c5 7a 10 25 c2 0d 00vmovss 0xdc2(%rip),%xmm12# 2054 <_IO_stdin_used+0x54> 1291: 00 1292: 48 89 7c 24 f8 mov%rdi,-0x8(%rsp) 1297: c7 44 24 f0 00 00 00movl $0x0,-0x10(%rsp) 129e: 00 129f: c7 44 24 f4 00 00 00movl $0x0,-0xc(%rsp) 12a6: 00 12a7: c5 7a 10 1d a9 0d 00vmovss 0xda9(%rip),%xmm11# 2058 <_IO_stdin_used+0x58> 12ae: 00 12af: 62 e1 7e 08 10 0d a3vmovss 0xda3(%rip),%xmm17# 205c <_IO_stdin_used+0x5c> 12b6: 0d 00 00 12b9: 0f 1f 80 00 00 00 00nopl 0x0(%rax) 12c0: 48 8b 6c 24 f8 mov-0x8(%rsp),%rbp 12c5: 31 f6 xor%esi,%esi 12c7: 31 db xor%ebx,%ebx 12c9: 62 31 7c 48 28 c2 vmovaps %zmm18,%zmm8 12cf: 90 nop 12d0: 8b 54 24 f0 mov-0x10(%rsp),%edx 12d4: 45 31 e4xor%r12d,%r12d 12d7: 62 b1 7c 48 28 f8 vmovaps %zmm16,%zmm7 12dd: 62 c1 7c 48 28 d9 vmovaps %zmm9,%zmm19 12e3: c5 32 11 cc vmovss %xmm9,%xmm9,%xmm4 12e7: eb 26 jmp130f 12e9: 0f 1f 80 00 00 00 00nopl 0x0(%rax) 12f0: c5 ba 59 c4 vmulss %xmm4,%xmm8,%xmm0 12f4: 62 f3 7d 08 0a c0 09vrndscaless $0x9,%xmm0,%xmm0,%xmm0 12fb: c5 fa 2c f0 vcvttss2si %xmm0,%esi 12ff: c4 c1 5a 59 c2 vmulss %xmm10,%xmm4,%xmm0 1304: 62 f3 7d 08 0a c0 09vrndscaless $0x9,%xmm0,%xmm0,%xmm0 130b: c5 fa 2c d0 vcvttss2si %xmm0,%edx 130f: 4c 89 e1mov%r12,%rcx 1312: 62 c1 7c 48 28 e8 vmovaps %zmm8,%zmm21 1318: 48 c1 e9 21 shr$0x21,%rcx 131c: 62 e1 7c 48 28 e4 vmovaps %zmm4,%zmm20 1322: c5 d2 2a ea vcvtsi2ss %edx,%xmm5,%xmm5 1326: 4c 31 e1xor%r12,%rcx 1329: 49 0f af ca imul %r10,%rcx 132d: 48 63 d2movslq %edx,%rdx 1330: c5 e2 2a de vcvtsi2ss %esi,%xmm3,%xmm3 1334: 4f 8d 24 0c lea(%r12,%r9,1),%r12 1338: 48 69 d2 53 42 41 4eimul $0x4e414253,%rdx,%rdx 133f: 62 c2 55 08 9b e2 vfmsub132ss %xmm10,%xmm5,%xmm20 1345: c4 c1 52 58 e9 vaddss %xmm9,%xmm5,%xmm5 134a: 48 8d 01lea(%rcx),%rax 134d: 48 c1 e8 21 shr$0x21,%rax 1351: 62 e2 65 08 9b ec vfmsub132ss %xmm4,%xmm3,%xmm21 1357: 48 31 c1xor%rax,%rcx 135a: 4c 8d ba 53 42 41 4elea0x4e414253(%rdx),%r15 1361: 48 89 cfmov%rcx,%rdi 1364: 48 89 c8mov%rcx,%rax 1367: 48 81 f7 70 46 ab 58xor$0x58ab4670,%rdi 136e: c4 c1 62 58 d9 vaddss %xmm9,%xmm3,%xmm3 1373: 48 c1 e8 21 shr$0x21,%
[Bug c/83584] "ISO C forbids conversion of object pointer to function pointer type" -- no, not really
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83584 Witold Baryluk changed: What|Removed |Added CC||witold.baryluk+gcc at gmail dot co ||m --- Comment #19 from Witold Baryluk --- This is still happening, even when using -std=c11 with gcc 9.2.1 C11 does state in annex J.5.7: http://port70.net/~nsz/c/c11/n1570.html#J.5.7p1 """ J.5.7 Function pointer casts 1 A pointer to an object or to void may be cast to a pointer to a function, allowing data to be invoked as a function (6.5.4). """ I am not sure how else I am supposed to use `dlsym(3)`. Maybe if there was a version of dlsym that instead of returning (void*), would return (void(*)()) or (void (*)(void)) it would help. Maybe it is a POSIX bug then? This issue however is not a duplicate of bug 11234. And the error message is incorrect anyway. Sorry if this was mentioned before.
[Bug c/83584] "ISO C forbids conversion of object pointer to function pointer type" -- no, not really
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83584 --- Comment #20 from Witold Baryluk --- FYI. http://austingroupbugs.net/view.php?id=74#c205 says Note that conversion from a void * pointer to a function pointer as in: fptr = (int (*)(int))dlsym(handle, "my_function"); is not defined by the ISO C Standard. This standard requires this conversion to work correctly on conforming implementations. This is published now as IEEE Std 1003.1-2017, aka POSIX.1-2017: https://pubs.opengroup.org/onlinepubs/9699919799/functions/dlsym.html POSIX standard is free to do so.
[Bug tree-optimization/63945] Missing vectorization optimization
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63945 Witold Baryluk changed: What|Removed |Added CC||witold.baryluk+gcc at gmail dot co ||m --- Comment #1 from Witold Baryluk --- It does vectorize for me on gcc 9.2.1: -march=skylake-avx512 aa.cpp:34:29: optimized: loop vectorized using 32 byte vectors aa.cpp:25:27: optimized: loop vectorized using 32 byte vectors if (val<100.) 1279: c5 fb 10 0b vmovsd (%rbx),%xmm1 127d: c5 fb 10 05 8b 0d 00vmovsd 0xd8b(%rip),%xmm0# 2010 <_IO_stdin_used+0x10> 1284: 00 1285: c5 f9 2f c1 vcomisd %xmm1,%xmm0 1289: 76 2b jbe12b6 <_ZN4TEST4testEv+0xc6> 128b: c4 e2 7d 19 c9 vbroadcastsd %xmm1,%ymm1 1290: 31 c0 xor%eax,%eax 1292: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) c[i] = val*a[i]+b[i]; 1298: c4 c1 7d 10 04 04 vmovupd (%r12,%rax,1),%ymm0 129e: c4 c2 f5 a8 44 05 00vfmadd213pd 0x0(%r13,%rax,1),%ymm1,%ymm0 12a5: c5 fd 11 04 07 vmovupd %ymm0,(%rdi,%rax,1) for (unsigned int i=0; i ::operator delete(__p); 12b6: c5 f8 77vzeroupper Similarly: -march=knm aa.cpp:34:29: optimized: loop vectorized using 64 byte vectors aa.cpp:25:27: optimized: loop vectorized using 64 byte vectors if (val<100.) 15bc: 31 c0 xor%eax,%eax 15be: 66 90 xchg %ax,%ax c[i] = val*a[i]+b[i]; 15c0: 62 f1 fd 48 28 04 01vmovapd (%rcx,%rax,1),%zmm0 15c7: 62 f2 ed 48 a8 04 06vfmadd213pd (%rsi,%rax,1),%zmm2,%zmm0 15ce: 62 d1 fd 48 11 04 01vmovupd %zmm0,(%r9,%rax,1) for (unsigned int i=0; i (plus a lot of handling for unaligned stack). -march=znver2 aa.cpp:34:29: optimized: loop vectorized using 32 byte vectors aa.cpp:25:27: optimized: loop vectorized using 32 byte vectors if (val<100.) 1279: c5 fb 10 0b vmovsd (%rbx),%xmm1 127d: c5 fb 10 05 8b 0d 00vmovsd 0xd8b(%rip),%xmm0# 2010 <_IO_stdin_used+0x10> 1284: 00 1285: c5 f9 2f c1 vcomisd %xmm1,%xmm0 1289: 76 33 jbe12be <_ZN4TEST4testEv+0xce> 128b: c4 e2 7d 19 c9 vbroadcastsd %xmm1,%ymm1 1290: 31 c0 xor%eax,%eax 1292: 66 66 2e 0f 1f 84 00data16 nopw %cs:0x0(%rax,%rax,1) 1299: 00 00 00 00 129d: 0f 1f 00nopl (%rax) c[i] = val*a[i]+b[i]; 12a0: c4 c1 7d 10 04 04 vmovupd (%r12,%rax,1),%ymm0 12a6: c4 c2 f5 a8 44 05 00vfmadd213pd 0x0(%r13,%rax,1),%ymm1,%ymm0 12ad: c5 fd 11 04 07 vmovupd %ymm0,(%rdi,%rax,1) for (unsigned int i=0; i -march=core2 aa.cpp:34:29: optimized: loop vectorized using 16 byte vectors aa.cpp:25:27: optimized: loop vectorized using 16 byte vectors if (val<100.) 1276: f2 0f 10 13 movsd (%rbx),%xmm2 127a: f2 0f 10 05 8e 0d 00movsd 0xd8e(%rip),%xmm0# 2010 <_IO_stdin_used+0x10> 1281: 00 1282: 66 0f 2f c2 comisd %xmm2,%xmm0 1286: 76 40 jbe12c8 <_ZN4TEST4testEv+0xd8> 1288: 31 c0 xor%eax,%eax 128a: 66 0f 14 d2 unpcklpd %xmm2,%xmm2 128e: 66 90 xchg %ax,%ax c[i] = val*a[i]+b[i]; 1290: f3 0f 7e 44 05 00 movq 0x0(%rbp,%rax,1),%xmm0 1296: f3 41 0f 7e 0c 04 movq (%r12,%rax,1),%xmm1 129c: 66 0f 16 44 05 08 movhpd 0x8(%rbp,%rax,1),%xmm0 12a2: 66 0f 59 c2 mulpd %xmm2,%xmm0 12a6: 66 41 0f 16 4c 04 08movhpd 0x8(%r12,%rax,1),%xmm1 12ad: 66 0f 58 c1 addpd %xmm1,%xmm0 12b1: 66 0f 13 04 07 movlpd %xmm0,(%rdi,%rax,1) 12b6: 66 0f 17 44 07 08 movhpd %xmm0,0x8(%rdi,%rax,1) for (unsigned int i=0; i Looks all pretty optimally vectorized to me. The code can be made even better, if you ensure proper alignment of std::vector arrrays, which they might not be at the moment.
[Bug tree-optimization/92130] New: Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130 Bug ID: 92130 Summary: Missed vectorization for iteration dependent loads and simple multiplicative accumulators Product: gcc Version: 9.2.1 Status: UNCONFIRMED Severity: normal Priority: P3 Component: tree-optimization Assignee: unassigned at gcc dot gnu.org Reporter: witold.baryluk+gcc at gmail dot com Target Milestone: --- Created attachment 47051 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47051&action=edit Perlin2D noise mesh generation So, I do have pretty complex multi level loop spread across many functions, but it can be all vectorized, but under certain scenarios gcc does not vectorize it with gcc 9.2.1 I am attaching somehow simplified code with few defines inside to play with it. The one exposed by default present the biggest challenge to gcc, despite me able to vectorize it manually. I tested this on SSE2, AVX2 (cascadelake and znver2), AVX512 (-march=knm and -march=skylake-avx512) and ARM SVE, with all same effects. I am using associative math and other flags mentioned in the sourcefile at the top. The high level overview is like this: input: A, F, W, maxO, sufficiently aligned d. foreach y: foreach x: float v = 0.0 float a = 1.0 float f = 1.0 foreach o in [0, maxO): v += a * g(f * x, f * y, o, h(o, p)) a *= A f *= F d[y*W + x] = v where both g and h are pure functions (relatively complex tho) with no control flow or data dependent flow. In some situations if a and f are replaced by a precomputed table of coefficient for every o, and then used as v += a[o] * g(f[o] * x, f[o] * y, h(o, p)), it does vectorize, but not always. h(o, p) could also be precomputed, but I didn't bother as it appears to not have any bad effect on vectorizer. Vectorizater should vectorize along the 'foreach x', and compute multiple x-s per-lane completely independently. It is true that when updating a and f, each lane need to be duplicated, but that can be done by computing it scalarly, and then broadcasting, or by repeating same constants updates in each lane.
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130 --- Comment #1 from Witold Baryluk --- Created attachment 47052 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47052&action=edit Minimized test case
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130 --- Comment #2 from Witold Baryluk --- Added a minimized test case that has only one outer loop, and f and h are removed for simple inlined replacement. Example diagnostic: $ gcc -std=c17 -march=knm -O3 -ffast-math -fassociative-math -ftree-vectorizer-verbose=2 -fopt-info-vec-all -ggdb -Wall mesh_minimal.c -o mesh_minimal_knm -lm mesh_minimal.c:34:3: missed: couldn't vectorize loop mesh_minimal.c:34:3: missed: not vectorized: latch block not empty. mesh_minimal.c:33:13: note: vectorized 0 loops in function.
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130
--- Comment #3 from Witold Baryluk ---
If only the frequency is updated in the inner loop:
frequency *= 2.131f;
function fill_data is vectorized:
mesh_minimal.c:34:3: optimized: loop vectorized using 64 byte vectors
mesh_minimal.c:33:13: note: vectorized 1 loops in function.
However if amplitude is updated in the inner loop:
amplitude *= 0.781f;
function fill_data is NOT vectorized.
mesh_minimal.c:34:3: missed: couldn't vectorize loop
mesh_minimal.c:34:3: missed: not vectorized: latch block not empty.
mesh_minimal.c:33:13: note: vectorized 0 loops in function.
Here for reference:
/* line 20 */ static float perlin1d(float x) {
float accum = 0.0;
float frequency = 1.0;
float amplitude = 1.0;
for (int i = 0; i < 8; i++) {
accum += amplitude * (sinf(x * frequency + (float)i));
frequency *= 2.131f;
amplitude *= 0.781f;
}
return accum;
}
__attribute__((noinline))
/* line 33 */ static void fill_data(int width, float * __restrict__
height_data, float scale) {
/* line 34 */ for (int i = 0; i < width; i++) {
height_data[i] = perlin1d(i);
}
}
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130
--- Comment #4 from Witold Baryluk ---
If I reduce minimized test case even further:
only frequency update: VECTORIZED:
static float perlin1d(float x) {
float accum = 0.0f;
float amplitude = 1.0f;
float frequency = 1.0f;
for (int i = 0; i < 8; i++) {
accum += amplitude * sinf(x * frequency);
frequency *= 2.131f;
}
return accum;
}
__attribute__((noinline))
static void fill_data(int width, float * __restrict__ height_data, float scale)
{
for (int i = 0; i < width; i++) {
height_data[i] = perlin1d(i);
}
}
only amplitude update: VECTORIZED:
static float perlin1d(float x) {
float accum = 0.0f;
float amplitude = 1.0f;
float frequency = 1.0f;
for (int i = 0; i < 8; i++) {
accum += amplitude * sinf(x * frequency);
amplitude *= 0.781f;
}
return accum;
}
__attribute__((noinline))
static void fill_data(int width, float * __restrict__ height_data, float scale)
{
for (int i = 0; i < width; i++) {
height_data[i] = perlin1d(i);
}
}
both frequency and amplitude update: NOT VECTORIZED:
static float perlin1d(float x) {
float accum = 0.0f;
float amplitude = 1.0f;
float frequency = 1.0f;
for (int i = 0; i < 8; i++) {
accum += amplitude * sinf(x * frequency);
amplitude *= 0.781f;
frequency *= 2.131f;
}
return accum;
}
__attribute__((noinline))
static void fill_data(int width, float * __restrict__ height_data, float scale)
{
for (int i = 0; i < width; i++) {
height_data[i] = perlin1d(i);
}
}
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130
--- Comment #5 from Witold Baryluk ---
As a bonus:
static float perlin1d(float x) {
float accum = 0.0f;
for (int i = 0; i < 8; i++) {
accum += powf(0.781f, i) * sinf(x * powf(2.131f, i));
}
return accum;
}
claims to be vectorized, but really isn't, and has non inline or lowered calls
to sinf and expf_finite.
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130
--- Comment #6 from Witold Baryluk ---
I also tested clang with LLVM 10~svn374655 and it does vectorize the loop
properly, even when both frequency and amplitude variables are updated every
loop.
It still doesn't inline calls to sinf, even if I set -fno-math-errno and other
things from -ffast-math. My random guess is that it is because there is no
hardware support for vectorized sinf, and there is no vectorized variant of
sinf software implementation either. If I provide my own version of sinf using
simple Taylor expansion, clang fully vectorized the code:
401320: 62 e1 7d 58 fe 3d 56vpaddd 0xd56(%rip){1to16},%zmm0,%zmm23
# 402080 <_IO_stdin_used+0x80>
401327: 0d 00 00
40132a: 62 61 7c 48 5b c0 vcvtdq2ps %zmm0,%zmm24
401330: 62 a1 7c 48 5b ff vcvtdq2ps %zmm23,%zmm23
401336: 62 f1 7c 48 10 4c 24vmovups 0x140(%rsp),%zmm1
40133d: 05
40133e: 62 61 3c 40 59 d1 vmulps %zmm1,%zmm24,%zmm26
401344: 62 61 44 40 59 f9 vmulps %zmm1,%zmm23,%zmm31
40134a: 62 f1 7c 48 10 4c 24vmovups 0x100(%rsp),%zmm1
401351: 04
401352: 62 61 3c 40 59 d9 vmulps %zmm1,%zmm24,%zmm27
401358: 62 f1 44 40 59 c9 vmulps %zmm1,%zmm23,%zmm1
40135e: 62 01 2c 40 59 ca vmulps %zmm26,%zmm26,%zmm25
401364: 62 f1 7c 48 10 54 24vmovups 0x80(%rsp),%zmm2
40136b: 02
40136c: 62 61 3c 40 59 e2 vmulps %zmm2,%zmm24,%zmm28
401372: 62 f1 44 40 59 d2 vmulps %zmm2,%zmm23,%zmm2
401378: 62 02 25 40 ac ca vfnmadd213ps %zmm26,%zmm27,%zmm25
40137e: 62 f1 7c 48 10 5c 24vmovups 0x40(%rsp),%zmm3
401385: 01
401386: 62 61 3c 40 59 eb vmulps %zmm3,%zmm24,%zmm29
40138c: 62 f1 44 40 59 db vmulps %zmm3,%zmm23,%zmm3
401392: 62 01 1c 40 59 d4 vmulps %zmm28,%zmm28,%zmm26
401398: 62 01 04 40 59 df vmulps %zmm31,%zmm31,%zmm27
40139e: 62 02 15 40 ac d4 vfnmadd213ps %zmm28,%zmm29,%zmm26
4013a4: 62 f1 7c 48 10 6c 24vmovups -0x40(%rsp),%zmm5
4013ab: ff
4013ac: 62 f1 3c 40 59 e5 vmulps %zmm5,%zmm24,%zmm4
4013b2: 62 f1 44 40 59 ed vmulps %zmm5,%zmm23,%zmm5
4013b8: 62 61 6c 48 59 e2 vmulps %zmm2,%zmm2,%zmm28
4013be: 62 f1 7c 48 10 7c 24vmovups -0x80(%rsp),%zmm7
4013c5: fe
4013c6: 62 f1 3c 40 59 f7 vmulps %zmm7,%zmm24,%zmm6
4013cc: 62 f1 44 40 59 ff vmulps %zmm7,%zmm23,%zmm7
4013d2: 62 61 5c 48 59 ec vmulps %zmm4,%zmm4,%zmm29
4013d8: 62 61 54 48 59 f5 vmulps %zmm5,%zmm5,%zmm30
4013de: 62 62 4d 48 ac ec vfnmadd213ps %zmm4,%zmm6,%zmm29
4013e4: 62 d1 3c 40 59 e3 vmulps %zmm11,%zmm24,%zmm4
4013ea: 62 d1 44 40 59 f3 vmulps %zmm11,%zmm23,%zmm6
4013f0: 62 02 75 48 ac df vfnmadd213ps %zmm31,%zmm1,%zmm27
4013f6: 62 d1 3c 40 59 cc vmulps %zmm12,%zmm24,%zmm1
4013fc: 62 41 44 40 59 fc vmulps %zmm12,%zmm23,%zmm31
401402: 62 71 5c 48 59 c4 vmulps %zmm4,%zmm4,%zmm8
401408: 62 62 65 48 ac e2 vfnmadd213ps %zmm2,%zmm3,%zmm28
40140e: 62 72 75 48 ac c4 vfnmadd213ps %zmm4,%zmm1,%zmm8
401414: 62 d1 3c 40 59 ce vmulps %zmm14,%zmm24,%zmm1
40141a: 62 d1 44 40 59 d6 vmulps %zmm14,%zmm23,%zmm2
401420: 62 62 45 48 ac f5 vfnmadd213ps %zmm5,%zmm7,%zmm30
401426: 62 d1 3c 40 59 df vmulps %zmm15,%zmm24,%zmm3
40142c: 62 d1 44 40 59 e7 vmulps %zmm15,%zmm23,%zmm4
401432: 62 f1 74 48 59 e9 vmulps %zmm1,%zmm1,%zmm5
401438: 62 f1 4c 48 59 fe vmulps %zmm6,%zmm6,%zmm7
40143e: 62 71 6c 48 59 ca vmulps %zmm2,%zmm2,%zmm9
401444: 62 f2 65 48 ac e9 vfnmadd213ps %zmm1,%zmm3,%zmm5
40144a: 62 b1 3c 40 59 c9 vmulps %zmm17,%zmm24,%zmm1
401450: 62 f2 05 40 ac fe vfnmadd213ps %zmm6,%zmm31,%zmm7
401456: 62 b1 44 40 59 d9 vmulps %zmm17,%zmm23,%zmm3
40145c: 62 b1 3c 40 59 f2 vmulps %zmm18,%zmm24,%zmm6
401462: 62 21 44 40 59 fa vmulps %zmm18,%zmm23,%zmm31
401468: 62 72 5d 48 ac ca vfnmadd213ps %zmm2,%zmm4,%zmm9
40146e: 62 f1 74 48 59 d1 vmulps %zmm1,%zmm1,%zmm2
401474: 62 f1 64 48 59 e3 vmulps %zmm3,%zmm3,%zmm4
40147a: 62 f2 4d 48 ac d1 vfnmadd213ps %zmm1,%zmm6,%zmm2
401480: 62 f2 05 40 ac e3 vfnmadd213ps %zmm3,%zmm31,%zmm4
401486: 62 b1 3c 40 59 cc vmulps %zmm20,%zmm24,%zmm1
40148c: 62 b1 3c 40 59 dd vmulps %zmm21,%zmm24,%zmm3
401492: 62 f1 74 48 59 f1 vmulps %zmm1,%zmm1,%zmm6
401498: 62 21 44 40 59 fc vmulps %zmm20,%zmm23,%zmm31
40149e: 62 f2 65 48 ac f1 vfnmadd213ps %zmm1,%zmm3,%z
[Bug tree-optimization/92130] Missed vectorization for iteration dependent loads and simple multiplicative accumulators
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130 --- Comment #7 from Witold Baryluk --- Online examples: https://gcc.godbolt.org/z/Nyjty3
[Bug c/100257] New: poor codegen with vcvtph2ps / stride of 6
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100257
Bug ID: 100257
Summary: poor codegen with vcvtph2ps / stride of 6
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
gcc (Compiler-Explorer-Build) 12.0.0 20210424 (experimental)
https://godbolt.org/z/n6ooMdnz8
This C code:
```
#include
#include
#include
struct float3 {
float f1;
float f2;
float f3;
};
struct util_format_r16g16b16_float {
uint16_t r;
uint16_t g;
uint16_t b;
};
static inline struct float3 _mesa_half3_to_float3(uint16_t val_0, uint16_t
val_1, uint16_t val_2) {
#if defined(__F16C__)
//const __m128i in = {val_0, val_1, val_2};
//__m128 out;
//__asm volatile("vcvtph2ps %1, %0" : "=v"(out) : "v"(in));
const __m128i in = _mm_setr_epi16(val_0, val_1, val_2, 0, 0, 0, 0, 0);
const __m128 out = _mm_cvtph_ps(in);
const struct float3 r = {out[0], out[1], out[2]};
return r;
#endif
}
void
util_format_r16g16b16_float_unpack_rgba_float(void *restrict dst_row, const
uint8_t *restrict src, unsigned width)
{
float *dst = dst_row;
for (unsigned x = 0; x < width; x += 1) {
const struct util_format_r16g16b16_float pixel;
memcpy(&pixel, src, sizeof pixel);
struct float3 r = _mesa_half3_to_float3(pixel.r, pixel.g, pixel.b);
dst[0] = r.f1; /* r */
dst[1] = r.f2; /* g */
dst[2] = r.f3; /* b */
dst[3] = 1; /* a */
src += 6;
dst += 4;
}
}
```
Is compiled "poorly" by gcc, even worse when compiled on i386 (with -mf16c
enabled) when using -FPIE.
Example:
gcc -O3 -m32 -march=znver2 -mfpmath=sse -fPIE
util_format_r16g16b16_float_unpack_rgba_float:
pushebp
pushedi
pushesi
pushebx
sub esp, 28
mov ecx, DWORD PTR 56[esp]
mov edx, DWORD PTR 48[esp]
call__x86.get_pc_thunk.ax
add eax, OFFSET FLAT:_GLOBAL_OFFSET_TABLE_
mov ebx, DWORD PTR 52[esp]
testecx, ecx
je .L8
vmovss xmm3, DWORD PTR .LC0@GOTOFF[eax]
xor esi, esi
xor ebp, ebp
vpxor xmm2, xmm2, xmm2
.L3:
mov eax, DWORD PTR [ebx]
vmovss DWORD PTR 12[edx], xmm3
add ebx, 6
add edx, 16
inc esi
mov ecx, eax
vmovd xmm0, eax
shr ecx, 16
mov edi, ecx
movzx ecx, WORD PTR -2[ebx]
vpinsrw xmm0, xmm0, edi, 1
vmovd xmm1, ecx
vpinsrw xmm1, xmm1, ebp, 1
vpunpckldq xmm0, xmm0, xmm1
vpunpcklqdq xmm0, xmm0, xmm2
vcvtph2ps xmm0, xmm0
vmovss DWORD PTR -16[edx], xmm0
vextractps DWORD PTR -12[edx], xmm0, 1
vextractps DWORD PTR -8[edx], xmm0, 2
cmp DWORD PTR 56[esp], esi
jne .L3
.L8:
add esp, 28
pop ebx
pop esi
pop edi
pop ebp
ret
.LC0:
.long 1065353216
__x86.get_pc_thunk.ax:
mov eax, DWORD PTR [esp]
ret
clang:
util_format_r16g16b16_float_unpack_rgba_float: #
@util_format_r16g16b16_float_unpack_rgba_float
mov eax, dword ptr [esp + 12]
testeax, eax
je .LBB0_3
mov ecx, dword ptr [esp + 8]
mov edx, dword ptr [esp + 4]
.LBB0_2:# =>This Inner Loop Header: Depth=1
vmovd xmm0, dword ptr [ecx] # xmm0 = mem[0],zero,zero,zero
vpinsrw xmm0, xmm0, word ptr [ecx + 4], 2
add ecx, 6
vcvtph2ps xmm0, xmm0
vmovss dword ptr [edx], xmm0
vextractps dword ptr [edx + 4], xmm0, 1
vextractps dword ptr [edx + 8], xmm0, 2
mov dword ptr [edx + 12], 1065353216
add edx, 16
dec eax
jne .LBB0_2
.LBB0_3:
ret
clang code is essentially optimal.
The issue persist if I use `vcvtph2ps` directly via asm, or via intrinsics.
The issue might be the src stride, of 6, instead 8, that is confusing gcc.
Additionally, constant 1065353216 (which is weird, I would expect it to be 0),
is stored in data section, instead inline as immediate, this makes code
actually larger, and in PIE mode, requires extra pointer trickery, and on -m32,
even calling extra function.
Even without -fPIE the main loop has poor codegen even on x86-64 / amd64
compared to clang or what I would considered good code.
gcc -m64 -O3 -march=native
util_format_r16g16b16_float_unpack_rgba_float:
testedx, edx
je .L8
mov edx, edx
sal rdx, 4
vmovss xmm3,
[Bug tree-optimization/96275] Vectorizer doesn't take into account bitmask condition from branch conditions.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96275 --- Comment #3 from Witold Baryluk --- Thanks for looking into that. I just wanted to update that this still suboptimal in current gcc trunk 20201226. While clang produces superior code.
[Bug d/98457] New: [d] writef!"%s" doesn't work with MonoTime / SysTick
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98457
Bug ID: 98457
Summary: [d] writef!"%s" doesn't work with MonoTime / SysTick
Product: gcc
Version: 10.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
void main() {
import std.stdio;
import core.time : MonoTime;
writef!"%s"(MonoTime.currTime());
}
Doesn't compile with gdc 10.2.1:
$ gdc test_monotime.d
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/core/time.d:2405:16: error: static
variable _ticksPerSecond cannot be read at compile time
2405 | return _ticksPerSecond[_clockIdx];
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/core/time.d:2418:99: note: called
from here: ticksPerSecond()
2418 | return "MonoTime(" ~ signedToTempString(_ticks, 10) ~ "
ticks, " ~ signedToTempString(ticksPerSecond, 10) ~ " ticks per second)";
|
^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/core/time.d:2418:98: note: called
from here: signedToTempString(ticksPerSecond(), 10u)
2418 | return "MonoTime(" ~ signedToTempString(_ticks, 10) ~ "
ticks, " ~ signedToTempString(ticksPerSecond, 10) ~ " ticks per second)";
|
^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:3353:28: note: called
from here: val.toString()
3353 | put(w, val.toString());
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:3353:12: note: called
from here: put(w, val.toString())
3353 | put(w, val.toString());
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:3672:21: note: called
from here: formatObject(w, val, f)
3672 | formatObject(w, val, f);
| ^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:568:28: note: called
from here: formatValue(w, _param_2, spec)
568 | formatValue(w, args[i], spec);
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:5767:28: note: called
from here: formattedWrite(w, fmt, _param_1)
5767 | auto n = formattedWrite(w, fmt, args);
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:5729:16: note: called
from here: format("%s", MonoTimeImpl(0L))
5729 | .format(fmt, Args.init);
|^
/usr/lib/gcc/x86_64-linux-gnu/10/include/d/std/format.d:5733:2: note: called
from here: (*function () => null)()
5733 | }();
| ^
(null):0: confused by earlier errors, bailing out
Adding manually .toString() makes it work (at the expense of possible extra
allocation).
No issues in ldc2 1.24.0 or dmd2 2.095.0-beta.1
It doesn't look like issue in phobos, but something deeper.
[Bug d/98457] [d] writef!"%s" doesn't work with MonoTime / SysTick
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98457
--- Comment #1 from Witold Baryluk ---
Godbolt link: https://godbolt.org/z/q3bzhP
with gcc trunk 20201217 and a bit more diagnostic
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/core/time.d:2405:16:
error: static variable _ticksPerSecond cannot be read at compile time
2405 | return _ticksPerSecond[_clockIdx];
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/core/time.d:2418:99:
note: called from here: ticksPerSecond()
2418 | return "MonoTime(" ~ signedToTempString(_ticks, 10) ~ "
ticks, " ~ signedToTempString(ticksPerSecond, 10) ~ " ticks per second)";
|
^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/core/time.d:2418:98:
note: called from here: signedToTempString(ticksPerSecond(), 10u)
2418 | return "MonoTime(" ~ signedToTempString(_ticks, 10) ~ "
ticks, " ~ signedToTempString(ticksPerSecond, 10) ~ " ticks per second)";
|
^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:3353:28:
note: called from here: val.toString()
3353 | put(w, val.toString());
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:3353:12:
note: called from here: put(w, val.toString())
3353 | put(w, val.toString());
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:3672:21:
note: called from here: formatObject(w, val, f)
3672 | formatObject(w, val, f);
| ^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:568:28:
note: called from here: formatValue(w, _param_2, spec)
568 | formatValue(w, args[i], spec);
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:5767:28:
note: called from here: formattedWrite(w, fmt, _param_1)
5767 | auto n = formattedWrite(w, fmt, args);
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:5729:16:
note: called from here: format("%s", MonoTimeImpl(0L))
5729 | .format(fmt, Args.init);
|^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/format.d:5733:2:
note: called from here: (*function () => null)()
5733 | }();
| ^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/stdio.d:3754:15:
error: template instance std.format.checkFormatException!("�}�",
MonoTimeImpl!cast(ClockType)0) error instantiating
3754 | alias e = checkFormatException!(fmt, A);
| ^
:4:14: note: instantiated from here: writef!("%s",
MonoTimeImpl!cast(ClockType)0)
4 | writef!"%s"(MonoTime.currTime());
| ^
/opt/compiler-explorer/gcc-trunk-20201227/lib/gcc/x86_64-linux-gnu/11.0.0/include/d/std/stdio.d:3755:5:
note: while evaluating: static assert(!e)
3755 | static assert(!e, e.msg);
| ^
Compiler returned: 1
[Bug d/98494] New: libphobos: std.process Config.stderrPassThrough missing
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98494
Bug ID: 98494
Summary: libphobos: std.process Config.stderrPassThrough
missing
Product: gcc
Version: 10.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
It appears that gdc version of libphobos is somehow lagging in some aspects
behind upstream.
One of the things I see missing, is `Config.stderrPassThrough` in std.process.
I see it was added upstream about 12 months ago:
enum Config {
...
/**
By default, the $(LREF execute) and $(LREF executeShell) functions
will capture child processes' both stdout and stderr. This can be
undesirable if the standard output is to be processed or otherwise
used by the invoking program, as `execute`'s result would then
contain a mix of output and warning/error messages.
Specify this flag when calling `execute` or `executeShell` to
cause invoked processes' stderr stream to be sent to $(REF stderr,
std,stdio), and only capture and return standard output.
This flag has no effect on $(LREF spawnProcess) or $(LREF spawnShell).
*/
stderrPassThrough = 128,
}
The implementation usage of this is relatively small and easy to backport:
in executeImpl:
-auto p = pipeFunc(commandLine, Redirect.stdout | Redirect.stderrToStdout,
- env, config, workDir, extraArgs);
+auto redirect = (config & Config.stderrPassThrough)
+? Redirect.stdout
+: Redirect.stdout | Redirect.stderrToStdout;
+
+auto p = pipeFunc(commandLine, redirect,
+ env, config, workDir, extraArgs);
There are some other minor changes there, but nothing functionally significant.
Mostly unittests and minor signature changes (adding `scope` to many input
parameters).
Thank you.
[Bug d/100769] New: [D] memcmp() == 0 for small constant strings not folded
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100769
Bug ID: 100769
Summary: [D] memcmp() == 0 for small constant strings not
folded
Product: gcc
Version: 10.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
I expect this D code to be quite optimal, but it isn't.
```
extern(C) int memcmp(const void *s1, const void *s2, size_t n);
int recognize3(const char* s) {
return memcmp(s, "stract class", 12) == 0;
}
```
https://godbolt.org/z/vx17WK9rs
It produces a call to memcmp, instead of inlining and specializing the code for
this specific case.
int example.recognize3(const(char*)):
sub rsp, 8
mov edx, 12
mov esi, OFFSET FLAT:.LC0
callmemcmp
testeax, eax
seteal
add rsp, 8
movzx eax, al
ret
ldc2 1.24.0 (for D) and clang 11.0.1-2 (for C and C++), and gcc 10.2.1 (for C
and C++) produce close to optimal codes. Similarly ldc2 1.26.0 (for D), and gcc
11.1 (for C and C++):
int example.recognize3(const(char*)):
movabs rcx, 7142836979195081843
xor rcx, qword ptr [rdi]
mov edx, dword ptr [rdi + 8]
xor rdx, 1936941420
xor eax, eax
or rdx, rcx
seteal
ret
and
recognize3:
movabs rax, 7142836979195081843
cmp QWORD PTR [rdi], rax
je .L6
.L2:
mov eax, 1
xor eax, 1
ret
.L6:
xor eax, eax
cmp DWORD PTR [rdi+8], 1936941420
jne .L2
xor eax, 1
ret
Notice, how both gcc, clang and ldc2, compare first 8 bytes of input, then 4
bytes of input. clang and ldc2 just xor/or the result, then return, with no
conditional jumps. gcc does a bit poorer, with more conditionals and more
jumps, but still pretty good and same idea.
gdc however, calls the generic memcmp, that does looping and does about 12
jumps and/or 13 exists.
[Bug d/100769] [D] memcmp() == 0 for small constant strings not folded
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100769
--- Comment #1 from Witold Baryluk ---
A typo in the example (godbolt is good), I forgot the `.ptr`:
extern(C) int memcmp(const void *s1, const void *s2, size_t n);
int recognize3(const char* s) {
return memcmp(s, "stract class".ptr, 12) == 0;
}
casting to ubyte*, or void*, doesn't change anything really.
options: -O3 -frelease -fno-semantic-interposition
tested on amd64, Debian / Linux.
[Bug d/100769] [D] memcmp() == 0 for small constant strings not folded
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100769 --- Comment #2 from Witold Baryluk --- Hmm. It appears that using `import core.stdc.string : memcmp;` actually resolves the problem. It looks like my manually declaration of memcmp for some reason disabled optimisations for memcmp.
[Bug d/100769] [D] memcmp() == 0 for small constant strings not folded
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100769 Witold Baryluk changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |FIXED --- Comment #4 from Witold Baryluk --- Ok. That makes sense. Thanks.
[Bug d/100769] [D] memcmp() == 0 for small constant strings not folded
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100769 Witold Baryluk changed: What|Removed |Added Resolution|FIXED |INVALID
[Bug d/105360] New: Inlined lazy parameters / delegate literals, still emitted
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105360
Bug ID: 105360
Summary: Inlined lazy parameters / delegate literals, still
emitted
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
```
extern bool g();
extern void f(int n);
void log(lazy int num) {
if (g()) {
const n = num();
f(n);
}
}
void p(int n) {
log(n * 137);
}
```
This should emit the same (or close to the same) as code with no `lazy` (and
num reference changed accordingly) on `log` function. (Because compiler knows
that `num ` is called once, has no side effects, is moderately expensive, etc).
And the code for p is exactly the same - log and `n * 137` fully inlined.
However, the anonymous dgliteral code is still emitted, despite not being
referenced anywhere:
```
pure nothrow @nogc @safe int example.p(int).__dgliteral2(): # < This should
not be in object file
imuleax, DWORD PTR [rdi], 137
ret
```
Rest of the object file is correct and optimal:
```
void example.log(lazy int):
pushrbp
pushrbx
mov rbp, rdi
mov rbx, rsi
sub rsp, 8
callbool example.g()
testal, al
je .L3
mov rdi, rbp
callrbx
add rsp, 8
pop rbx
pop rbp
mov edi, eax
jmp void example.f(int)
.L3:
add rsp, 8
pop rbx
pop rbp
ret
void example.p(int):
pushrbx
mov ebx, edi
callbool example.g()
testal, al
je .L6
imuledi, ebx, 137
pop rbx
jmp void example.f(int)
.L6:
pop rbx
ret
```
gdc
(Compiler-Explorer-Build-gcc-748d46cd049c89a799f99f14547267ebae915af6-binutils-2.36.1)
12.0.1 20220421 (experimental) via godbolt.org
For a code passing reasonably big literals, this can lead to object file code
duplication.
ldc2 shows no such problem.
[Bug d/105360] Inlined lazy parameters / delegate literals, still emitted
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105360 --- Comment #1 from Witold Baryluk --- https://godbolt.org/z/c8oT6E4cf
[Bug d/105413] New: gdc extended assembler cannot constraints r8 - r15
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105413
Bug ID: 105413
Summary: gdc extended assembler cannot constraints r8 - r15
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
gcc in C does not support directly register constraints for x86_64 registers r8
- r15.
In C this can be done however using local register variables and asm
attributes.
https://gcc.gnu.org/onlinedocs/gcc/Local-Register-Variables.html
There is no way to use this in GDC extended assembler.
version (linux) {
version (GNU) {
enum SYSCALL {
OPENAT = 56,
}
@nogc:
nothrow:
size_t syscall(SYSCALL ident)(size_t arg1, size_t arg2, size_t arg3, size_t
arg4) {
version (X86_64) {
asm @nogc nothrow {
"syscall"
// output:
: "=a" (arg1)
// inputs:
: "a" (ident), // rax - syscall number
"D" (arg1), // rdi - arg1
"S" (arg2), // rsi - arg2
"d" (arg3), // rdx - arg3
"r10" (arg4), // r10 - arg4
"m"( *cast(ubyte*)arg1) // "dummy" input instead of full memory
clobber
// clobers
: "c", "r11"; // Clobers rax, and rcx and r11.
}
return arg1;
} else {
static assert(false, "This platform/architecture is not supported when
using GDC compiler");
}
}
}
private int openatdummy() @nogc nothrow {
return cast(int)syscall!(SYSCALL.OPENAT)(0, 0, 0, 0);
}
}
myio.d: In function ‘syscall’:
myio.d:232:10: error: matching constraint references invalid operand number
232 | ;
https://godbolt.org/z/xGzxa6orc
[Bug d/105413] gdc extended assembler cannot constraints r8 - r15
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105413
--- Comment #3 from Witold Baryluk ---
It works. Thank you.
Any chance this will be in gcc 12.x? I work a lot on Debian Linux, and I doubt
I will have gcc trunk or gcc 13 available any time soon.
Also weirdly gcc does not inline this function, unless I add
@attribute("always_inline") on syscall, or @attribute("flatten") on
openatdummy.
[Bug d/107241] New: std.bitmanip.bigEndianToNative et al not inlined
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107241
Bug ID: 107241
Summary: std.bitmanip.bigEndianToNative et al not inlined
Product: gcc
Version: 12.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
gdc fails to inline number of small functions that should fully inline and end
in single instruction.
on amd64 / x86, for example std.bitmanip.bigEndianToNative causes a chain of
calls / jumps, even with @attribute("flatten")
import std.bitmanip;
import gcc.attributes;
@attribute("flatten")
size_t f(char[] b) {
return std.bitmanip.bigEndianToNative!(size_t,
8)(cast(ubyte[8])(b[2..10]));
}
gcc -O3 -march=znver2 -frelease
pure nothrow @nogc @safe ulong
std.bitmanip.swapEndian!(ulong).swapEndian(const(ulong)):
mov rax, rdi
bswap rax
ret
pure nothrow @nogc @safe ulong std.bitmanip.endianToNativeImpl!(true, ulong,
8uL).endianToNativeImpl(ubyte[8]):
jmp pure nothrow @nogc @safe ulong
std.bitmanip.swapEndian!(ulong).swapEndian(const(ulong))
pure nothrow @nogc @safe ulong std.bitmanip.bigEndianToNative!(ulong,
8uL).bigEndianToNative(ubyte[8]):
jmp pure nothrow @nogc @safe ulong
std.bitmanip.endianToNativeImpl!(true, ulong, 8uL).endianToNativeImpl(ubyte[8])
ulong example.f(char[]):
mov rdi, QWORD PTR [rsi+2]
jmp pure nothrow @nogc @safe ulong
std.bitmanip.bigEndianToNative!(ulong, 8uL).bigEndianToNative(ubyte[8])
No issues with LDC.
ulong example.f(char[]):
mov rax, qword ptr [rsi + 2]
bswap rax
ret
godbolt: https://godbolt.org/z/Pj3f7oGso
[Bug c++/103966] New: std::atomic relaxed load, inc, store sub-optimal codegen
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103966
Bug ID: 103966
Summary: std::atomic relaxed load, inc, store sub-optimal
codegen
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
Both functions below, should compile to the same assembly on x86:
#include
#include
uint64_t x;
void inc_a() {
x++;
}
std::atomic y;
void inc_b_non_atomic() {
y.store(y.load(std::memory_order_relaxed) + 1, std::memory_order_relaxed);
}
and it does so in clang.
It does not in gcc 12 (and earlier).
https://godbolt.org/z/GcM67xz8T
This pattern is very popular in approximate statistical counters / metrics,
where the flow of information is unidirectional (i.e. from one thread that does
updates, to another thread that only reads the counters), and its performance
is critical in many codebases.
[Bug c++/103966] std::atomic relaxed load, inc, store sub-optimal codegen
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103966 --- Comment #1 from Witold Baryluk --- Current codegen on gcc 12 on 64-bit x86: inc_a(): inc QWORD PTR x[rip] ret inc_b_non_atomic(): mov rax, QWORD PTR y[rip] inc rax mov QWORD PTR y[rip], rax ret y: .zero 8 x: .zero 8
[Bug c++/103966] std::atomic relaxed load, inc, store sub-optimal codegen
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103966
--- Comment #2 from Witold Baryluk ---
Similarly, dec, add, sub, are affected, as well mul.
Example:
#include
#include
uint64_t x;
void add_a() {
x += 5;
}
std::atomic y;
void add_b_non_atomic() {
y.store(y.load(std::memory_order_relaxed) + 5, std::memory_order_relaxed);
}
Producing:
add_a():
add QWORD PTR x[rip], 5
ret
add_b_non_atomic():
mov rax, QWORD PTR y[rip]
add rax, 5
mov QWORD PTR y[rip], rax
ret
y:
.zero 8
x:
.zero 8
[Bug middle-end/35560] Missing CSE/PRE for memory operations involved in virtual call.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=35560
Witold Baryluk changed:
What|Removed |Added
CC||witold.baryluk+gcc at gmail
dot co
||m
--- Comment #15 from Witold Baryluk ---
I know this is a pretty old bug, but I was exploring some assembly of gcc and
clang on godbolt, and also stumbled into same issue.
https://godbolt.org/z/qPzMhWse1
class A {
public:
virtual int f7(int x) const;
};
int g(const A * const a, int x) {
int r = 0;
for (int i = 0; i < 1; i++)
r += a->f7(x);
return r;
}
(same happens without loop, when just calling a->f7 multiple times)
g(A const*, int):
pushr13
mov r13d, esi
pushr12
xor r12d, r12d
pushrbp
mov rbp, rdi
pushrbx
mov ebx, 1
sub rsp, 8
.L2:
mov rax, QWORD PTR [rbp+0] # a vtable deref
mov esi, r13d
mov rdi, rbp
call[QWORD PTR [rax]]# f7 indirect call
add r12d, eax
dec ebx
jne .L2
add rsp, 8
pop rbx
pop rbp
mov eax, r12d
pop r12
pop r13
ret
I was expecting mov rax, QWORD PTR [rbp+0] and call[QWORD PTR [rax]],
to be hoisted out of the loop (call converted to lea, and call register).
A bit sad.
Is there some recent work done on this optimization?
Are there at least some cases where it is valid to do CSE, or change code so it
is moved out of the loop?
[Bug c/108255] New: Repeated address-of (lea) not optimized for size.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108255
Bug ID: 108255
Summary: Repeated address-of (lea) not optimized for size.
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
https://godbolt.org/z/q5sx9e49j
void f(int *);
int g(int of) {
int x = 13;
f(&x);
f(&x);
f(&x);
f(&x);
f(&x);
f(&x);
f(&x);
f(&x);
return 0;
}
Got:
g(int):
sub rsp, 24
lea rdi, [rsp+12]
mov DWORD PTR [rsp+12], 13
callf(int*)
lea rdi, [rsp+12] # compute, 5 bytes
callf(int*)
lea rdi, [rsp+12] # recompute, 5 bytes
callf(int*)
lea rdi, [rsp+12] # recompute, 5 bytes
callf(int*)
lea rdi, [rsp+12]
callf(int*)
lea rdi, [rsp+12]
callf(int*)
lea rdi, [rsp+12]
callf(int*)
lea rdi, [rsp+12]
callf(int*)
xor eax, eax
add rsp, 24
ret
But, note that lea is 5 bytes.
Expected (generated by clang 3.0 - 15.0):
g(int): # @g(int)
pushrbx # extra, but just 1 byte
sub rsp, 16
mov dword ptr [rsp + 12], 13 # CSE temp
lea rbx, [rsp + 12]
mov rdi, rbx # use
callf(int*)@PLT
mov rdi, rbx # reuse, 3 bytes
callf(int*)@PLT
mov rdi, rbx # reuse, 3 bytes
callf(int*)@PLT
mov rdi, rbx
callf(int*)@PLT
mov rdi, rbx
callf(int*)@PLT
mov rdi, rbx
callf(int*)@PLT
mov rdi, rbx
callf(int*)@PLT
mov rdi, rbx
callf(int*)@PLT
xor eax, eax
add rsp, 16
pop rbx # extra, but just 1 byte
ret
Technically this is more instructions.
But
mov rdi, rbx is 3 bytes, which is shorter than 5 bytes of lea. This is at minor
expense of needing to save and restore rbx.
PS. Same happens when using temporary `int *const y = &x;`
Also same when optimizing for size (`-Os`).
It looks like gcc 4.8.5 produced expected code, but gcc 4.9.0 does not.
It is possible that the code produced by gcc 4.9.0 is faster, but it is also
likely it contributes quite a bit to binary size.
clang uses CSE even if there are even just two uses of `&x` in the above
example. It is likely a bit higher threshold is (3 or 4) is actually optimal
(can be calculated knowing encoding sizes).
Weirdly tho, gcc -m32 does this:
g():
pushebp
mov ebp, esp
pushebx
lea ebx, [ebp-12]
sub esp, 32
mov DWORD PTR [ebp-12], 13
pushebx
callf(int*)
mov DWORD PTR [esp], ebx
callf(int*)
mov DWORD PTR [esp], ebx
callf(int*)
mov ebx, DWORD PTR [ebp-4]
xor eax, eax
leave
ret
Where, it does compute address and stores it in temporary. But does it on a
stack, instead in a register (my guess is there are no free register to store
it and it is spilled)., but in fact lea here would be likely faster (mov
DWORD PTR [esp], ebx, but requires memory/cache access, lea is 5 bytes, but
does not require memory access)
[Bug d/109221] New: std.math.floor, core.math.ldexp, std.math.poly poor inlining
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109221
Bug ID: 109221
Summary: std.math.floor, core.math.ldexp, std.math.poly poor
inlining
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
Example:
static float sRGB_case4(float x) {
// import std.math : exp;
return 1.055f * expImpl(x) - 0.055f; // expImpl not inlined by default
// (inlined when using pragma(inline, true), but that fails to inline in
DMD)
}
// pragma(inline, true)
// This is borrowed from phobos/exponential.d to help gcc inline it fully.
// Only T == float case is here (as some traits are private to phobos).
// Also isNaN and range checks are removed, as sRGB performs own checks.
static private T expImpl(T)(T x) @safe pure nothrow @nogc
{
//import std.math : floatTraits, RealFormat;
//import std.math.traits : isNaN;
//import std.math.rounding : floor;
//import std.math.algebraic : poly;
//import std.math.constants : LOG2E;
import std.math;
import core.math;
static immutable T[6] P = [
5.001201E-1,
1.665459E-1,
4.1665795894E-2,
8.3334519073E-3,
1.3981999507E-3,
1.9875691500E-4,
];
enum T C1 = 0.693359375;
enum T C2 = -2.12194440e-4;
// Overflow and Underflow limits.
enum T OF = 88.72283905206835;
enum T UF = -103.278929903431851103; // ln(2^-149)
// Special cases.
//if (isNaN(x))
//return x;
//if (x > OF)
//return real.infinity;
//if (x < UF)
//return 0.0;
// Express: e^^x = e^^g * 2^^n
// = e^^g * e^^(n * LOG2E)
// = e^^(g + n * LOG2E)
T xx = floor((cast(T) LOG2E) * x + cast(T) 0.5); // NOT INLINED!
const int n = cast(int) xx;
x -= xx * C1;
x -= xx * C2;
xx = x * x;
x = poly(x, P) * xx + x + 1.0f; // poly is generated optimally, but
not inlined
// Scale by power of 2.
x = core.math.ldexp(x, n);// NOT INLINED
return x;
}
gdc gdc
(Compiler-Explorer-Build-gcc-454a4d5041f53cd1f7d902f6c0017b7ce95b36df-binutils-2.38)
13.0.1 20230318 (experimental)
gdc -O3 -march=znver2 -frelease -fbounds-check=off
pure nothrow @nogc @safe float std.math.algebraic.poly!(float, float,
6).poly(float, ref const(float[6])):
vmovss xmm1, DWORD PTR [rdi+20]
vfmadd213ss xmm1, xmm0, DWORD PTR [rdi+16]
vfmadd213ss xmm1, xmm0, DWORD PTR [rdi+12]
vfmadd213ss xmm1, xmm0, DWORD PTR [rdi+8]
vfmadd213ss xmm1, xmm0, DWORD PTR [rdi+4]
vfmadd213ss xmm0, xmm1, DWORD PTR [rdi]
ret
pure nothrow @nogc @safe float example.expImpl!(float).expImpl(float):
pushrbx
vmovaps xmm1, xmm0
sub rsp, 16
vmovss xmm0, DWORD PTR .LC0[rip]
vfmadd213ss xmm0, xmm1, DWORD PTR .LC1[rip]
vmovss DWORD PTR [rsp+8], xmm1
callpure nothrow @nogc @trusted float
std.math.rounding.floor(float)
vmovss xmm1, DWORD PTR [rsp+8]
mov edi, OFFSET FLAT:immutable(float[6])
example.expImpl!(float).expImpl(float).P
vfnmadd231ssxmm1, xmm0, DWORD PTR .LC2[rip]
vmovss DWORD PTR [rsp+12], xmm0
vfnmadd231ssxmm1, xmm0, DWORD PTR .LC3[rip]
vmulss xmm3, xmm1, xmm1
vmovaps xmm0, xmm1
vmovss DWORD PTR [rsp+8], xmm1
vmovd ebx, xmm3
callpure nothrow @nogc @safe float std.math.algebraic.poly!(float,
float, 6).poly(float, ref const(float[6]))
vmovss xmm1, DWORD PTR [rsp+8]
vmovd xmm4, ebx
vmovss xmm2, DWORD PTR [rsp+12]
vfmadd132ss xmm0, xmm1, xmm4
vaddss xmm0, xmm0, DWORD PTR .LC4[rip]
add rsp, 16
pop rbx
vcvttss2si edi, xmm2
jmp ldexpf
float example.sRGB_case4(float):
sub rsp, 8
callpure nothrow @nogc @safe float
example.expImpl!(float).expImpl(float)
vmovss xmm1, DWORD PTR .LC6[rip]
vfmadd132ss xmm0, xmm1, DWORD PTR .LC5[rip]
add rsp, 8
ret
https://godbolt.org/z/YMoMPdjn5
Additionally
std.math.exp itself, is never inlined by gcc. This is important, as some early
checks (isNaN, OF, UF checks) in exp could be removed by proper inlining.
[Bug d/109221] std.math.floor, core.math.ldexp, std.math.poly poor inlining
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109221 --- Comment #1 from Witold Baryluk --- PS. LDC 1.23.0 - 1.32.0 produce optimal code. LDC 1.22.0 a bit worse (due to use of x87 codegen), and 1.21 and older fail to inline `ldexp`, but still inline `poly` and `floor` perfectly.
[Bug d/109221] std.math.floor, core.math.ldexp, std.math.poly poor inlining
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109221 --- Comment #2 from Witold Baryluk --- Interesting enough, GDC 10.2 does inline `poly` instantiation with all the constants.
[Bug d/110113] New: gdc -fpreview=dip1021 crash in d/dmd/root/aav.d:127 dmd_aaGetRvalue from DsymbolTable::lookup(Identifier const*)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110113
Bug ID: 110113
Summary: gdc -fpreview=dip1021 crash in d/dmd/root/aav.d:127
dmd_aaGetRvalue from DsymbolTable::lookup(Identifier
const*)
Product: gcc
Version: 13.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
Created attachment 55254
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55254&action=edit
Minimized test case with dustmite
Debian Linux amd64, experimental gcc-13, gdc 13.1.0-3
This is not very deterministic. Run few times to trigger.
```
user@debian:~$ cat lup.d
class LUBench {
}
float lup(ulong , ulong , int , int = 1) {
double[] solution;
new LUBench;
return solution[0] ;
}
float lup_3200(ulong iters, ulong flops) {
return lup(iters, flops, 3200);
}
float raytrace() {
struct V {
float x, y, z;
auto normalize() {
}
import std;
auto cross() {
}
auto norm2() {
}
auto norm() {
}
auto opBinary(){
}
}
}
user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d
lup.d:11:7: error: function ‘lup.raytrace’ has no ‘return’ statement, but is
expected to return a value of type ‘float’
11 | float raytrace() {
| ^
user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d
lup.d:11:7: error: function ‘lup.raytrace’ has no ‘return’ statement, but is
expected to return a value of type ‘float’
11 | float raytrace() {
| ^
user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d
lup.d:11:7: error: function ‘lup.raytrace’ has no ‘return’ statement, but is
expected to return a value of type ‘float’
11 | float raytrace() {
| ^
user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d
lup.d:11:7: error: function ‘lup.raytrace’ has no ‘return’ statement, but is
expected to return a value of type ‘float’
11 | float raytrace() {
| ^
user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d
/usr/lib/gcc/x86_64-linux-gnu/13/include/d/std/math/algebraic.d:968:47:
internal compiler error: Segmentation fault
968 | return cast(Unqual!T) (T(1) << bsr(val) + type);
| ^
0xd32f86 crash_signal
../../src/gcc/toplev.cc:314
0x7f53b651cf8f ???
./signal/../sysdeps/unix/sysv/linux/x86_64/libc_sigaction.c:0
0x17f7d10 _D3dmd4root3aav15dmd_aaGetRvalueFNaNbNiPSQBnQBmQBk2AAPvZQd
../../src/gcc/d/dmd/root/aav.d:127
0x1706b25 DsymbolTable::lookup(Identifier const*)
../../src/gcc/d/dmd/dsymbol.d:2408
0x1706b25 ScopeDsymbol::search(Loc const&, Identifier*, int)
../../src/gcc/d/dmd/dsymbol.d:1470
0x17ef5b3
_D3dmd6opover15search_functionFCQBe7dsymbol12ScopeDsymbolCQCe10identifier10IdentifierZCQDhQCd7Dsymbol
../../src/gcc/d/dmd/opover.d:1435
0x1701fe0 search_toString(StructDeclaration*)
../../src/gcc/d/dmd/dstruct.d:51
0x180310a semanticTypeInfoMembers(StructDeclaration*)
../../src/gcc/d/dmd/semantic3.d:1650
0x1803394 Semantic3Visitor::visit(AggregateDeclaration*)
../../src/gcc/d/dmd/semantic3.d:1590
0x17fef19 semantic3(Dsymbol*, Scope*)
../../src/gcc/d/dmd/semantic3.d:83
0x175dc89 ExpressionSemanticVisitor::visit(DeclarationExp*)
../../src/gcc/d/dmd/expressionsem.d:5572
0x175dc89 ExpressionSemanticVisitor::visit(DeclarationExp*)
../../src/gcc/d/dmd/expressionsem.d:5407
0x175eb82 expressionSemantic(Expression*, Scope*)
../../src/gcc/d/dmd/expressionsem.d:12706
0x18096fa StatementSemanticVisitor::visit(ExpStatement*)
../../src/gcc/d/dmd/statementsem.d:207
0x18228c1 statementSemantic(Statement*, Scope*)
../../src/gcc/d/dmd/statementsem.d:149
0x18228c1 StatementSemanticVisitor::visit(CompoundStatement*)
../../src/gcc/d/dmd/statementsem.d:270
0x1809112 statementSemantic(Statement*, Scope*)
../../src/gcc/d/dmd/statementsem.d:149
0x18002a1 Semantic3Visitor::visit(FuncDeclaration*)
../../src/gcc/d/dmd/semantic3.d:598
0x17feae4 semantic3(Dsymbol*, Scope*)
../../src/gcc/d/dmd/semantic3.d:83
0x17feae4 Semantic3Visitor::visit(Module*)
../../src/gcc/d/dmd/semantic3.d:205
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See for instructions.
user@debian:~$
```
Could not reduce further, as it is sensitive to identifiers, and due to
non-deterministic nature testing requires many repetitions.
[Bug d/110113] gdc -fpreview=dip1021 crash in d/dmd/root/aav.d:127 dmd_aaGetRvalue from DsymbolTable::lookup(Identifier const*)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110113 --- Comment #1 from Witold Baryluk --- BTW. Adding return statement in `raytrace`, does not change anything: ``` user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d user@debian:~$ gdc-13 -c -fpreview=dip1021 lup.d /usr/lib/gcc/x86_64-linux-gnu/13/include/d/std/math/algebraic.d:968:47: internal compiler error: Segmentation fault 968 | return cast(Unqual!T) (T(1) << bsr(val) + type); | ^ 0xd32f86 crash_signal ../../src/gcc/toplev.cc:314 0x7f7144273f8f ??? ./signal/../sysdeps/unix/sysv/linux/x86_64/libc_sigaction.c:0 0x17f7d10 _D3dmd4root3aav15dmd_aaGetRvalueFNaNbNiPSQBnQBmQBk2AAPvZQd ../../src/gcc/d/dmd/root/aav.d:127 0x1706b25 DsymbolTable::lookup(Identifier const*) ../../src/gcc/d/dmd/dsymbol.d:2408 0x1706b25 ScopeDsymbol::search(Loc const&, Identifier*, int) ../../src/gcc/d/dmd/dsymbol.d:1470 ... ... ```
[Bug d/110113] gdc -fpreview=dip1021 crash in d/dmd/root/aav.d:127 dmd_aaGetRvalue from DsymbolTable::lookup(Identifier const*)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110113 --- Comment #2 from Witold Baryluk --- Also FYI, I was not able to trigger this on DMD64 D Compiler v2.104.0
[Bug d/110113] gdc -fpreview=dip1021 crash in d/dmd/root/aav.d:127 dmd_aaGetRvalue from DsymbolTable::lookup(Identifier const*)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110113 --- Comment #10 from Witold Baryluk --- Thank you Iain. Amazing debugging skills. BTW. `import std;` was because dustmite reduced original import to just that. Original import was `import std.math.algebraic : sqrt;` But you already figured this out without even using Phobos.
[Bug d/110516] New: core.volatile.volatileLoad is broken
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110516
Bug ID: 110516
Summary: core.volatile.volatileLoad is broken
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: d
Assignee: ibuclaw at gdcproject dot org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
gcc 12.2.0 (from Debian stable) and gcc trunk 14.0.0 (in godbolt) tested.
core.volatile.volatileLoad simply does not work.
1) It merges loads.
2) It removes unused loads at -O1 and higher.
Example:
void actualRun(ubyte* ptr1) {
import core.volatile : volatileLoad;
volatileLoad(ptr1);
volatileLoad(ptr1);
volatileLoad(ptr1);
volatileLoad(ptr1);
}
Without optimisations:
void example.actualRun(ubyte*):
pushrbp
mov rbp, rsp
mov QWORD PTR [rbp-8], rdi
nop
pop rbp
ret
Incorrect.
With optimisations:
void example.actualRun(ubyte*):
ret
Incorrect.
Expected:
void example.actualRun(ubyte*):
movzx eax, byte ptr [rdi]
movzx eax, byte ptr [rdi]
movzx eax, byte ptr [rdi]
movzx eax, byte ptr [rdi]
ret
dmd and ldc behave properly.
It looks like it never worked properly.
Would be good to have a test case for this, so it does not become a regression
later.
I did not test volatileStore, but I would not be surprised it is also broken.
[Bug d/110516] core.volatile.volatileLoad discarded if result is unused
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110516 --- Comment #8 from Witold Baryluk --- I see. Point 1 is definitively incorrect. I interpreted asembler wrong: void example.actualRun(ubyte*): pushrbp mov rbp, rsp mov QWORD PTR [rbp-8], rdi nop pop rbp ret The move there, is just some stack manipulation, it has nothing to do with volatileLoad. You are right about the side effect visibility and volatileStore. Still, there should be a way to express real memory read, with result not stored anywhere in program (just written to register, then discarded). This has some (not very common) uses in memmory-mapped IO, i.e. in drivers for devices where the read itself could indicate something (this of course usually also require setting proper page table attributes to disable caching or other optimizations, etc, not just volatile load in machine code). I do not have specific examples at hand, but afaik I saw some examples in the past (mostly on older architectures), as well some watchdog chips that reset timer on read. Another use is for doing memory and cache read benchmarks and profiling. We want to invoke read (to register) from some memory location, but we do not need the value for anything else. And more esoteric use might be memory probing. On some level systems, kernel or bootloader, might not know the memory layout, and resort to just doing reads, and relaying on CPU fault handlers to report invalid reads. And some people might use load without destination, as a prefetch hint, or to prefault some memory pages.
[Bug d/110516] core.volatile.volatileLoad discarded if result is unused
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110516 --- Comment #9 from Witold Baryluk --- Thank you for a quick fix Iain!
[Bug d/113125] New: [D] internal compiler error: in make_import, at d/imports.cc:48
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113125 Bug ID: 113125 Summary: [D] internal compiler error: in make_import, at d/imports.cc:48 Product: gcc Version: 13.2.1 Status: UNCONFIRMED Severity: normal Priority: P3 Component: d Assignee: ibuclaw at gdcproject dot org Reporter: witold.baryluk+gcc at gmail dot com Target Milestone: --- Debian testing, amd64, gcc version 13.2.0 (Debian 13.2.0-7) meta.d: ``` module objc.meta; struct A; ``` runtime.d: ``` module objc.runtime; public import meta : A; ``` gdc -v -c -I. runtime.d ``` $ gdc -v -c -I. runtime.d Using built-in specs. COLLECT_GCC=gdc OFFLOAD_TARGET_NAMES=nvptx-none:amdgcn-amdhsa OFFLOAD_TARGET_DEFAULT=1 Target: x86_64-linux-gnu Configured with: ../src/configure -v --with-pkgversion='Debian 13.2.0-7' --with-bugurl=file:///usr/share/doc/gcc-13/README.Bugs --enable-languages=c,ada,c++,go,d,fortran,objc,obj-c++,m2 --prefix=/usr --with-gcc-major-version-only --program-suffix=-13 --program-prefix=x86_64-linux-gnu- --enable-shared --enable-linker-build-id --libexecdir=/usr/libexec --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --enable-bootstrap --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --disable-vtable-verify --enable-plugin --enable-default-pie --with-system-zlib --enable-libphobos-checking=release --with-target-system-zlib=auto --enable-objc-gc=auto --enable-multiarch --disable-werror --enable-cet --with-arch-32=i686 --with-abi=m64 --with-multilib-list=m32,m64,mx32 --enable-multilib --with-tune=generic --enable-offload-targets=nvptx-none=/build/reproducible-path/gcc-13-13.2.0/debian/tmp-nvptx/usr,amdgcn-amdhsa=/build/reproducible-path/gcc-13-13.2.0/debian/tmp-gcn/usr --enable-offload-defaulted --without-cuda-driver --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu --with-build-config=bootstrap-lto-lean --enable-link-serialization=3 Thread model: posix Supported LTO compression algorithms: zlib zstd gcc version 13.2.0 (Debian 13.2.0-7) COLLECT_GCC_OPTIONS='-v' '-c' '-I' '.' '-o' 'runtime.o' '-shared-libgcc' '-mtune=generic' '-march=x86-64' /usr/libexec/gcc/x86_64-linux-gnu/13/d21 runtime.d -quiet -dumpbase runtime.d -dumpbase-ext .d -mtune=generic -march=x86-64 -version -imultiarch x86_64-linux-gnu -I . -v -o /tmp/ccPyiN0m.s GNU D (Debian 13.2.0-7) version 13.2.0 (x86_64-linux-gnu) compiled by GNU C version 13.2.0, GMP version 6.3.0, MPFR version 4.2.1, MPC version 1.3.1, isl version isl-0.26-GMP GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072 binary/usr/libexec/gcc/x86_64-linux-gnu/13/d21 version v2.103.1 predefs GNU D_Version2 LittleEndian GNU_DWARF2_Exceptions GNU_StackGrowsDown GNU_InlineAsm D_LP64 D_PIC D_PIE assert D_PreConditions D_PostConditions D_Invariants D_ModuleInfo D_Exceptions D_TypeInfo all X86_64 D_HardFloat Posix linux CRuntime_Glibc CppRuntime_Gcc parse runtime importall runtime importmeta (meta.d) importobject(/usr/lib/gcc/x86_64-linux-gnu/13/include/d/object.d) importcore.attribute (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/attribute.d) importgcc.attributes (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/gcc/attributes.d) importcore.internal.hash (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/hash.d) importcore.internal.traits (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/traits.d) importcore.internal.entrypoint (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/entrypoint.d) importcore.internal.array.appending (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/appending.d) importcore.internal.array.comparison (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/comparison.d) importcore.internal.array.equality (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/equality.d) importcore.internal.array.casting (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/casting.d) importcore.internal.array.concatenation (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/concatenation.d) importcore.internal.array.construction (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/construction.d) importcore.internal.array.arrayassign (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/arrayassign.d) importcore.internal.array.capacity (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/array/capacity.d) importcore.internal.dassert (/usr/lib/gcc/x86_64-linux-gnu/13/include/d/core/internal/dassert.d) importcore.atomic (/usr/lib/gcc/x86_64-linu
