Skip to content

R
riscv-gcc-1
Commit e9ddec4e by Uros Bizjak Committed by Uros Bizjak
Options

re PR target/33329 (ICE in expand_simple_binop, at optabs.c:1294) 

        PR target/33329
        PR target/26449
        * config/i386/sse.md (mulv4si3): Do not expand sse2 sequence.
        (*sse2_mulv4si3): New define_insn_and_split pattern. Split insn in
        split1 pass.
        (mulv16qi3): Implement as define_insn_and_split pattern instead of
        define_expand, to split insn in split1 pass.
        (mulv2di3): Ditto.

testsuite/ChangeLog:

        PR target/33329
        PR target/26449
        * gcc.target/i386/pr33329.c: New file.

From-SVN: r128269
parent 79646678
Showing with 100 additions and 48 deletions
gcc/ChangeLog
2007-09-08 Uros Bizjak <ubizjak@gmail.com>
PR target/33329
PR target/26449
* config/i386/sse.md (mulv4si3): Do not expand sse2 sequence.
(*sse2_mulv4si3): New define_insn_and_split pattern. Split insn in
split1 pass.
(mulv16qi3): Implement as define_insn_and_split pattern instead of
define_expand, to split insn in split1 pass.
(mulv2di3): Ditto.
2007-09-08 Dorit Nuzman <dorit@il.ibm.com> 2007-09-08 Dorit Nuzman <dorit@il.ibm.com>
PR tree-optimization/33301 PR tree-optimization/33301
gcc/config/i386/sse.md
...@@ -2899,11 +2899,15 @@ ...@@ -2899,11 +2899,15 @@
(set_attr "prefix_data16" "1") (set_attr "prefix_data16" "1")
(set_attr "mode" "TI")]) (set_attr "mode" "TI")])
(define_expand "mulv16qi3" (define_insn_and_split "mulv16qi3"
[(set (match_operand:V16QI 0 "register_operand" "") [(set (match_operand:V16QI 0 "register_operand" "")
(mult:V16QI (match_operand:V16QI 1 "register_operand" "") (mult:V16QI (match_operand:V16QI 1 "register_operand" "")
(match_operand:V16QI 2 "register_operand" "")))] (match_operand:V16QI 2 "register_operand" "")))]
"TARGET_SSE2" "TARGET_SSE2
&& !(reload_completed || reload_in_progress)"
"#"
"&& 1"
[(const_int 0)]
{ {
rtx t[12], op0; rtx t[12], op0;
int i; int i;
...@@ -3097,50 +3101,6 @@ ...@@ -3097,50 +3101,6 @@
{ {
if (TARGET_SSE4_1) if (TARGET_SSE4_1)
ix86_fixup_binary_operands_no_copy (MULT, V4SImode, operands); ix86_fixup_binary_operands_no_copy (MULT, V4SImode, operands);
else
{
rtx t1, t2, t3, t4, t5, t6, thirtytwo;
rtx op0, op1, op2;
op0 = operands[0];
op1 = operands[1];
op2 = operands[2];
t1 = gen_reg_rtx (V4SImode);
t2 = gen_reg_rtx (V4SImode);
t3 = gen_reg_rtx (V4SImode);
t4 = gen_reg_rtx (V4SImode);
t5 = gen_reg_rtx (V4SImode);
t6 = gen_reg_rtx (V4SImode);
thirtytwo = GEN_INT (32);
/* Multiply elements 2 and 0. */
emit_insn (gen_sse2_umulv2siv2di3 (gen_lowpart (V2DImode, t1),
op1, op2));
/* Shift both input vectors down one element, so that elements 3
and 1 are now in the slots for elements 2 and 0. For K8, at
least, this is faster than using a shuffle. */
emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode, t2),
gen_lowpart (TImode, op1),
thirtytwo));
emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode, t3),
gen_lowpart (TImode, op2),
thirtytwo));
/* Multiply elements 3 and 1. */
emit_insn (gen_sse2_umulv2siv2di3 (gen_lowpart (V2DImode, t4),
t2, t3));
/* Move the results in element 2 down to element 1; we don't care
what goes in elements 2 and 3. */
emit_insn (gen_sse2_pshufd_1 (t5, t1, const0_rtx, const2_rtx,
const0_rtx, const0_rtx));
emit_insn (gen_sse2_pshufd_1 (t6, t4, const0_rtx, const2_rtx,
const0_rtx, const0_rtx));
/* Merge the parts back together. */
emit_insn (gen_sse2_punpckldq (op0, t5, t6));
DONE;
}
}) })
(define_insn "*sse4_1_mulv4si3" (define_insn "*sse4_1_mulv4si3"
...@@ -3153,11 +3113,68 @@ ...@@ -3153,11 +3113,68 @@
(set_attr "prefix_extra" "1") (set_attr "prefix_extra" "1")
(set_attr "mode" "TI")]) (set_attr "mode" "TI")])
(define_expand "mulv2di3" (define_insn_and_split "*sse2_mulv4si3"
[(set (match_operand:V4SI 0 "register_operand" "")
(mult:V4SI (match_operand:V4SI 1 "register_operand" "")
(match_operand:V4SI 2 "register_operand" "")))]
"TARGET_SSE2 && !TARGET_SSE4_1
&& !(reload_completed || reload_in_progress)"
"#"
"&& 1"
[(const_int 0)]
{
rtx t1, t2, t3, t4, t5, t6, thirtytwo;
rtx op0, op1, op2;
op0 = operands[0];
op1 = operands[1];
op2 = operands[2];
t1 = gen_reg_rtx (V4SImode);
t2 = gen_reg_rtx (V4SImode);
t3 = gen_reg_rtx (V4SImode);
t4 = gen_reg_rtx (V4SImode);
t5 = gen_reg_rtx (V4SImode);
t6 = gen_reg_rtx (V4SImode);
thirtytwo = GEN_INT (32);
/* Multiply elements 2 and 0. */
emit_insn (gen_sse2_umulv2siv2di3 (gen_lowpart (V2DImode, t1),
op1, op2));
/* Shift both input vectors down one element, so that elements 3
and 1 are now in the slots for elements 2 and 0. For K8, at
least, this is faster than using a shuffle. */
emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode, t2),
gen_lowpart (TImode, op1),
thirtytwo));
emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode, t3),
gen_lowpart (TImode, op2),
thirtytwo));
/* Multiply elements 3 and 1. */
emit_insn (gen_sse2_umulv2siv2di3 (gen_lowpart (V2DImode, t4),
t2, t3));
/* Move the results in element 2 down to element 1; we don't care
what goes in elements 2 and 3. */
emit_insn (gen_sse2_pshufd_1 (t5, t1, const0_rtx, const2_rtx,
const0_rtx, const0_rtx));
emit_insn (gen_sse2_pshufd_1 (t6, t4, const0_rtx, const2_rtx,
const0_rtx, const0_rtx));
/* Merge the parts back together. */
emit_insn (gen_sse2_punpckldq (op0, t5, t6));
DONE;
})
(define_insn_and_split "mulv2di3"
[(set (match_operand:V2DI 0 "register_operand" "") [(set (match_operand:V2DI 0 "register_operand" "")
(mult:V2DI (match_operand:V2DI 1 "register_operand" "") (mult:V2DI (match_operand:V2DI 1 "register_operand" "")
(match_operand:V2DI 2 "register_operand" "")))] (match_operand:V2DI 2 "register_operand" "")))]
"TARGET_SSE2" "TARGET_SSE2
&& !(reload_completed || reload_in_progress)"
"#"
"&& 1"
[(const_int 0)]
{ {
rtx t1, t2, t3, t4, t5, t6, thirtytwo; rtx t1, t2, t3, t4, t5, t6, thirtytwo;
rtx op0, op1, op2; rtx op0, op1, op2;
......
gcc/testsuite/ChangeLog
2007-09-08 Uros Bizjak <ubizjak@gmail.com>
PR target/33329
PR target/26449
* gcc.target/i386/pr33329.c: New file.
2007-09-08 Eric Botcazou <ebotcazou@adacore.com> 2007-09-08 Eric Botcazou <ebotcazou@adacore.com>
* gnat.dg/renaming3.adb, renaming4.ads: New test. * gnat.dg/renaming3.adb, renaming4.ads: New test.
gcc/testsuite/gcc.target/i386/pr33329.c 0 → 100644
/* { dg-do compile } */
/* { dg-options "-O2 -ftree-vectorize -msse2" } */
extern void g (int *);
void f (void)
{
int tabs[8], tabcount;
for (tabcount = 1; tabcount <= 8; tabcount += 7)
{
int i;
for (i = 0; i < 8; i++)
tabs[i] = i * 2;
g (tabs);
}
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment