Hi!
Before avx512vl we don't have a single instruction to do V2DImode and
V4DImode abs, but that isn't much different from say V4SImode before SSE3
where we also just emit a short sequence that is better than elementwise
expansion. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2018-04-30 Jakub Jelinek <ja...@redhat.com>
PR target/85572
* config/i386/i386.c (ix86_expand_sse2_abs): Handle E_V2DImode and
E_V4DImode.
* config/i386/sse.md (abs<mode>2): Use VI_AVX2 iterator instead of
VI1248_AVX512VL_AVX512BW. Handle V2DImode and V4DImode if not
TARGET_AVX512VL using ix86_expand_sse2_abs. Formatting fixes.
* g++.dg/other/sse2-pr85572-1.C: New test.
* g++.dg/other/sse2-pr85572-2.C: New test.
* g++.dg/other/sse4-pr85572-1.C: New test.
* g++.dg/other/avx2-pr85572-1.C: New test.
--- gcc/config/i386/i386.c.jj 2018-04-25 15:09:29.895453703 +0200
+++ gcc/config/i386/i386.c 2018-04-30 18:31:56.027101932 +0200
@@ -49806,39 +49806,74 @@ ix86_expand_sse2_abs (rtx target, rtx in
switch (mode)
{
+ case E_V2DImode:
+ case E_V4DImode:
+ /* For 64-bit signed integer X, with SSE4.2 use
+ pxor t0, t0; pcmpgtq X, t0; pxor t0, X; psubq t0, X.
+ Otherwise handle it similarly to V4SImode, except use 64 as W instead
of
+ 32 and use logical instead of arithmetic right shift (which is
+ unimplemented) and subtract. */
+ if (TARGET_SSE4_2)
+ {
+ tmp0 = gen_reg_rtx (mode);
+ tmp1 = gen_reg_rtx (mode);
+ emit_move_insn (tmp1, CONST0_RTX (mode));
+ if (mode == E_V2DImode)
+ emit_insn (gen_sse4_2_gtv2di3 (tmp0, tmp1, input));
+ else
+ emit_insn (gen_avx2_gtv4di3 (tmp0, tmp1, input));
+
+ tmp1 = expand_simple_binop (mode, XOR, tmp0, input,
+ NULL, 0, OPTAB_DIRECT);
+ x = expand_simple_binop (mode, MINUS, tmp1, tmp0,
+ target, 0, OPTAB_DIRECT);
+ break;
+ }
+
+ tmp0 = expand_simple_binop (mode, LSHIFTRT, input,
+ GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1),
+ NULL, 0, OPTAB_DIRECT);
+ tmp0 = expand_simple_unop (mode, NEG, tmp0, NULL, false);
+
+ tmp1 = expand_simple_binop (mode, XOR, tmp0, input,
+ NULL, 0, OPTAB_DIRECT);
+ x = expand_simple_binop (mode, MINUS, tmp1, tmp0,
+ target, 0, OPTAB_DIRECT);
+ break;
+
+ case E_V4SImode:
/* For 32-bit signed integer X, the best way to calculate the absolute
value of X is (((signed) X >> (W-1)) ^ X) - ((signed) X >> (W-1)). */
- case E_V4SImode:
- tmp0 = expand_simple_binop (mode, ASHIFTRT, input,
- GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1),
- NULL, 0, OPTAB_DIRECT);
- tmp1 = expand_simple_binop (mode, XOR, tmp0, input,
- NULL, 0, OPTAB_DIRECT);
- x = expand_simple_binop (mode, MINUS, tmp1, tmp0,
- target, 0, OPTAB_DIRECT);
- break;
+ tmp0 = expand_simple_binop (mode, ASHIFTRT, input,
+ GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1),
+ NULL, 0, OPTAB_DIRECT);
+ tmp1 = expand_simple_binop (mode, XOR, tmp0, input,
+ NULL, 0, OPTAB_DIRECT);
+ x = expand_simple_binop (mode, MINUS, tmp1, tmp0,
+ target, 0, OPTAB_DIRECT);
+ break;
+ case E_V8HImode:
/* For 16-bit signed integer X, the best way to calculate the absolute
value of X is max (X, -X), as SSE2 provides the PMAXSW insn. */
- case E_V8HImode:
- tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0);
+ tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0);
- x = expand_simple_binop (mode, SMAX, tmp0, input,
- target, 0, OPTAB_DIRECT);
- break;
+ x = expand_simple_binop (mode, SMAX, tmp0, input,
+ target, 0, OPTAB_DIRECT);
+ break;
+ case E_V16QImode:
/* For 8-bit signed integer X, the best way to calculate the absolute
value of X is min ((unsigned char) X, (unsigned char) (-X)),
as SSE2 provides the PMINUB insn. */
- case E_V16QImode:
- tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0);
+ tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0);
- x = expand_simple_binop (V16QImode, UMIN, tmp0, input,
- target, 0, OPTAB_DIRECT);
- break;
+ x = expand_simple_binop (V16QImode, UMIN, tmp0, input,
+ target, 0, OPTAB_DIRECT);
+ break;
- default:
- gcc_unreachable ();
+ default:
+ gcc_unreachable ();
}
if (x != target)
--- gcc/config/i386/sse.md.jj 2018-04-26 22:03:29.000000000 +0200
+++ gcc/config/i386/sse.md 2018-04-30 19:09:18.740838508 +0200
@@ -15174,12 +15174,14 @@ (define_insn "abs<mode>2_mask"
(set_attr "mode" "<sseinsnmode>")])
(define_expand "abs<mode>2"
- [(set (match_operand:VI1248_AVX512VL_AVX512BW 0 "register_operand")
- (abs:VI1248_AVX512VL_AVX512BW
- (match_operand:VI1248_AVX512VL_AVX512BW 1 "vector_operand")))]
+ [(set (match_operand:VI_AVX2 0 "register_operand")
+ (abs:VI_AVX2
+ (match_operand:VI_AVX2 1 "vector_operand")))]
"TARGET_SSE2"
{
- if (!TARGET_SSSE3)
+ if (!TARGET_SSSE3
+ || ((<MODE>mode == V2DImode || <MODE>mode == V4DImode)
+ && !TARGET_AVX512VL))
{
ix86_expand_sse2_abs (operands[0], operands[1]);
DONE;
--- gcc/testsuite/g++.dg/other/sse2-pr85572-1.C.jj 2018-04-30
18:40:29.566619137 +0200
+++ gcc/testsuite/g++.dg/other/sse2-pr85572-1.C 2018-04-30 19:04:04.111419300
+0200
@@ -0,0 +1,14 @@
+// PR target/85572
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O2 -msse2 -mno-sse3" }
+// { dg-final { scan-assembler-times {\mpxor\M} 2 } }
+// { dg-final { scan-assembler-times {\mpsubq\M} 2 } }
+// { dg-final { scan-assembler-times {\mpsrlq\M} 1 } }
+
+typedef long long V __attribute__((vector_size (16)));
+
+V
+foo (V x)
+{
+ return x < 0 ? -x : x;
+}
--- gcc/testsuite/g++.dg/other/sse2-pr85572-2.C.jj 2018-04-30
18:43:59.057879751 +0200
+++ gcc/testsuite/g++.dg/other/sse2-pr85572-2.C 2018-04-30 19:04:55.717493565
+0200
@@ -0,0 +1,45 @@
+// PR target/85572
+// { dg-do run { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O2 -msse2" }
+// { dg-require-effective-target sse2_runtime }
+
+typedef long long V __attribute__((vector_size (16)));
+typedef long long W __attribute__((vector_size (32)));
+
+__attribute__((noipa)) V
+foo (V x)
+{
+ return x < 0 ? -x : x;
+}
+
+__attribute__((noipa)) void
+bar (W *x, W *y)
+{
+ *y = *x < 0 ? -*x : *x;
+}
+
+int
+main ()
+{
+ V a = { 11LL, -15LL };
+ V b = foo (a);
+ if (b[0] != 11LL || b[1] != 15LL)
+ __builtin_abort ();
+ V c = { -123456789123456LL, 654321654321654LL };
+ V d = foo (c);
+ if (d[0] != 123456789123456LL || d[1] != 654321654321654LL)
+ __builtin_abort ();
+ V e = { 0, 1 };
+ V f = foo (e);
+ if (f[0] != 0 || f[1] != 1)
+ __builtin_abort ();
+ W g = { 17LL, -32LL, -123456789123456LL, 654321654321654LL }, h;
+ bar (&g, &h);
+ if (h[0] != 17LL || h[1] != 32LL
+ || h[2] != 123456789123456LL || h[3] != 654321654321654LL)
+ __builtin_abort ();
+ W i = { 0, 1, -1, 0 }, j;
+ bar (&i, &j);
+ if (j[0] != 0 || j[1] != 1 || j[2] != 1 || j[3] != 0)
+ __builtin_abort ();
+}
--- gcc/testsuite/g++.dg/other/sse4-pr85572-1.C.jj 2018-04-30
18:50:56.569420770 +0200
+++ gcc/testsuite/g++.dg/other/sse4-pr85572-1.C 2018-04-30 19:03:02.687330914
+0200
@@ -0,0 +1,14 @@
+// PR target/85572
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O2 -msse4 -mno-avx" }
+// { dg-final { scan-assembler-times {\mpxor\M} 2 } }
+// { dg-final { scan-assembler-times {\mpcmpgtq\M} 1 } }
+// { dg-final { scan-assembler-times {\mpsubq\M} 1 } }
+
+typedef long long V __attribute__((vector_size (16)));
+
+V
+foo (V x)
+{
+ return x < 0 ? -x : x;
+}
--- gcc/testsuite/g++.dg/other/avx2-pr85572-1.C.jj 2018-04-30
18:53:14.020598885 +0200
+++ gcc/testsuite/g++.dg/other/avx2-pr85572-1.C 2018-04-30 19:02:47.316308798
+0200
@@ -0,0 +1,21 @@
+// PR target/85572
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O2 -mavx2 -mno-avx512f" }
+// { dg-final { scan-assembler-times {\mvpxor\M} 4 } }
+// { dg-final { scan-assembler-times {\mvpcmpgtq\M} 2 } }
+// { dg-final { scan-assembler-times {\mvpsubq\M} 2 } }
+
+typedef long long V __attribute__((vector_size (16)));
+typedef long long W __attribute__((vector_size (32)));
+
+V
+foo (V x)
+{
+ return x < 0 ? -x : x;
+}
+
+W
+bar (W x)
+{
+ return x < 0 ? -x : x;
+}
Jakub
