Re: [patch] more simplify_subreg SIMD bugs

[Aldy Hernandez <aldyh at redhat dot com>]

> I don't think I know what's going on here.  What's the test case?

Testcase attached.  powerpc-eabispe.

The skinny of the problem is:

(insn 37 34 38 0x0 (set (reg:V2SF 141)
        (subreg:V2SF (reg/v:V2SI 134 [ a ]) 0)) -1 (nil)

CSE tries to substitute the original constant vector into insn 37 ala:

(insn 37 blah (set (reg:V2SF 141)
	(subreg:V2SF (const_vector:V2SI [
        		    (const_int 0 [0x0])
        			(const_int 2147483647 [0x7fffffff])

...and if you see the logic in simplify_subreg(), you'll see insn 37 
eventually being turned into:

(insn 37 7 38 0 0x1230dc0 (set (reg:V2SF 141 [ a ])
        (const_vector:V2SF [
                (const_double:SF 0.0 [0x0.0p+0])
                (const_double:SF 0.0 [0x0.0p+0])
            ])) 983 {*movv2sf_internal} (nil)

...because we drop down to...

  /* Attempt to simplify constant to non-SUBREG expression.  */
  if (CONSTANT_P (op))

...where poop rolls down hill.

I hope this makes the patch obviously OK ;-).
Aldy

Attachment: f.c
Description: Text document