| Message ID | CAOvf_xym6ajUDA2caCCtKr1=vVzOYZwaiV-0UO=Xyxhti2bDYg@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Oct 22, 2014 at 07:55:57PM +0400, Evgeny Stupachenko wrote: > There is a test for Linux x86 that also fails without the changes in the patch: > > ChangeLog: > > 2014-10-22 Evgeny Stupachenko <evstupac@gmail.com> > > PR rtl-optimization/63618 > * gcc.target/i386/pr63618.c: New. > > diff --git a/gcc/testsuite/gcc.target/i386/pr63618.c > b/gcc/testsuite/gcc.target/i386/pr63618.c > new file mode 100644 > index 0000000..cf10a2b > --- /dev/null > +++ b/gcc/testsuite/gcc.target/i386/pr63618.c > @@ -0,0 +1,34 @@ > +/* PR rtl-optimization/63618 */ > +/* { dg-do run } */ > +/* { dg-require-effective-target sse4 } */ > +/* { dg-require-effective-target ia32 } */ > +/* { dg-options "-O2 -msse4.2 -mtune=corei7 -fno-inline -fPIC" } */ What is ia32 specific on the test? As for -msse4.2 and sse4 effective target, I'd say better would be to stick that to /* { dg-additional-options "-msse4.2" { target sse4 } } */ Similarly, you want -fPIC only if { target fpic }, thus I'd say you want to start the test with: /* { dg-do run } */ /* { dg-options "-O2 -mtune=corei7 -fno-inline" } */ /* { dg-additional-options "-msse4.2" { target sse4 } } */ /* { dg-additional-options "-fpic" { target fpic } } */ Jakub
diff --git a/gcc/testsuite/gcc.target/i386/pr63618.c b/gcc/testsuite/gcc.target/i386/pr63618.c new file mode 100644 index 0000000..cf10a2b --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr63618.c @@ -0,0 +1,34 @@ +/* PR rtl-optimization/63618 */ +/* { dg-do run } */ +/* { dg-require-effective-target sse4 } */ +/* { dg-require-effective-target ia32 } */ +/* { dg-options "-O2 -msse4.2 -mtune=corei7 -fno-inline -fPIC" } */ + +static const __float128 cf = 0.1E+10Q; + +__float128 +f128_square(__float128 x) +{ + return x * x; +} + +__float128 +f128_p3(__float128 x) +{ + return x * x * x; +} + +__float128 +cond_f128_square (__float128 x, int p) +{ + x = f128_p3 (x); + if (p) + x = f128_square(cf); + return x; +} + +int main() +{ + __float128 x = cond_f128_square (cf, 1); + return (int)(x < cf); +}
There is a test for Linux x86 that also fails without the changes in the patch: ChangeLog: 2014-10-22 Evgeny Stupachenko <evstupac@gmail.com> PR rtl-optimization/63618 * gcc.target/i386/pr63618.c: New. On Fri, Oct 17, 2014 at 6:16 PM, Evgeny Stupachenko <evstupac@gmail.com> wrote: > Hi, > > Some instructions (like one in PR63534) could have hidden use of PIC register. > Therefore we need to leave SET_GOT not deleted till reload completed. > The patch prevents SET_GOT from deleting while PIC register is pseudo. > > Is it ok? > > ChangeLog: > > 2014-10-17 Evgeny Stupachenko <evstupac@gmail.com> > > PR target/63534 > * cse.c (delete_trivially_dead_insns): Consider PIC register is used > while it is pseudo. > * dse.c (deletable_insn_p): Likewise. > > diff --git a/gcc/cse.c b/gcc/cse.c > index be2f31b..062ba45 100644 > --- a/gcc/cse.c > +++ b/gcc/cse.c > @@ -6953,6 +6953,11 @@ delete_trivially_dead_insns (rtx_insn *insns, int nreg) > /* If no debug insns can be present, COUNTS is just an array > which counts how many times each pseudo is used. */ > } > + /* Pseudo PIC register should be considered as used due to possible > + new usages generated. */ > + if (pic_offset_table_rtx > + && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER) > + counts[REGNO (pic_offset_table_rtx)]++; > /* Go from the last insn to the first and delete insns that only set unused > registers or copy a register to itself. As we delete an insn, remove > usage counts for registers it uses. > diff --git a/gcc/dce.c b/gcc/dce.c > index 5b7d36e..a52a59c 100644 > --- a/gcc/dce.c > +++ b/gcc/dce.c > @@ -127,6 +127,10 @@ deletable_insn_p (rtx_insn *insn, bool fast, > bitmap arg_stores) > if (HARD_REGISTER_NUM_P (DF_REF_REGNO (def)) > && global_regs[DF_REF_REGNO (def)]) > return false; > + /* Initialization of pseudo PIC register should never be removed. */ > + else if (DF_REF_REG (def) == pic_offset_table_rtx > + && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER) > + return false; > > body = PATTERN (insn); > switch (GET_CODE (body))