This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Fix up arch= handling in x86 target attribute (PR target/90867)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Uros Bizjak <ubizjak at gmail dot com>, hjl dot tools at gmail dot com
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Wed, 20 Nov 2019 00:53:47 +0100
- Subject: [PATCH] Fix up arch= handling in x86 target attribute (PR target/90867)
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
The arch= handling in target attribute right now clears almost all isa_flags
and all isa_flags2, so that later call to ix86_option_override_internal
can set just the isa options for the specific arch and nothing else.
Unfortunately, it doesn't work, because next to the ix86_isa_flags{,2}
we have also ix86_isa_flags{,2}_explicit bitmask and in
ix86_option_override_internal and say for arch=x86_64 will not try to
set sse2 isa when we cleared it in ix86_isa_flags but kept it set in
ix86_isa_flags_explicit.
So, the testcase works fine with -O2, but doesn't work with -O2 -msse2,
in the former case ix86_isa_flags_explicit doesn't have MASK_SSE2 bit set,
but in the latter case it does, so in the former case we end up with
MASK_SSE2 in ix86_isa_flags, in the latter not in the function with target
attribute.
The following patch thus clears both ix86_isa_flags{,2} and corresponding
ix86_isa_flags{,2}_explicit. Also, so that say
target ("arch=x86_64", "avx") works, the clearing is done when actually
seeing the arch=, not at the end. target ("avx", "arch=x86_64") will still
not enable avx, like before, though.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2019-11-19 Jakub Jelinek <jakub@redhat.com>
PR target/90867
* config/i386/i386-options.c (ix86_valid_target_attribute_tree): Don't
clear opts->x_ix86_isa_flags{,2} here...
(ix86_valid_target_attribute_inner_p): ... but here when seeing
arch=. Also clear opts->x_ix86_isa_flags{,2}_explicit.
* gcc.target/i386/pr90867.c: New test.
--- gcc/config/i386/i386-options.c.jj 2019-11-18 12:07:54.672405129 +0100
+++ gcc/config/i386/i386-options.c 2019-11-19 18:43:36.426606458 +0100
@@ -1147,7 +1147,25 @@ ix86_valid_target_attribute_inner_p (tre
ret = false;
}
else
- p_strings[opt] = xstrdup (p + opt_len);
+ {
+ p_strings[opt] = xstrdup (p + opt_len);
+ if (opt == IX86_FUNCTION_SPECIFIC_ARCH)
+ {
+ /* If arch= is set, clear all bits in x_ix86_isa_flags,
+ except for ISA_64BIT, ABI_64, ABI_X32, and CODE16
+ and all bits in x_ix86_isa_flags2. */
+ opts->x_ix86_isa_flags &= (OPTION_MASK_ISA_64BIT
+ | OPTION_MASK_ABI_64
+ | OPTION_MASK_ABI_X32
+ | OPTION_MASK_CODE16);
+ opts->x_ix86_isa_flags_explicit &= (OPTION_MASK_ISA_64BIT
+ | OPTION_MASK_ABI_64
+ | OPTION_MASK_ABI_X32
+ | OPTION_MASK_CODE16);
+ opts->x_ix86_isa_flags2 = 0;
+ opts->x_ix86_isa_flags2_explicit = 0;
+ }
+ }
}
else if (type == ix86_opt_enum)
@@ -1225,18 +1243,8 @@ ix86_valid_target_attribute_tree (tree f
/* If we are using the default tune= or arch=, undo the string assigned,
and use the default. */
if (option_strings[IX86_FUNCTION_SPECIFIC_ARCH])
- {
- opts->x_ix86_arch_string
- = ggc_strdup (option_strings[IX86_FUNCTION_SPECIFIC_ARCH]);
-
- /* If arch= is set, clear all bits in x_ix86_isa_flags,
- except for ISA_64BIT, ABI_64, ABI_X32, and CODE16. */
- opts->x_ix86_isa_flags &= (OPTION_MASK_ISA_64BIT
- | OPTION_MASK_ABI_64
- | OPTION_MASK_ABI_X32
- | OPTION_MASK_CODE16);
- opts->x_ix86_isa_flags2 = 0;
- }
+ opts->x_ix86_arch_string
+ = ggc_strdup (option_strings[IX86_FUNCTION_SPECIFIC_ARCH]);
else if (!orig_arch_specified)
opts->x_ix86_arch_string = NULL;
--- gcc/testsuite/gcc.target/i386/pr90867.c.jj 2019-11-19 18:49:11.746592189 +0100
+++ gcc/testsuite/gcc.target/i386/pr90867.c 2019-11-19 18:48:17.271406789 +0100
@@ -0,0 +1,30 @@
+/* PR target/90867 */
+/* { dg-do run { target lp64 } } */
+/* { dg-options "-O2 -msse2" } */
+
+unsigned long long freq = 3600000000UL; /* 3.6 GHz = 3600.0 MHz */
+
+__attribute__((noipa)) void
+bar (double x)
+{
+ static double d = 3600000000.0;
+ if (x != d)
+ __builtin_abort ();
+ d /= 1000.0;
+}
+
+__attribute__ ((target ("arch=x86-64"))) int
+foo ()
+{
+ bar ((double) freq);
+ bar (1e-3 * freq);
+ bar (1e-6 * freq);
+ bar (1e-9 * freq);
+ return 0;
+}
+
+int
+main ()
+{
+ return foo ();
+}
Jakub