| Message ID | 000001cfd258$57b8fd50$072af7f0$@arm.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Sep 17, 2014 at 05:18:27PM +0800, Zhenqiang Chen wrote: > Hi, > > When assign_hard_reg, we always check_hard_reg_p, which has code > > if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno)) > return false; > > i.e. If a hard_regno is not in profitable_regs, we can not allocate it to > the ira_allocno_t A. > > So the conflict on a hard_regno, which does not belong to the > profitable_regs, can be ignored. Please refer > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63210 for the detail of the > test case. > > Bootstrap and no make check regression on X86-64 and ARM Chromebook. > NO spec2k performance regression on X86-64 and ARM Chromebook. > CSiBE code size is 0.01% better for ARM Cortex-M0. Note that your MUA malformed the patch, so it can't be easily applied. Marek
> -----Original Message----- > From: Marek Polacek [mailto:polacek@redhat.com] > Sent: Wednesday, September 17, 2014 5:29 PM > To: Zhenqiang Chen > Cc: gcc-patches@gcc.gnu.org > Subject: Re: [PATCH, ira] Ignore some conflict cost > > On Wed, Sep 17, 2014 at 05:18:27PM +0800, Zhenqiang Chen wrote: > > Hi, > > > > When assign_hard_reg, we always check_hard_reg_p, which has code > > > > if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno)) > > return false; > > > > i.e. If a hard_regno is not in profitable_regs, we can not allocate it > > to the ira_allocno_t A. > > > > So the conflict on a hard_regno, which does not belong to the > > profitable_regs, can be ignored. Please refer > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63210 for the detail of > > the test case. > > > > Bootstrap and no make check regression on X86-64 and ARM Chromebook. > > NO spec2k performance regression on X86-64 and ARM Chromebook. > > CSiBE code size is 0.01% better for ARM Cortex-M0. > > Note that your MUA malformed the patch, so it can't be easily applied. Thanks for the comments. Resend it as an attachment. -Zhenqiang
On 09/17/14 03:18, Zhenqiang Chen wrote: > Hi, > > When assign_hard_reg, we always check_hard_reg_p, which has code > > if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno)) > return false; > > i.e. If a hard_regno is not in profitable_regs, we can not allocate it to > the ira_allocno_t A. > > So the conflict on a hard_regno, which does not belong to the > profitable_regs, can be ignored. Please refer > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63210 for the detail of the > test case. > > Bootstrap and no make check regression on X86-64 and ARM Chromebook. > NO spec2k performance regression on X86-64 and ARM Chromebook. > CSiBE code size is 0.01% better for ARM Cortex-M0. > > OK for trunk? > Thanks! > -Zhenqiang > > ChangeLog: > 2014-09-17 Zhenqiang Chen <zhenqiang.chen@arm.com> > > PR rtl-optimization/63210 > * ira-color.c (assign_hard_reg): Ignore conflict cost if the > HARD_REGNO is not availabe for CONFLICT_A. > > testsuite/ChangeLog: > 2014-09-17 Zhenqiang Chen <zhenqiang.chen@arm.com> > > * gcc.target/arm/pr63210.c: New test. OK for the trunk. Thanks, Jeff
diff --git a/gcc/ira-color.c b/gcc/ira-color.c index e2ea359..e3fea54 100644 --- a/gcc/ira-color.c +++ b/gcc/ira-color.c @@ -1709,6 +1709,7 @@ assign_hard_reg (ira_allocno_t a, bool retry_p) { ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); enum reg_class conflict_aclass; + allocno_color_data_t data = ALLOCNO_COLOR_DATA (conflict_a); /* Reload can give another class so we need to check all allocnos. */ @@ -1780,7 +1781,12 @@ assign_hard_reg (ira_allocno_t a, bool retry_p) hard_regno = ira_class_hard_regs[aclass][j]; ira_assert (hard_regno >= 0); k = ira_class_hard_reg_index[conflict_aclass][hard_regno]; - if (k < 0) + if (k < 0 + /* If HARD_REGNO is not available for CONFLICT_A, + the conflict would be ignored, since HARD_REGNO + will never be assigned to CONFLICT_A. */ + || !TEST_HARD_REG_BIT (data->profitable_hard_regs, + hard_regno)) continue; full_costs[j] -= conflict_costs[k]; } diff --git a/gcc/testsuite/gcc.target/arm/pr63210.c b/gcc/testsuite/gcc.target/arm/pr63210.c new file mode 100644 index 0000000..c3ae928 --- /dev/null +++ b/gcc/testsuite/gcc.target/arm/pr63210.c @@ -0,0 +1,12 @@ +/* { dg-do assemble } */ +/* { dg-options "-mthumb -Os " } */ +/* { dg-require-effective-target arm_thumb1_ok } */ + +int foo1 (int c); +int foo2 (int c); + +int test (int c) +{ + return (foo1 (c) || foo2 (c)); +} +/* { dg-final { object-size text <= 28 } } */
Hi, When assign_hard_reg, we always check_hard_reg_p, which has code if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno)) return false; i.e. If a hard_regno is not in profitable_regs, we can not allocate it to the ira_allocno_t A. So the conflict on a hard_regno, which does not belong to the profitable_regs, can be ignored. Please refer https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63210 for the detail of the test case. Bootstrap and no make check regression on X86-64 and ARM Chromebook. NO spec2k performance regression on X86-64 and ARM Chromebook. CSiBE code size is 0.01% better for ARM Cortex-M0. OK for trunk? Thanks! -Zhenqiang ChangeLog: 2014-09-17 Zhenqiang Chen <zhenqiang.chen@arm.com> PR rtl-optimization/63210 * ira-color.c (assign_hard_reg): Ignore conflict cost if the HARD_REGNO is not availabe for CONFLICT_A. testsuite/ChangeLog: 2014-09-17 Zhenqiang Chen <zhenqiang.chen@arm.com> * gcc.target/arm/pr63210.c: New test.