Fixed on 4.3.2-1 Debian gnueabi and on trunk at rev 142808 (tested native)

Testing sample with gcc-4.3.3 (released after 2009-01-02 12:06) also fail:

sergeyhost:/tmp$ cat ./main.c 
#include <mmintrin.h>
void foo()
{
    _mm_srli_pi16(_mm_setzero_si64(), 8);
}
sergeyhost:/tmp$

sergeyhost:/tmp$ arm-iwmmx-linux-gnueabi-gcc -flax-vector-conversions -c -O2 ./main.c
/opt/toolchain/arm-iwmmx-linux-gnueabi/gcc-4.3.3-glibc-2.8-binutils-2.18-kernel-2.6.22-sanitized/lib/gcc/arm-iwmmx-linux-gnueabi/4.3.3/include/mmintrin.h: In function 'foo':
/opt/toolchain/arm-iwmmx-linux-gnueabi/gcc-4.3.3-glibc-2.8-binutils-2.18-kernel-2.6.22-sanitized/lib/gcc/arm-iwmmx-linux-gnueabi/4.3.3/include/mmintrin.h:529: internal compiler error: in arm_expand_binop_builtin, at config/arm/arm.c:15380
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
sergeyhost:/tmp$ 

I am use OSELAS.Toolchan-1.99.3 arm-iwmmx-linux-gnueabi_gcc-4.3.2_glibc-2.8_binutils-2.18_kernel-2.6.27-sanitized with small changes for me (update compiler because in contains "regression fixes and docs only" and downgrade kernel headers). I think this changes unimportant for bug.

Can anybody reopen bug?

same problem with gcc-4.3.4, sources comes from gentoo package

iris ~ $ arm-iwmmxt-linux-gnueabi-g++ -flax-vector-conversions -c -O2  1.cpp
/vol/marvel-4.3.2/bin/../lib/gcc/arm-iwmmxt-linux-gnueabi/4.3.4/include/mmintrin.h: In function 'void foo()':
/vol/marvel-4.3.2/bin/../lib/gcc/arm-iwmmxt-linux-gnueabi/4.3.4/include/mmintrin.h:529: internal compiler error: in arm_expand_binop_builtin, at config/arm/arm.c:15429
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
iris ~ $ arm-iwmmxt-linux-gnueabi-g++ -v
Using built-in specs.
Target: arm-iwmmxt-linux-gnueabi
Configured with: /vol/build/gcc-4.3.4/configure --target=arm-iwmmxt-linux-gnueabi --with-cpu=iwmmxt --with-float=soft --enable-cxx-flags=-msoft-float --disable-nls --enable-threads=posix --enable-symvers=gnu --enable-__cxa_atexit --enable-languages=c,c++ --enable-shared --enable-c99 --enable-long-long --with-sysroot=/opt/marvel/sysroot --prefix=/opt/marvel/
Thread model: posix
gcc version 4.3.4 (GCC) 

Denis, nobody wants to open this thread. First, we must do it.

PS You say in Russian?

for russian - yes.

Should we open new issue? :)

ß êîãäà-òî ïî ïîâîäó ýòîé îøèáêè îáèëüíî ãóãëèë è íàõîäèë íåñêîëüêî òðåäîâ â ýòîé áàãçèëëå. Èõ çàêðûâàëè ïîä ïðåäëîãîì "duplicate". ß äóìàþ, åñëè òû çàõî÷åøü, òû èõ òîæå íàéä¸øü. Òàê ÷òî - ýòî ãëàâíûé òðåä.
À òàê íàâåðíîå íàäî îòêðûâàòü íîâûé áàã è ñêîïèïàñòèòü òóäà ÷òî åñòü. :)

The assertion in the source is clearly broken if a constant can legally be passed to it.  If a constant can't legally be passed, then the caller needs to be fixed to diagnose it before we reach this point.

Appears with gcc 4.4.5 (builded with crosstool-ng 1.9.0)

$ /opt/arm-iwmmx-linux-uclibcgnueabi/bin/arm-iwmmx-linux-uclibcgnueabi-gcc iwmmx_test.c -flax-vector-conversions -O1
In file included from iwmmx_test.c:1:
/opt/arm-iwmmx-linux-uclibcgnueabi/lib/gcc/arm-iwmmx-linux-uclibcgnueabi/4.4.5/include/mmintrin.h: In function 'main':
/opt/arm-iwmmx-linux-uclibcgnueabi/lib/gcc/arm-iwmmx-linux-uclibcgnueabi/4.4.5/include/mmintrin.h:526: internal compiler error: in arm_expand_binop_builtin, at config/arm/arm.c:16168
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

iwmmx_test.c:

#include <mmintrin.h>
int main()
{
    _mm_srli_pi16(_mm_setzero_si64(), 8);
	return 0;
}

Compiler fail only when there is optimization(-O1).

It seems that this bug affects the whole iwmmxt shift intrinsics group(_mm_slli_pi16, _mm_slli_pi32, _mm_slli_si64, _mm_srai_pi16, _mm_srai_pi32, _mm_srai_si64, _mm_srli_pi16, _mm_srli_pi32, _mm_srli_si64, _mm_rori_pi16, _mm_rori_pi32, _mm_rori_si64). I tested it.

These functions have the same prototype:
static __inline __m64 func_name (__m64 __m, int __count)

But it is very strange that abort did not work in gcc 3.4.5(This version looks like is not affected by this bug and it contains abort instead gcc_assert).

My suggestion:

static rtx
arm_expand_binop_builtin_iwmmx_shift (enum insn_code icode,
			  tree exp, rtx target)
{
  rtx pat;
  tree arg0 = CALL_EXPR_ARG (exp, 0);
  tree arg1 = CALL_EXPR_ARG (exp, 1);
  rtx op0 = expand_normal (arg0);
  rtx op1 = expand_normal (arg1);
  enum machine_mode tmode = insn_data[icode].operand[0].mode;
  enum machine_mode mode0 = insn_data[icode].operand[1].mode;
  enum machine_mode mode1 = insn_data[icode].operand[2].mode;

  gcc_assert (VECTOR_MODE_P (mode0))
  op0 = safe_vector_operand (op0, mode0);

  gcc_assert (!VECTOR_MODE_P (mode1))

  if (! target
      || GET_MODE (target) != tmode
      || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
    target = gen_reg_rtx (tmode);

  if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
    op0 = copy_to_mode_reg (mode0, op0);
  if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
    op1 = copy_to_mode_reg (mode1, op1);

  pat = GEN_FCN (icode) (target, op0, op1);
  if (! pat)
    return 0;
  emit_insn (pat);
  return target;
}

And in
arm_expand_builtin
...
  switch (fcode)
    {
...
	case ARM_BUILTIN_WSLLHI:
	case ARM_BUILTIN_WSLLWI:
	case ARM_BUILTIN_WSLLDI:

	case ARM_BUILTIN_WSRAHI:
	case ARM_BUILTIN_WSRAWI:
	case ARM_BUILTIN_WSRADI:

	case ARM_BUILTIN_WSRLHI:
	case ARM_BUILTIN_WSRLWI:
	case ARM_BUILTIN_WSRLDI:

	case ARM_BUILTIN_WRORHI:
	case ARM_BUILTIN_WRORWI:
	case ARM_BUILTIN_WRORDI:
		arm_expand_binop_builtin_iwmmx_shift(fcode, exp, target);
...

Created attachment 22799 [details]
gcc 35294 possible Fix

Created attachment 22800 [details]
gcc 4.4.5 possible fix

I have attached two versions of the patch (gcc 4.2.4 and gcc 4.4.5). Patched gcc 4.2.4 seems to work. With the patched gcc 4.4.5, I ran into bug 44332, and can not check. For arm platform newer versions of gcc seems more completely ubusabled.

Created attachment 25391 [details]
[PATCH] Wire-up missing ARM iwmmxt intrinsics

Fixes it for me for gcc-4.6.1. Allows me to build an iwmmxt-optimized pixman using the standard _mm_* intrinsics.

The previous attachment 25391 [details] fixes it for me for gcc-4.5.3.

It allowed me to build pixman-0.24.0 with the --enable-arm-iwmmxt configure option.

Author: nickc
Date: Tue Jun 26 14:59:45 2012
New Revision: 188988

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=188988
Log:
	* doc/extend.texi (__builtin_arm_tinsrb): Add missing second
	parameter.
	(__builtin_arm_tinsrh): Likewise.
	(__builtin_arm_tinsrw): Likewise.
	(__builtin_arm_wsadb): Add missing v2si parameter.
	(__builtin_arm_wsadh): Likewise.
	(__builtin_arm_getwcx): Delete.
	(__builtin_arm_setwcx): Delete.
	(__builtin_arm_getwcgr0): Add.
	(__builtin_arm_getwcgr1): Add.
	(__builtin_arm_getwcgr2): Add.
	(__builtin_arm_getwcgr3): Add.
	(__builtin_arm_setwcgr0): Add.
	(__builtin_arm_setwcgr1): Add.
	(__builtin_arm_setwcgr2): Add.
	(__builtin_arm_setwcgr3): Add.

	PR target/35294
	* gcc.target/arm/mmx-2.c: New.

Added:
    trunk/gcc/testsuite/gcc.target/arm/mmx-2.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/doc/extend.texi
    trunk/gcc/testsuite/ChangeLog

Author: nickc
Date: Mon Jan 28 15:07:41 2013
New Revision: 195510

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195510
Log:
	PR target/35294
	* config/arm/arm.c (arm_expand_binop_builtin): Add new parameter.
	Fix mode checks to allow for the passing of constants in the
	second parameter.
	(arm_expand_builtin): Adjust calls to arm_expand_binop_builtin.
	Add entries in the switch statement for builtin iwmmxt vector
	shift and rotate builtins.

Modified:
    branches/gcc-4_7-branch/gcc/ChangeLog
    branches/gcc-4_7-branch/gcc/config/arm/arm.c

I have checked in a modified version of serowk's patch to the 4.7 branch.  The modification was to produce an error message in the cases where the modes do not match, and to allow for the automatic forcing of constants into registers for those builtins that do not accept immediates as their third operand.

Cheers
  Nick

*** Bug 36798 has been marked as a duplicate of this bug. ***

*** Bug 36966 has been marked as a duplicate of this bug. ***

Fixed.