On Thu, Feb 20, 2014 at 7:39 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> As discussed in the PR, gen_reg_rtx from when init_emit has not been
> initialized is highly undesirable. The following patch makes sure that
> for d->testing_p we never call gen_reg_rtx (i.e. from within
> ix86_vectorize_vec_perm_const_ok) and never try to emit insns.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux (together with Uros'
> patch to assert that gen_reg_rtx is not called when init_emit is not
> active) with RTL checking, further tested with
> GCC_TEST_RUN_EXPENSIVE=1 make -j16 -k check
> RUNTESTFLAGS='--target_board=unix\{-msse2,-msse3,-mssse3,-msse4,-mavx,-mavx2,-mavx512f\}
> dg-torture.exp=*vshuf*'
> (on AVX HW, so -mavx2 and -mavx512f tests expectedly failed execution,
> but at least didn't fail compilation, with the exception of
> gcc.dg/torture/vshuf-v8sf.c which ICEs with -mavx2 -DEXPENSIVE, but
> both without this patch and with this patch - will look at it eventually).
Attached is the complete backport of the patch to 4.7 branch. Jakub,
there is a slight churn in expand_vec_parm_interleave2 (there is no
dfinal.one_operand_p), can you please check if everything is OK there?
2014-02-21 Uros Bizjak <ubiz...@gmail.com>
Backport from mainline
2014-02-20 Jakub Jelinek <ja...@redhat.com>
* config/i386/i386.c (ix86_expand_vec_perm): Use V8SImode
mode for mask of V8SFmode permutation.
Backport from 4.8 branch
2014-02-20 Jakub Jelinek <ja...@redhat.com>
PR target/57896
* config/i386/i386.c (expand_vec_perm_interleave2): Don't call
gen_reg_rtx if d->testing_p.
(expand_vec_perm_pshufb2, expand_vec_perm_even_odd_1,
expand_vec_perm_broadcast_1): Return early if d->testing_p and
we will certainly return true.
Tested on x86_64-pc-linux-gnu {,-m32} with additional gen_reg_rtx
assert added, also with GCC_TEST_RUN_EXPENSIVE=1 -k check
RUNTESTFLAGS='--target_board=unix\{-msse2,-msse3,-mssse3,-msse4,-mavx,-mavx2\}
dg-torture.exp=*vshuf*' on AVX target, which uncovered the above
mentioned unrelated ICE, which is also fixed with this patch.
Uros.
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 207970)
+++ config/i386/i386.c (working copy)
@@ -20073,7 +20073,7 @@ ix86_expand_vec_perm (rtx operands[])
return;
case V8SFmode:
- mask = gen_lowpart (V8SFmode, mask);
+ mask = gen_lowpart (V8SImode, mask);
if (one_operand_shuffle)
emit_insn (gen_avx2_permvarv8sf (target, op0, mask));
else
@@ -36655,7 +36655,9 @@ expand_vec_perm_interleave2 (struct expand_vec_per
else
dfinal.perm[i] = e;
}
- dfinal.op0 = gen_reg_rtx (dfinal.vmode);
+
+ if (!d->testing_p)
+ dfinal.op0 = gen_reg_rtx (dfinal.vmode);
dfinal.op1 = dfinal.op0;
dremap.target = dfinal.op0;
@@ -37052,6 +37054,8 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
switch (d->vmode)
{
case V4DFmode:
+ if (d->testing_p)
+ break;
t1 = gen_reg_rtx (V4DFmode);
t2 = gen_reg_rtx (V4DFmode);
@@ -37071,6 +37075,8 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
{
int mask = odd ? 0xdd : 0x88;
+ if (d->testing_p)
+ break;
t1 = gen_reg_rtx (V8SFmode);
t2 = gen_reg_rtx (V8SFmode);
t3 = gen_reg_rtx (V8SFmode);
@@ -37112,6 +37118,8 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
return expand_vec_perm_pshufb2 (d);
else
{
+ if (d->testing_p)
+ break;
/* We need 2*log2(N)-1 operations to achieve odd/even
with interleave. */
t1 = gen_reg_rtx (V8HImode);
@@ -37133,6 +37141,8 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
return expand_vec_perm_pshufb2 (d);
else
{
+ if (d->testing_p)
+ break;
t1 = gen_reg_rtx (V16QImode);
t2 = gen_reg_rtx (V16QImode);
t3 = gen_reg_rtx (V16QImode);
@@ -37165,6 +37175,9 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
return expand_vec_perm_even_odd_1 (&d_copy, odd);
}
+ if (d->testing_p)
+ break;
+
t1 = gen_reg_rtx (V4DImode);
t2 = gen_reg_rtx (V4DImode);
@@ -37191,6 +37204,9 @@ expand_vec_perm_even_odd_1 (struct expand_vec_perm
return expand_vec_perm_even_odd_1 (&d_copy, odd);
}
+ if (d->testing_p)
+ break;
+
t1 = gen_reg_rtx (V8SImode);
t2 = gen_reg_rtx (V8SImode);
@@ -37283,6 +37299,8 @@ expand_vec_perm_broadcast_1 (struct expand_vec_per
case V16QImode:
/* These can be implemented via interleave. We save one insn by
stopping once we have promoted to V4SImode and then use pshufd. */
+ if (d->testing_p)
+ return true;
do
{
rtx dest;
