[PATCH v6 33/34] Drop single-precision Thumb-1 soft-float functions
Daniel Engel
gnu@danielengel.com
Mon Dec 27 19:05:29 GMT 2021
With the complete CM0 library integrated, regression testing showed new
failures with the message "compilation failed to produce executable":
gcc.dg/fixed-point/convert-float-1.c
gcc.dg/fixed-point/convert-float-3.c
gcc.dg/fixed-point/convert-sat.c
Investigating, this appears to be caused by the linker. I can't find a
comprehensive linker specification to claim this is actually a bug, but it
certainly doesn't match my expectations. Investigating, I found issues
with the link order of these symbols:
* __aeabi_fmul()
* __aeabi_f2d()
* __aeabi_f2iz()
Specifically, I expect the linker to import the _first_ definition of any
symbol. This is the basic behavior that allows the soft-float library to
supply missing symbols on architectures without optimized routines.
Comparing the v6-m multilib with the default, I see symbol exports for all
of the affect symbols:
gcc-obj/gcc/libgcc.a:
// assembly routines
_arm_mulsf3.o:
00000000 W __aeabi_fmul
00000000 W __mulsf3
_arm_addsubdf3.o:
00000368 T __aeabi_f2d
00000368 T __extendsfdf2
_arm_fixsfsi.o:
00000000 T __aeabi_f2iz
00000000 T __fixsfsi
mulsf3.o:
<empty>
fixsfsi.o:
<empty>
extendsfdf2.o.o:
<empty>
gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a:
// assembly routines
_arm_mulsf3.o:
00000000 T __aeabi_fmul
U __fp_assemble
U __fp_exception
U __fp_infinity
U __fp_zero
00000000 T __mulsf3
U __umulsidi3
_arm_fixsfsi.o:
00000000 T __aeabi_f2iz
00000000 T __fixsfsi
00000002 T __internal_f2iz
_arm_f2d.o:
00000000 T __aeabi_f2d
00000000 T __extendsfdf2
U __fp_normalize2
// soft-float library
mulsf3.o:
00000000 T __aeabi_fmul
fixsfsi.o:
00000000 T __aeabi_f2iz
extendsfdf2.o:
00000000 T __aeabi_f2d
Given the order of the archive file, I expect the linker to import the affected
functions from the _arm_* archive elements.
For "convert-sat.c", all is well with -march=armv7-m.
...
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_muldf3.o
OK> (/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_mulsf3.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_cmpsf2.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_fixsfsi.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_fixunssfsi.o
OK> (/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_addsubdf3.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_cmpdf2.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_fixdfsi.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_arm_fixunsdfsi.o
OK> (/home/mirdan/gcc-obj/gcc/libgcc.a)_fixsfdi.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_fixdfdi.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_fixunssfdi.o
(/home/mirdan/gcc-obj/gcc/libgcc.a)_fixunsdfdi.o
...
However, with -march=armv6s-m, the linker imports these symbols from the soft-
float library. (NOTE: The CM0 library only implements single-precision float
operations, so imports from muldf3.o, fixdfsi.o, etc are expected.)
...
??> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)mulsf3.o
??> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)fixsfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)muldf3.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)fixdfsi.o
??> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)extendsfdf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_clzsi2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fcmpge.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fcmple.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixsfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunssfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunssfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_cmpdf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunsdfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixdfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunsdfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)eqdf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)gedf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)ledf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)subdf3.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)floatunsidf.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_cmpsf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixsfsi.o
...
It seems that the order in which the linker resolves symbols matters. In the
affected test cases, the linker begins searching for fixed-point function
symbols first: _subQQ.o, _cmpQQ.o, etc. The fixed-point archive elements
appear after the _arm_* archive elements, so the initial definitions of the
floating point functions are discarded. However, the fixed-point functions
contain unresolved symbol references which the linker registers progressively.
Given that the default libgcc.a does not build the soft-point library [1],
the linker cannot import any floating point objects until the second pass.
However, when v6-m/nofp/libgcc.a _does_ include the soft-point library, the
linker proceeds to import some floating point objects during the first pass.
To test this theory, add explicit symbol references to convert-sat.c:
--- a/gcc/testsuite/gcc.dg/fixed-point/convert-sat.c
+++ b/gcc/testsuite/gcc.dg/fixed-point/convert-sat.c
@@ -11,6 +11,12 @@ extern void abort (void);
int main ()
{
+ volatile float a = 1.0;
+ volatile float b = 2.0;
+ volatile float c = a * b;
+ volatile double d = a;
+ volatile int e = a;
+
SAT_CONV1 (short _Accum, hk);
SAT_CONV1 (_Accum, k);
SAT_CONV1 (long _Accum, lk);
Afterwards, the linker imports the expected symbols:
...
==> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_mulsf3.o
==> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_muldi3.o
==> (/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fixsfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_f2d.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fp_exceptionf.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fp_assemblef.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fp_normalizef.o
...
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)muldf3.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)fixdfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_clzsi2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fixunssfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fcmpge.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fcmple.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fixsfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_fixunssfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_cmpdf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunsdfsi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixdfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_fixunsdfdi.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)eqdf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)gedf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)ledf2.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)subdf3.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)floatunsidf.o
(/home/mirdan/gcc-obj/gcc/thumb/v6-m/nofp/libgcc.a)_arm_cmpsf2.o
...
At a minimum this behavior results in the use of non-preferred code in an
affected application. However, as long as each object exports a single
entry point, this does not automatically result in a build failure.
Indeed, in the case of __aeabi_fmul() and __aeabi_f2d(), all references seem
to resolve uniformly in favor of the soft-float library. The first pass that
imports the soft-float version of __aeabi_f2iz() also succeeds.
However, the first pass fails to find __aeabi_f2uiz(), since the soft-float
library does not implement this variant. So, this symbol remains undefined
until the second pass. However, the assembly version of __aeabi_f2uiz()
the linker finds happens to be implemented as a branch to __internal_f2iz() [2].
But the linker, importing __internal_f2iz(), also finds the main entry point
__aeabi_f2iz(). And, since __aeabi_f2iz() was already found in the soft-float
library, the linker throws an error.
The solution is two-fold. First, the assembly routines have separately been
made robust against this potential error condition (by weakening and splitting
symbols). Second, this commit to block single-precision functions from the
soft-float library makes it impossible for the linker to select a non-preferred
version. Two duplicate symbols remain (extendsfdf2) and (truncdfsf2), but the
situation is much improved.
[1] softfp_wrap_start = "#if !__ARM_ARCH_ISA_ARM && __ARM_ARCH_ISA_THUMB == 1"
[2] (These operations share a substantial portion of their code path, so this
choice leads to a size reduction in programs that use both functions.)
gcc/libgcc/ChangeLog:
2021-01-13 Daniel Engel <gnu@danielengel.com>
* config/arm/t-softfp (softfp_float_modes): Added as "df".
---
libgcc/config/arm/t-softfp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/libgcc/config/arm/t-softfp b/libgcc/config/arm/t-softfp
index 554ec9bc47b..bd6a4642e5f 100644
--- a/libgcc/config/arm/t-softfp
+++ b/libgcc/config/arm/t-softfp
@@ -1,2 +1,4 @@
softfp_wrap_start := '\#if !__ARM_ARCH_ISA_ARM && __ARM_ARCH_ISA_THUMB == 1'
softfp_wrap_end := '\#endif'
+softfp_float_modes := df
+
--
2.25.1
More information about the Gcc-patches
mailing list