On Wed, Jun 15, 2011 at 4:54 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> All of these _mm{,256}_permute2_p[sd] and _mm_roti_epi{8,16,32,64}
> intrinsics ICE if the last argument is constant integer, but not in the
> expected range.
>
> I could only find MSFT documentation for these intrinsics, where for
> *permute2* it says that the last argument must be 0, 1, 2 or 3,
> for *roti* it says that the last argument is integer rotation count,
> preferrably constant and that if count is negative, it performs right
> rotation instead of left rotation.
> This patch adjusts the builtins to match that, if we want to instead
> e.g. always mandate _mm_roti_epi* last argument is constant integer,
> or constant integer in the range -N+1 .. N-1 where N is the number
> after _mm_roti_epi, or in the range 0 .. N-1, it can be easily adjusted.
>
> Regtested on x86_64-linux {-m32,-m64}, unfortunately on a SandyBridge
> box, so I couldn't verify if xop-rotate[12]-int.c actually succeeds
> on xop capable HW.
>
> 2011-06-15 Jakub Jelinek <ja...@redhat.com>
>
> PR target/49411
> * config/i386/i386.c (ix86_expand_multi_arg_builtins): If
> last_arg_constant and last argument doesn't match its predicate,
> for xop_vpermil2<mode>3 error out and for xop_rotl<mode>3
> if it is CONST_INT, mask it, otherwise expand using rotl<mode>3.
>
> * gcc.target/i386/xop-vpermil2px-1.c: New test.
> * gcc.target/i386/xop-vpermil2px-2.c: New test.
> * gcc.target/i386/xop-rotate1-int.c: New test.
> * gcc.target/i386/xop-rotate2-int.c: New test.
>
> --- gcc/config/i386/i386.c.jj 2011-06-09 16:56:56.000000000 +0200
> +++ gcc/config/i386/i386.c 2011-06-15 11:17:12.000000000 +0200
> @@ -26149,16 +26149,66 @@ ix86_expand_multi_arg_builtin (enum insn
> int adjust = (comparison_p) ? 1 : 0;
> enum machine_mode mode = insn_data[icode].operand[i+adjust+1].mode;
>
> - if (last_arg_constant && i == nargs-1)
> + if (last_arg_constant && i == nargs - 1)
> {
> - if (!CONST_INT_P (op))
> + if (!insn_data[icode].operand[i + 1].predicate (op, mode))
> {
> - error ("last argument must be an immediate");
> - return gen_reg_rtx (tmode);
> + enum insn_code new_icode = icode;
> + switch (icode)
> + {
> + case CODE_FOR_xop_vpermil2v2df3:
> + case CODE_FOR_xop_vpermil2v4sf3:
> + case CODE_FOR_xop_vpermil2v4df3:
> + case CODE_FOR_xop_vpermil2v8sf3:
> + if (!CONST_INT_P (op))
> + {
> + error ("last argument must be an immediate");
> + return gen_reg_rtx (tmode);
> + }
> + error ("last argument must be in the range 0 .. 3");
> + return gen_reg_rtx (tmode);
> + case CODE_FOR_xop_rotlv2di3:
> + new_icode = CODE_FOR_rotlv2di3;
> + goto xop_rotl;
> + case CODE_FOR_xop_rotlv4si3:
> + new_icode = CODE_FOR_rotlv4si3;
> + goto xop_rotl;
> + case CODE_FOR_xop_rotlv8hi3:
> + new_icode = CODE_FOR_rotlv8hi3;
> + goto xop_rotl;
> + case CODE_FOR_xop_rotlv16qi3:
> + new_icode = CODE_FOR_rotlv16qi3;
> + xop_rotl:
> + if (CONST_INT_P (op))
> + {
> + int mask = GET_MODE_BITSIZE (GET_MODE_INNER (tmode)) -
> 1;
> + op = GEN_INT (INTVAL (op) & mask);
> + gcc_checking_assert
> + (insn_data[icode].operand[i + 1].predicate (op,
> mode));
> + }
> + else
> + {
> + gcc_checking_assert
> + (nargs == 2
> + && insn_data[new_icode].operand[0].mode == tmode
> + && insn_data[new_icode].operand[1].mode == tmode
> + && insn_data[new_icode].operand[2].mode == mode
> + && insn_data[new_icode].operand[0].predicate
> + == insn_data[icode].operand[0].predicate
> + && insn_data[new_icode].operand[1].predicate
> + == insn_data[icode].operand[1].predicate);
> + icode = new_icode;
> + goto non_constant;
> + }
> + break;
> + default:
> + gcc_unreachable ();
> + }
> }
> }
> else
> {
> + non_constant:
> if (VECTOR_MODE_P (mode))
> op = safe_vector_operand (op, mode);
>
> --- gcc/testsuite/gcc.target/i386/xop-vpermil2px-1.c.jj 2011-06-15
> 10:18:29.000000000 +0200
> +++ gcc/testsuite/gcc.target/i386/xop-vpermil2px-1.c 2011-06-15
> 10:41:13.000000000 +0200
> @@ -0,0 +1,25 @@
> +/* PR target/49411 */
> +/* { dg-do compile } */
> +/* { dg-options "-O0 -mxop" } */
> +
> +#include <x86intrin.h>
> +
> +__m128d a1, a2, a3;
> +__m256d b1, b2, b3;
> +__m128 c1, c2, c3;
> +__m256 d1, d2, d3;
> +__m128i s;
> +__m256i t;
> +
> +void
> +foo (int i)
> +{
> + a1 = _mm_permute2_pd (a2, a3, s, 3);
> + b1 = _mm256_permute2_pd (b2, b3, t, 3);
> + c1 = _mm_permute2_ps (c2, c3, s, 3);
> + d1 = _mm256_permute2_ps (d2, d3, t, 3);
> + a1 = _mm_permute2_pd (a2, a3, s, 17); /* { dg-error "last
> argument must be in the range 0 .. 3" } */
> + b1 = _mm256_permute2_pd (b2, b3, t, 17); /* { dg-error "last argument
> must be in the range 0 .. 3" } */
> + c1 = _mm_permute2_ps (c2, c3, s, 17); /* { dg-error "last
> argument must be in the range 0 .. 3" } */
> + d1 = _mm256_permute2_ps (d2, d3, t, 17); /* { dg-error "last argument
> must be in the range 0 .. 3" } */
> +}
> --- gcc/testsuite/gcc.target/i386/xop-vpermil2px-2.c.jj 2011-06-15
> 10:39:36.000000000 +0200
> +++ gcc/testsuite/gcc.target/i386/xop-vpermil2px-2.c 2011-06-15
> 10:39:44.000000000 +0200
> @@ -0,0 +1,21 @@
> +/* PR target/49411 */
> +/* { dg-do compile } */
> +/* { dg-options "-O0 -mxop" } */
> +
> +#include <x86intrin.h>
> +
> +__m128d a1, a2, a3;
> +__m256d b1, b2, b3;
> +__m128 c1, c2, c3;
> +__m256 d1, d2, d3;
> +__m128i s;
> +__m256i t;
> +
> +void
> +foo (int i)
> +{
> + a1 = _mm_permute2_pd (a2, a3, s, i); /* { dg-error "last argument
> must be an immediate" } */
> + b1 = _mm256_permute2_pd (b2, b3, t, i); /* { dg-error "last argument
> must be an immediate" } */
> + c1 = _mm_permute2_ps (c2, c3, s, i); /* { dg-error "last argument
> must be an immediate" } */
> + d1 = _mm256_permute2_ps (d2, d3, t, i); /* { dg-error "last argument
> must be an immediate" } */
> +}
> --- gcc/testsuite/gcc.target/i386/xop-rotate1-int.c.jj 2011-06-15
> 10:47:29.000000000 +0200
> +++ gcc/testsuite/gcc.target/i386/xop-rotate1-int.c 2011-06-15
> 11:25:25.000000000 +0200
> @@ -0,0 +1,63 @@
> +/* PR target/49411 */
> +/* { dg-do run } */
> +/* { dg-require-effective-target xop } */
> +/* { dg-options "-O2 -mxop" } */
> +
> +#include "xop-check.h"
> +
> +#include <x86intrin.h>
> +
> +extern void abort (void);
> +
> +union
> +{
> + __m128i v;
> + unsigned char c[16];
> + unsigned short s[8];
> + unsigned int i[4];
> + unsigned long long l[2];
> +} a, b, c, d;
> +
> +#define TEST1(F, N, S, SS) \
> +do { \
> + for (i = 0; i < sizeof (a.F) / sizeof (a.F[0]); i++) \
> + a.F[i] = i * 17; \
> + s = _mm_set1_epi##SS (N); \
> + b.v = _mm_roti_epi##S (a.v, N); \
> + c.v = _mm_rot_epi##S (a.v, s); \
> + for (i = 0; i < sizeof (a.F) / sizeof (a.F[0]); i++) \
> + { \
> + int mask = __CHAR_BIT__ * sizeof (a.F[i]) - 1; \
> + d.F[i] = a.F[i] << (N & mask); \
> + if (N & mask) \
> + d.F[i] |= a.F[i] >> (mask + 1 - (N & mask)); \
> + if (b.F[i] != c.F[i] || b.F[i] != d.F[i]) \
> + abort (); \
> + } \
> +} while (0)
> +#define TEST(N) \
> + TEST1 (c, N, 8, 8); \
> + TEST1 (s, N, 16, 16); \
> + TEST1 (i, N, 32, 32); \
> + TEST1 (l, N, 64, 64x)
> +
> +volatile int n;
> +
> +static void
> +xop_test (void)
> +{
> + unsigned int i;
> + __m128i s;
> +
> +#ifndef NON_CONST
> + TEST (5);
> + TEST (-5);
> + TEST (0);
> + TEST (31);
> +#else
> + n = 5; TEST (n);
> + n = -5; TEST (n);
> + n = 0; TEST (n);
> + n = 31; TEST (n);
> +#endif
> +}
> --- gcc/testsuite/gcc.target/i386/xop-rotate2-int.c.jj 2011-06-15
> 11:25:42.000000000 +0200
> +++ gcc/testsuite/gcc.target/i386/xop-rotate2-int.c 2011-06-15
> 11:26:03.000000000 +0200
> @@ -0,0 +1,7 @@
> +/* PR target/49411 */
> +/* { dg-do run } */
> +/* { dg-require-effective-target xop } */
> +/* { dg-options "-O2 -mxop" } */
> +
> +#define NON_CONST 1
> +#include "xop-rotate1-int.c"
>
> Jakub
>
I will test on AMD HW.
--
Quentin
