[PATCH] Fix x86 ICE with -mtune=amdfam10 -mno-sse2 (PR target/81121)

[PATCH] Fix x86 ICE with -mtune=amdfam10 -mno-sse2 (PR target/81121)

[Jakub Jelinek]

Hi!

This testcase started to ICE when PR70873 fix changed the splitter:
@@ -5153,11 +5147,11 @@
 ;; slots when !TARGET_INTER_UNIT_MOVES_TO_VEC disables the general_regs
 ;; alternative in sse2_loadld.
 (define_split
-  [(set (match_operand:MODEF 0 "register_operand")
+  [(set (match_operand:MODEF 0 "sse_reg_operand")
 	(float:MODEF (match_operand:SI 1 "nonimmediate_operand")))]
-  "TARGET_SSE2 && TARGET_SSE_MATH
-   && TARGET_USE_VECTOR_CONVERTS && optimize_function_for_speed_p (cfun)
-   && reload_completed && SSE_REG_P (operands[0])
+  "TARGET_USE_VECTOR_CONVERTS
+   && optimize_function_for_speed_p (cfun)
+   && reload_completed
    && (MEM_P (operands[1]) || TARGET_INTER_UNIT_MOVES_TO_VEC)
    && (!EXT_REX_SSE_REG_P (operands[0])
        || TARGET_AVX512VL)"
Having sse_reg_operand match the output operand does not imply
TARGET_SSE2 is enabled, but we need it for both the
  if (<ssevecmode>mode == V4SFmode)
    emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
  else
    emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
instructions that we want to use in the splitter.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk
(or do you want TARGET_SSE2 first or right after
TARGET_USE_VECTOR_CONVERTS)?

2017-06-19  Jakub Jelinek  <jakub@redhat.com>

	PR target/81121
	* config/i386/i386.md (TARGET_USE_VECTOR_CONVERTS float si->{sf,df}
	splitter): Require TARGET_SSE2 in the condition.

	* gcc.target/i386/pr81121.c: New test.

--- gcc/config/i386/i386.md.jj	2017-06-08 20:50:46.000000000 +0200
+++ gcc/config/i386/i386.md	2017-06-19 11:30:38.937491668 +0200
@@ -5294,6 +5294,7 @@ (define_split
    && optimize_function_for_speed_p (cfun)
    && reload_completed
    && (MEM_P (operands[1]) || TARGET_INTER_UNIT_MOVES_TO_VEC)
+   && TARGET_SSE2
    && (!EXT_REX_SSE_REG_P (operands[0])
        || TARGET_AVX512VL)"
   [(const_int 0)]
--- gcc/testsuite/gcc.target/i386/pr81121.c.jj	2017-06-19 11:36:06.545501324 +0200
+++ gcc/testsuite/gcc.target/i386/pr81121.c	2017-06-19 11:35:40.000000000 +0200
@@ -0,0 +1,10 @@
+/* PR target/81121 */
+/* { dg-do compile } */
+/* { dg-options "-O0 -march=amdfam10 -mno-sse2" } */
+
+void
+foo (short *x, short *y)
+{
+  float a = 0;
+  y[0] = x[0] * a;
+}

	Jakub

Re: [PATCH] Fix x86 ICE with -mtune=amdfam10 -mno-sse2 (PR target/81121)

[Uros Bizjak]

On Mon, Jun 19, 2017 at 5:37 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> This testcase started to ICE when PR70873 fix changed the splitter:
> @@ -5153,11 +5147,11 @@
>  ;; slots when !TARGET_INTER_UNIT_MOVES_TO_VEC disables the general_regs
>  ;; alternative in sse2_loadld.
>  (define_split
> -  [(set (match_operand:MODEF 0 "register_operand")
> +  [(set (match_operand:MODEF 0 "sse_reg_operand")
>         (float:MODEF (match_operand:SI 1 "nonimmediate_operand")))]
> -  "TARGET_SSE2 && TARGET_SSE_MATH
> -   && TARGET_USE_VECTOR_CONVERTS && optimize_function_for_speed_p (cfun)
> -   && reload_completed && SSE_REG_P (operands[0])
> +  "TARGET_USE_VECTOR_CONVERTS
> +   && optimize_function_for_speed_p (cfun)
> +   && reload_completed
>     && (MEM_P (operands[1]) || TARGET_INTER_UNIT_MOVES_TO_VEC)
>     && (!EXT_REX_SSE_REG_P (operands[0])
>         || TARGET_AVX512VL)"
> Having sse_reg_operand match the output operand does not imply
> TARGET_SSE2 is enabled, but we need it for both the
>   if (<ssevecmode>mode == V4SFmode)
>     emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
>   else
>     emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
> instructions that we want to use in the splitter.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk
> (or do you want TARGET_SSE2 first or right after
> TARGET_USE_VECTOR_CONVERTS)?

Please put TARGET_SSE2 first in the insn condition.

> 2017-06-19  Jakub Jelinek  <jakub@redhat.com>
>
>         PR target/81121
>         * config/i386/i386.md (TARGET_USE_VECTOR_CONVERTS float si->{sf,df}
>         splitter): Require TARGET_SSE2 in the condition.
>
>         * gcc.target/i386/pr81121.c: New test.

OK with the above change.

Thanks,
Uros.

> --- gcc/config/i386/i386.md.jj  2017-06-08 20:50:46.000000000 +0200
> +++ gcc/config/i386/i386.md     2017-06-19 11:30:38.937491668 +0200
> @@ -5294,6 +5294,7 @@ (define_split
>     && optimize_function_for_speed_p (cfun)
>     && reload_completed
>     && (MEM_P (operands[1]) || TARGET_INTER_UNIT_MOVES_TO_VEC)
> +   && TARGET_SSE2
>     && (!EXT_REX_SSE_REG_P (operands[0])
>         || TARGET_AVX512VL)"
>    [(const_int 0)]
> --- gcc/testsuite/gcc.target/i386/pr81121.c.jj  2017-06-19 11:36:06.545501324 +0200
> +++ gcc/testsuite/gcc.target/i386/pr81121.c     2017-06-19 11:35:40.000000000 +0200
> @@ -0,0 +1,10 @@
> +/* PR target/81121 */
> +/* { dg-do compile } */
> +/* { dg-options "-O0 -march=amdfam10 -mno-sse2" } */
> +
> +void
> +foo (short *x, short *y)
> +{
> +  float a = 0;
> +  y[0] = x[0] * a;
> +}
>
>         Jakub