https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115673
Peter Bergner <bergner at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |vmakarov at gcc dot gnu.org
--- Comment #13 from Peter Bergner <bergner at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #12)
> Re-confirmed.
> +FAIL: gcc.target/i386/pr91384.c scan-assembler-not testl
int
test (int a)
{
int r;
if (r = -a)
foo ();
else
bar ();
return r;
}
Looking at just this test case, the first (x86_64) RTL difference between trunk
and trunk with Surya's patch reverted, I see:
Hard reg set forest:
0:( 0-6 8-15 20-51)@0
1:( 0-6 36-43)@36000
2:( 0 3 6 40-43)@18000
Allocno a0r98 of GENERAL_REGS(15) has 7 avail. regs 0 3 6 40-43, node:
0 3 6 40-43 (confl regs = 7-35 44-91)
Allocno a1r101 of GENERAL_REGS(15) has 15 avail. regs 0-6 36-43, node:
0-6 36-43 (confl regs = 7-35 44-91)
Forming thread from colorable bucket:
Pushing a0(r98,l0)(cost 0)
Pushing a1(r101,l0)(cost 0)
Popping a1(r101,l0) -- assign reg 5
- Popping a0(r98,l0) -- assign reg 0
+ Popping a0(r98,l0) -- assign reg 3
Disposition:
- 0:r98 l0 0 1:r101 l0 5
+ 0:r98 l0 3 1:r101 l0 5
New iteration of spill/restore move
-+++Costs: overall 14000, reg 14000, mem 0, ld 0, st 0, move 0
++++Costs: overall 5000, reg 5000, mem 0, ld 0, st 0, move 0
+++ move loops 0, new jumps 0
So with Surya's patch, we assign hardreg 0 (%rax and volatile) to variable "r"
versus hardreg 3 (%rbx and non-volatile). I think the increase of the
allocation cost of using the non-volatile hardreg 3 by Surya's patch is
justified. However, I'm wondering whether we should also be increasing the
allocation cost of using a volatile hardreg for pseudos that are live across
calls, since those too will need to be saved/restored.
