diff mbox series

[v1] Update check_effective_target_vect_int_mod according to LoongArch SX/ASX capabilities.

Message ID 20230920011557.106634-1-panchenghui@loongson.cn
State New
Headers show
Series [v1] Update check_effective_target_vect_int_mod according to LoongArch SX/ASX capabilities. | expand

Commit Message

Chenghui Pan Sept. 20, 2023, 1:15 a.m. UTC 
LoongArch failed to pass gcc.dg/pr104992.c with -mlsx and -mlasx. This test uses
different dg-final directives depending on the vect_int_mod result, LoongArch
SX/ASX supports this operations but corresponding description is not defined in
target-supports.exp. This patch solves the problem above with some
modification in proc check_effective_target_vect_int_mod.

gcc/testsuite/ChangeLog:

	* lib/target-supports.exp: Update check_effective_target_vect_int_mod according to
	LoongArch SX/ASX capabilities.
---
 gcc/testsuite/lib/target-supports.exp | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

Comments

Xi Ruoyao Sept. 24, 2023, 10:05 a.m. UTC | #1 
On Wed, 2023-09-20 at 09:15 +0800, Chenghui Pan wrote:
> LoongArch failed to pass gcc.dg/pr104992.c with -mlsx and -mlasx. This test uses
> different dg-final directives depending on the vect_int_mod result, LoongArch
> SX/ASX supports this operations but corresponding description is not defined in
> target-supports.exp. This patch solves the problem above with some
> modification in proc check_effective_target_vect_int_mod.

I think we can just add -mdouble-float -mlasx into DEFAULT_VECTCFLAGS
and always enable vect_int_mod for LoongArch.  This will make vect.exp
tests automatically run for every "make check" on LoongArch.

> gcc/testsuite/ChangeLog:
> 
> 	* lib/target-supports.exp: Update check_effective_target_vect_int_mod according to
> 	LoongArch SX/ASX capabilities.
> ---
>  gcc/testsuite/lib/target-supports.exp | 18 ++++++++++++++++++
>  1 file changed, 18 insertions(+)
> 
> diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp
> index 2de41cef2f6..b253dc578d2 100644
> --- a/gcc/testsuite/lib/target-supports.exp
> +++ b/gcc/testsuite/lib/target-supports.exp
> @@ -8586,6 +8586,8 @@ proc check_effective_target_vect_int_mod { } {
>      return [check_cached_effective_target_indexed vect_int_mod {
>        expr { ([istarget powerpc*-*-*]
>  	      && [check_effective_target_has_arch_pwr10])
> +	     || ([istarget loongarch*-*-*]
> +		 && [check_effective_target_loongarch_sx])
>               || [istarget amdgcn-*-*] }}]
>  }
>  
> @@ -12656,6 +12658,22 @@ proc check_effective_target_const_volatile_readonly_section { } {
>    return 1
>  }
>  
> +proc check_effective_target_loongarch_sx { } {
> +    return [check_no_compiler_messages loongarch_lsx assembly {
> +       #if !defined(__loongarch_sx)
> +       #error "LSX not defined"
> +       #endif
> +    }]
> +}
> +
> +proc check_effective_target_loongarch_asx { } {
> +    return [check_no_compiler_messages loongarch_asx assembly {
> +       #if !defined(__loongarch_asx)
> +       #error "LASX not defined"
> +       #endif
> +    }]
> +}
> +
>  # Appends necessary Python flags to extra-tool-flags if Python.h is supported.
>  # Otherwise, modifies dg-do-what.
>  proc dg-require-python-h { args } {
Chenghui Pan Sept. 25, 2023, 3:33 a.m. UTC | #2 
Thanks for your advice! I will check it out and submit a new version of
patch.

On Sun, 2023-09-24 at 18:05 +0800, Xi Ruoyao wrote:
> On Wed, 2023-09-20 at 09:15 +0800, Chenghui Pan wrote:
> > LoongArch failed to pass gcc.dg/pr104992.c with -mlsx and -mlasx.
> > This test uses
> > different dg-final directives depending on the vect_int_mod result,
> > LoongArch
> > SX/ASX supports this operations but corresponding description is
> > not defined in
> > target-supports.exp. This patch solves the problem above with some
> > modification in proc check_effective_target_vect_int_mod.
> 
> I think we can just add -mdouble-float -mlasx into DEFAULT_VECTCFLAGS
> and always enable vect_int_mod for LoongArch.  This will make
> vect.exp
> tests automatically run for every "make check" on LoongArch.
> 
> > gcc/testsuite/ChangeLog:
> > 
> >         * lib/target-supports.exp: Update
> > check_effective_target_vect_int_mod according to
> >         LoongArch SX/ASX capabilities.
> > ---
> >  gcc/testsuite/lib/target-supports.exp | 18 ++++++++++++++++++
> >  1 file changed, 18 insertions(+)
> > 
> > diff --git a/gcc/testsuite/lib/target-supports.exp
> > b/gcc/testsuite/lib/target-supports.exp
> > index 2de41cef2f6..b253dc578d2 100644
> > --- a/gcc/testsuite/lib/target-supports.exp
> > +++ b/gcc/testsuite/lib/target-supports.exp
> > @@ -8586,6 +8586,8 @@ proc check_effective_target_vect_int_mod { }
> > {
> >      return [check_cached_effective_target_indexed vect_int_mod {
> >        expr { ([istarget powerpc*-*-*]
> >               && [check_effective_target_has_arch_pwr10])
> > +            || ([istarget loongarch*-*-*]
> > +                && [check_effective_target_loongarch_sx])
> >               || [istarget amdgcn-*-*] }}]
> >  }
> >  
> > @@ -12656,6 +12658,22 @@ proc
> > check_effective_target_const_volatile_readonly_section { } {
> >    return 1
> >  }
> >  
> > +proc check_effective_target_loongarch_sx { } {
> > +    return [check_no_compiler_messages loongarch_lsx assembly {
> > +       #if !defined(__loongarch_sx)
> > +       #error "LSX not defined"
> > +       #endif
> > +    }]
> > +}
> > +
> > +proc check_effective_target_loongarch_asx { } {
> > +    return [check_no_compiler_messages loongarch_asx assembly {
> > +       #if !defined(__loongarch_asx)
> > +       #error "LASX not defined"
> > +       #endif
> > +    }]
> > +}
> > +
> >  # Appends necessary Python flags to extra-tool-flags if Python.h
> > is supported.
> >  # Otherwise, modifies dg-do-what.
> >  proc dg-require-python-h { args } {
>
Chenghui Pan Sept. 25, 2023, 9:38 a.m. UTC | #3 
Hi!

After some attemptions, I think we still ne to check
"check_effective_target_loongarch_sx" in vect_int_mod. I wrote some
temp logics in gcc/testsuite/lib/target-supports.exp like this:

diff --git a/gcc/testsuite/lib/target-supports.exp
b/gcc/testsuite/lib/target-supports.exp
index 2de41cef2f6..91e1c22a6e1 100644
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -8586,7 +8586,8 @@ proc check_effective_target_vect_int_mod { } {
     return [check_cached_effective_target_indexed vect_int_mod {
       expr { ([istarget powerpc*-*-*]
              && [check_effective_target_has_arch_pwr10])
-             || [istarget amdgcn-*-*] }}]
+             || [istarget loongarch*-*-*]
+             || [istarget amdgcn-*-*] }}]
 }
 
 # Return 1 if the target supports vector even/odd elements extraction,
0 otherwise.
@@ -11174,6 +11175,12 @@ proc check_vect_support_and_set_flags { } {
            lappend DEFAULT_VECTCFLAGS "--param" "riscv-vector-abi"
            set dg-do-what-default compile
        }
+    } elseif [istarget loongarch*-*-*] {
+      if [check_effective_target_loongarch_asx_hw] {
+         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlasx"
+      } elseif [check_effective_target_loongarch_sx_hw] {
+         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlsx"
+      }
     } else {
         return 0
     }
\* temp impl of sx/asx hw proc *\

And then in make check without --target_board=unix/-mlasx, vect.exp is
invoked with expected vector isa options, but pr104992.c failed because
it expected result with "vect_int_mod returns 1" but it was compiled
without -mlsx/-mlasx. Seems pr104992.c is invoked by gcc.dg/dg.exp,
pr104992.c is not affected by DEFAULT_CFLAGS, so we still need to check
if LSX/LASX is available in vect_int_mod. 

Other parts of new patch is still WIP.

On Sun, 2023-09-24 at 18:05 +0800, Xi Ruoyao wrote:
> On Wed, 2023-09-20 at 09:15 +0800, Chenghui Pan wrote:
> > LoongArch failed to pass gcc.dg/pr104992.c with -mlsx and -mlasx.
> > This test uses
> > different dg-final directives depending on the vect_int_mod result,
> > LoongArch
> > SX/ASX supports this operations but corresponding description is
> > not defined in
> > target-supports.exp. This patch solves the problem above with some
> > modification in proc check_effective_target_vect_int_mod.
> 
> I think we can just add -mdouble-float -mlasx into DEFAULT_VECTCFLAGS
> and always enable vect_int_mod for LoongArch.  This will make
> vect.exp
> tests automatically run for every "make check" on LoongArch.
> 
> > gcc/testsuite/ChangeLog:
> > 
> >         * lib/target-supports.exp: Update
> > check_effective_target_vect_int_mod according to
> >         LoongArch SX/ASX capabilities.
> > ---
> >  gcc/testsuite/lib/target-supports.exp | 18 ++++++++++++++++++
> >  1 file changed, 18 insertions(+)
> > 
> > diff --git a/gcc/testsuite/lib/target-supports.exp
> > b/gcc/testsuite/lib/target-supports.exp
> > index 2de41cef2f6..b253dc578d2 100644
> > --- a/gcc/testsuite/lib/target-supports.exp
> > +++ b/gcc/testsuite/lib/target-supports.exp
> > @@ -8586,6 +8586,8 @@ proc check_effective_target_vect_int_mod { }
> > {
> >      return [check_cached_effective_target_indexed vect_int_mod {
> >        expr { ([istarget powerpc*-*-*]
> >               && [check_effective_target_has_arch_pwr10])
> > +            || ([istarget loongarch*-*-*]
> > +                && [check_effective_target_loongarch_sx])
> >               || [istarget amdgcn-*-*] }}]
> >  }
> >  
> > @@ -12656,6 +12658,22 @@ proc
> > check_effective_target_const_volatile_readonly_section { } {
> >    return 1
> >  }
> >  
> > +proc check_effective_target_loongarch_sx { } {
> > +    return [check_no_compiler_messages loongarch_lsx assembly {
> > +       #if !defined(__loongarch_sx)
> > +       #error "LSX not defined"
> > +       #endif
> > +    }]
> > +}
> > +
> > +proc check_effective_target_loongarch_asx { } {
> > +    return [check_no_compiler_messages loongarch_asx assembly {
> > +       #if !defined(__loongarch_asx)
> > +       #error "LASX not defined"
> > +       #endif
> > +    }]
> > +}
> > +
> >  # Appends necessary Python flags to extra-tool-flags if Python.h
> > is supported.
> >  # Otherwise, modifies dg-do-what.
> >  proc dg-require-python-h { args } {
>
Xi Ruoyao Sept. 25, 2023, 9:44 a.m. UTC | #4 
On Mon, 2023-09-25 at 17:38 +0800, Chenghui Pan wrote:
> Hi!
> 
> After some attemptions, I think we still ne to check
> "check_effective_target_loongarch_sx" in vect_int_mod. I wrote some
> temp logics in gcc/testsuite/lib/target-supports.exp like this:
> 
> diff --git a/gcc/testsuite/lib/target-supports.exp
> b/gcc/testsuite/lib/target-supports.exp
> index 2de41cef2f6..91e1c22a6e1 100644
> --- a/gcc/testsuite/lib/target-supports.exp
> +++ b/gcc/testsuite/lib/target-supports.exp
> @@ -8586,7 +8586,8 @@ proc check_effective_target_vect_int_mod { } {
>      return [check_cached_effective_target_indexed vect_int_mod {
>        expr { ([istarget powerpc*-*-*]
>               && [check_effective_target_has_arch_pwr10])
> -             || [istarget amdgcn-*-*] }}]
> +             || [istarget loongarch*-*-*]
> +             || [istarget amdgcn-*-*] }}]
>  }
>  
>  # Return 1 if the target supports vector even/odd elements extraction,
> 0 otherwise.
> @@ -11174,6 +11175,12 @@ proc check_vect_support_and_set_flags { } {
>             lappend DEFAULT_VECTCFLAGS "--param" "riscv-vector-abi"
>             set dg-do-what-default compile
>         }
> +    } elseif [istarget loongarch*-*-*] {
> +      if [check_effective_target_loongarch_asx_hw] {
> +         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlasx"
> +      } elseif [check_effective_target_loongarch_sx_hw] {
> +         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlsx"
> +      }

I think we can always enable LASX in DEFAULT_VECTCFLAGS, but set dg-do-
what-default to "run" only if both the hardware and the kernel supports
LASX.  If the kernel or the hardware is not capable we set dg-do-what-
default to "compile".

>      } else {
>          return 0
>      }
> \* temp impl of sx/asx hw proc *\
> 
> And then in make check without --target_board=unix/-mlasx, vect.exp is
> invoked with expected vector isa options, but pr104992.c failed because
> it expected result with "vect_int_mod returns 1" but it was compiled
> without -mlsx/-mlasx. Seems pr104992.c is invoked by gcc.dg/dg.exp,
> pr104992.c is not affected by DEFAULT_CFLAGS, so we still need to check
> if LSX/LASX is available in vect_int_mod. 
> 
> Other parts of new patch is still WIP.
Chenghui Pan Sept. 25, 2023, 9:51 a.m. UTC | #5 
Thanks! I will try to improve it.

On Mon, 2023-09-25 at 17:44 +0800, Xi Ruoyao wrote:
> On Mon, 2023-09-25 at 17:38 +0800, Chenghui Pan wrote:
> > Hi!
> > 
> > After some attemptions, I think we still ne to check
> > "check_effective_target_loongarch_sx" in vect_int_mod. I wrote some
> > temp logics in gcc/testsuite/lib/target-supports.exp like this:
> > 
> > diff --git a/gcc/testsuite/lib/target-supports.exp
> > b/gcc/testsuite/lib/target-supports.exp
> > index 2de41cef2f6..91e1c22a6e1 100644
> > --- a/gcc/testsuite/lib/target-supports.exp
> > +++ b/gcc/testsuite/lib/target-supports.exp
> > @@ -8586,7 +8586,8 @@ proc check_effective_target_vect_int_mod { }
> > {
> >      return [check_cached_effective_target_indexed vect_int_mod {
> >        expr { ([istarget powerpc*-*-*]
> >               && [check_effective_target_has_arch_pwr10])
> > -             || [istarget amdgcn-*-*] }}]
> > +             || [istarget loongarch*-*-*]
> > +             || [istarget amdgcn-*-*] }}]
> >  }
> >  
> >  # Return 1 if the target supports vector even/odd elements
> > extraction,
> > 0 otherwise.
> > @@ -11174,6 +11175,12 @@ proc check_vect_support_and_set_flags { }
> > {
> >             lappend DEFAULT_VECTCFLAGS "--param" "riscv-vector-abi"
> >             set dg-do-what-default compile
> >         }
> > +    } elseif [istarget loongarch*-*-*] {
> > +      if [check_effective_target_loongarch_asx_hw] {
> > +         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlasx"
> > +      } elseif [check_effective_target_loongarch_sx_hw] {
> > +         lappend DEFAULT_VECTCFLAGS "-mdouble-float" "-mlsx"
> > +      }
> 
> I think we can always enable LASX in DEFAULT_VECTCFLAGS, but set dg-
> do-
> what-default to "run" only if both the hardware and the kernel
> supports
> LASX.  If the kernel or the hardware is not capable we set dg-do-
> what-
> default to "compile".
> 
> >      } else {
> >          return 0
> >      }
> > \* temp impl of sx/asx hw proc *\
> > 
> > And then in make check without --target_board=unix/-mlasx, vect.exp
> > is
> > invoked with expected vector isa options, but pr104992.c failed
> > because
> > it expected result with "vect_int_mod returns 1" but it was
> > compiled
> > without -mlsx/-mlasx. Seems pr104992.c is invoked by gcc.dg/dg.exp,
> > pr104992.c is not affected by DEFAULT_CFLAGS, so we still need to
> > check
> > if LSX/LASX is available in vect_int_mod. 
> > 
> > Other parts of new patch is still WIP.
>
diff mbox series

Patch

diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp
index 2de41cef2f6..b253dc578d2 100644
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -8586,6 +8586,8 @@  proc check_effective_target_vect_int_mod { } {
     return [check_cached_effective_target_indexed vect_int_mod {
       expr { ([istarget powerpc*-*-*]
 	      && [check_effective_target_has_arch_pwr10])
+	     || ([istarget loongarch*-*-*]
+		 && [check_effective_target_loongarch_sx])
              || [istarget amdgcn-*-*] }}]
 }
 
@@ -12656,6 +12658,22 @@  proc check_effective_target_const_volatile_readonly_section { } {
   return 1
 }
 
+proc check_effective_target_loongarch_sx { } {
+    return [check_no_compiler_messages loongarch_lsx assembly {
+       #if !defined(__loongarch_sx)
+       #error "LSX not defined"
+       #endif
+    }]
+}
+
+proc check_effective_target_loongarch_asx { } {
+    return [check_no_compiler_messages loongarch_asx assembly {
+       #if !defined(__loongarch_asx)
+       #error "LASX not defined"
+       #endif
+    }]
+}
+
 # Appends necessary Python flags to extra-tool-flags if Python.h is supported.
 # Otherwise, modifies dg-do-what.
 proc dg-require-python-h { args } {