Hi!
In these patterns, we call simplify_gen_subreg on the input operand
to create paradoxical subregs that have 2x, 4x or 8x elements as the input
operand. That works fine if the input operand is a REG, but when it is a
SUBREG, RTL doesn't allow SUBREG of SUBREG and so relies on simplify_subreg
actually simplifying it. And e.g. if the input operand is a SUBREG that
changes the element mode (floating vs. non-floating) and then combined with
a paradoxical subreg (i.e. different size) this can easily fail, then
simplify_gen_subreg returns NULL but we still use it in instructions.
Fixed by forcing the operands into REG.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2021-02-10 Jakub Jelinek <ja...@redhat.com>
PR target/99025
* config/i386/sse.md (fix<fixunssuffix>_truncv2sfv2di2,
<insn>v8qiv8hi2, <insn>v8qiv8si2, <insn>v4qiv4si2, <insn>v4hiv4si2,
<insn>v8qiv8di2, <insn>v4qiv4di2, <insn>v2qiv2di2, <insn>v4hiv4di2,
<insn>v2hiv2di2, <insn>v2siv2di2): Force operands[1] into REG before
calling simplify_gen_subreg on it.
* gcc.target/i386/pr99025.c: New test.
--- gcc/config/i386/sse.md.jj 2021-02-10 07:52:32.673901634 +0100
+++ gcc/config/i386/sse.md 2021-02-10 10:57:37.229665371 +0100
@@ -6356,6 +6356,7 @@ (define_expand "fix<fixunssuffix>_truncv
(match_operand:V2SF 1 "register_operand")))]
"TARGET_AVX512DQ && TARGET_AVX512VL"
{
+ operands[1] = force_reg (V2SFmode, operands[1]);
operands[1] = simplify_gen_subreg (V4SFmode, operands[1], V2SFmode, 0);
emit_insn (gen_avx512dq_fix<fixunssuffix>_truncv2sfv2di2
(operands[0], operands[1]));
@@ -18013,6 +18014,7 @@ (define_expand "<insn>v8qiv8hi2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V8QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V8QImode, 0);
emit_insn (gen_sse4_1_<code>v8qiv8hi2 (operands[0], operands[1]));
DONE;
@@ -18090,6 +18092,7 @@ (define_expand "<insn>v8qiv8si2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V8QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V8QImode, 0);
emit_insn (gen_avx2_<code>v8qiv8si2 (operands[0], operands[1]));
DONE;
@@ -18153,6 +18156,7 @@ (define_expand "<insn>v4qiv4si2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V4QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V4QImode, 0);
emit_insn (gen_sse4_1_<code>v4qiv4si2 (operands[0], operands[1]));
DONE;
@@ -18279,6 +18283,7 @@ (define_expand "<insn>v4hiv4si2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V4HImode, operands[1]);
operands[1] = simplify_gen_subreg (V8HImode, operands[1], V4HImode, 0);
emit_insn (gen_sse4_1_<code>v4hiv4si2 (operands[0], operands[1]));
DONE;
@@ -18366,6 +18371,7 @@ (define_expand "<insn>v8qiv8di2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V8QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V8QImode, 0);
emit_insn (gen_avx512f_<code>v8qiv8di2 (operands[0], operands[1]));
DONE;
@@ -18427,6 +18433,7 @@ (define_expand "<insn>v4qiv4di2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V8QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V8QImode, 0);
emit_insn (gen_avx2_<code>v4qiv4di2 (operands[0], operands[1]));
DONE;
@@ -18453,6 +18460,7 @@ (define_expand "<insn>v2qiv2di2"
(match_operand:V2QI 1 "register_operand")))]
"TARGET_SSE4_1"
{
+ operands[1] = force_reg (V2QImode, operands[1]);
operands[1] = simplify_gen_subreg (V16QImode, operands[1], V2QImode, 0);
emit_insn (gen_sse4_1_<code>v2qiv2di2 (operands[0], operands[1]));
DONE;
@@ -18525,6 +18533,7 @@ (define_expand "<insn>v4hiv4di2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V4HImode, operands[1]);
operands[1] = simplify_gen_subreg (V8HImode, operands[1], V4HImode, 0);
emit_insn (gen_avx2_<code>v4hiv4di2 (operands[0], operands[1]));
DONE;
@@ -18586,6 +18595,7 @@ (define_expand "<insn>v2hiv2di2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V2HImode, operands[1]);
operands[1] = simplify_gen_subreg (V8HImode, operands[1], V2HImode, 0);
emit_insn (gen_sse4_1_<code>v2hiv2di2 (operands[0], operands[1]));
DONE;
@@ -18737,6 +18747,7 @@ (define_expand "<insn>v2siv2di2"
{
if (!MEM_P (operands[1]))
{
+ operands[1] = force_reg (V2SImode, operands[1]);
operands[1] = simplify_gen_subreg (V4SImode, operands[1], V2SImode, 0);
emit_insn (gen_sse4_1_<code>v2siv2di2 (operands[0], operands[1]));
DONE;
--- gcc/testsuite/gcc.target/i386/pr99025.c.jj 2021-02-09 19:17:29.705924814
+0100
+++ gcc/testsuite/gcc.target/i386/pr99025.c 2021-02-09 19:17:10.687137688
+0100
@@ -0,0 +1,17 @@
+/* PR target/99025 */
+/* { dg-do compile } */
+/* { dg-options "-O3 -msse4" } */
+
+long v[16];
+int w;
+union U { float u; int r; } x;
+
+void
+foo (float y)
+{
+ union U z;
+ x.u = w;
+ v[5] = x.r;
+ z.u = y;
+ v[6] = z.r;
+}
Jakub
