On 1/29/2022 9:23 AM, Jakub Jelinek wrote:
Hi!
On Fri, Jan 28, 2022 at 11:38:23AM -0700, Jeff Law wrote:
Thanks. Given the original submission and most of the review work was done
prior to stage3 closing, I went ahead and installed this on the trunk.
Unfortunately this breaks quite a lot of things.
The main problem is that GIMPLE allows EQ_EXPR etc. only with BOOLEAN_TYPE
or with TYPE_PRECISION == 1 integral type (or vector boolean).
Violating this causes verification failures in tree-cfg.cc in some cases,
in other cases wrong-code issues because before it is verified we e.g.
transform
1U / x
into
x == 1U
and later into
x (because we assume that == type must be one of the above cases and
when it is the same type as the type of the first operand, for boolean-ish
cases it should be equivalent).
Fixed by changing that
(eq @1 { build_one_cst (type); })
into
(convert (eq:boolean_type_node @1 { build_one_cst (type); }))
Note, I'm not 100% sure if :boolean_type_node is required in that case,
I see some spots in match.pd that look exactly like this, while there is
e.g. (convert (le ...)) that supposedly does the right thing too.
The signed integer 1/X case doesn't need changes changes, for
(cond (le ...) ...)
le gets correctly boolean_type_node and cond should use type.
I've also reformatted it, some lines were too long, match.pd uses
indentation by 1 column instead of 2 etc.
Bootstrapped/regtested on powerpc64le-linux and tested on the testcases
on x86_64-linux, ok for trunk?
2022-01-29 Jakub Jelinek <ja...@redhat.com>
Andrew Pinski <apin...@marvell.com>
PR tree-optimization/104279
PR tree-optimization/104280
PR tree-optimization/104281
* match.pd (1 / X -> X == 1 for unsigned X): Build eq with
boolean_type_node and convert to type. Formatting fixes.
* gcc.dg/torture/pr104279.c: New test.
* gcc.dg/torture/pr104280.c: New test.
* gcc.dg/torture/pr104281.c: New test.
Ugh. Sorry for the breakage. THat makes me wonder how the original
patch was tested at all.
I think both forms will DTRT (eq:boolean_type_node ...) vs (convert (le
...)) .
OK for the trunk as well as the testcase fix.
Jeff
