PowerPC TPREL_HA/LO optimisation
Alan Modra
amodra@gmail.com
Tue Nov 17 07:02:44 GMT 2020
On Thu, Nov 12, 2020 at 08:26:53PM +0300, Gleb Fotengauer-Malinovskiy wrote:
> Hi,
>
> On Mon, Aug 24, 2020 at 09:43:08PM +0930, Alan Modra via Binutils wrote:
> > ppc64 ld optimises sequences like the following
> > addis 3,13,wot@tprel@ha
> > lwz 3,wot@tprel@l(3)
> > to
> > nop
> > lwz 3,wot@tprel(13)
> > when "wot" is located near enough to the thread pointer.
> > However, the ABI doesn't require that R_PPC64_TPREL16_HA always be on
> > an addis rt,13,imm instruction, and while ld checked for that on the
> > high-part instruction it didn't disable the optimisation on the
> > low-part instruction. This patch fixes that problem, disabling the
> > tprel optimisation globally if high-part instructions don't pass
> > sanity checks. The optimisation is also enabled for ppc32, where
> > before ld.bfd had the code in the wrong place and ld.gold had it in a
> > block only enabled for ppc64.
>
> Looks like this commit broke incremental_test.sh test in gold on ppc64le.
> (I added abort(); to the end of do_gold_unreachable function to get this
> backtrace.)
Indeed. I'll be applying the following once some testsuite runs
finish.
Subject: [GOLD] powerpc incremental-dump assertion failure
incremental-dump wants to instantiate Target_powerpc without options
being set up. This patch fixes
internal error in options, at gold/parameters.h:92
* powerpc.cc (Target_powerpc::no_tprel_opt_): Rename from tprel_opt_.
Init to false.
(Target_powerpc::tprel_opt): Test parameters->options().tls_optimize().
(Target_powerpc::set_tprel_opt): Delete.
(Target_powerpc::set_no_tprel_opt): New function. Update all uses
of set_tprel_opt.
diff --git a/gold/powerpc.cc b/gold/powerpc.cc
index fcbe71fbd3..25b7dbd355 100644
--- a/gold/powerpc.cc
+++ b/gold/powerpc.cc
@@ -649,8 +649,7 @@ class Target_powerpc : public Sized_target<size, big_endian>
stub_tables_(), branch_lookup_table_(), branch_info_(), tocsave_loc_(),
power10_relocs_(false), plt_thread_safe_(false), plt_localentry0_(false),
plt_localentry0_init_(false), has_localentry0_(false),
- has_tls_get_addr_opt_(false),
- tprel_opt_(parameters->options().tls_optimize()),
+ has_tls_get_addr_opt_(false), no_tprel_opt_(false),
relax_failed_(false), relax_fail_count_(0),
stub_group_size_(0), savres_section_(0),
tls_get_addr_(NULL), tls_get_addr_opt_(NULL),
@@ -1154,11 +1153,11 @@ class Target_powerpc : public Sized_target<size, big_endian>
bool
tprel_opt() const
- { return this->tprel_opt_; }
+ { return !this->no_tprel_opt_ && parameters->options().tls_optimize(); }
void
- set_tprel_opt(bool val)
- { this->tprel_opt_ = val; }
+ set_no_tprel_opt()
+ { this->no_tprel_opt_ = true; }
// Remember any symbols seen with non-zero localentry, even those
// not providing a definition
@@ -1717,7 +1716,7 @@ class Target_powerpc : public Sized_target<size, big_endian>
bool plt_localentry0_init_;
bool has_localentry0_;
bool has_tls_get_addr_opt_;
- bool tprel_opt_;
+ bool no_tprel_opt_;
bool relax_failed_;
int relax_fail_count_;
@@ -8471,7 +8470,7 @@ Target_powerpc<size, big_endian>::Scan::local(
uint32_t insn = elfcpp::Swap<32, big_endian>::readval(view + off);
if ((insn & ((0x3fu << 26) | 0x1f << 16))
!= ((15u << 26) | ((size == 32 ? 2 : 13) << 16)))
- target->set_tprel_opt(false);
+ target->set_no_tprel_opt();
}
}
break;
@@ -8486,7 +8485,7 @@ Target_powerpc<size, big_endian>::Scan::local(
break;
// Fall through.
case elfcpp::R_POWERPC_TPREL16_HI:
- target->set_tprel_opt(false);
+ target->set_no_tprel_opt();
break;
default:
break;
@@ -9268,7 +9267,7 @@ Target_powerpc<size, big_endian>::Scan::global(
uint32_t insn = elfcpp::Swap<32, big_endian>::readval(view + off);
if ((insn & ((0x3fu << 26) | 0x1f << 16))
!= ((15u << 26) | ((size == 32 ? 2 : 13) << 16)))
- target->set_tprel_opt(false);
+ target->set_no_tprel_opt();
}
}
break;
@@ -9283,7 +9282,7 @@ Target_powerpc<size, big_endian>::Scan::global(
break;
// Fall through.
case elfcpp::R_POWERPC_TPREL16_HI:
- target->set_tprel_opt(false);
+ target->set_no_tprel_opt();
break;
default:
break;
--
Alan Modra
Australia Development Lab, IBM
More information about the Binutils
mailing list