| Message ID | 563B221D.2010201@arm.com |
|---|---|
| State | New |
| Headers | show |
Hi Ramana, On 10/11/15 14:33, Ramana Radhakrishnan wrote: > On Thu, Nov 5, 2015 at 9:32 AM, Kyrill Tkachov <kyrylo.tkachov@arm.com> wrote: >> Hi all, >> >> This cleanup patch removes handling of CONST_DOUBLE rtxes that carry large >> integers. >> These should never be passed down from the midend and the arm backend >> doesn't create them. >> The code has been there since 2007 but the arm backend was moved to >> TARGET_SUPPORTS_WIDE_INT >> in 2014, so this path should never be taken. >> >> Bootstrapped and tested on arm-none-linux-gnueabihf. >> >> Ok for trunk? > > This is OK - Thanks for reviewing. Sorry I had forgotten the ChangeLog in the initial submission. It is 2015-11-10 Kyrylo Tkachov <kyrylo.tkachov@arm.com> * config/arm/arm.c (neon_valid_immediate): Remove integer CONST_DOUBLE handling. It should never occur. I have committed the patch with that entry as r230115. Thanks, Kyrill > Ramana >> Thanks, >> Kyrill >>
commit 8a1c27b21e6662c3203b95c457c0df5c0239805e
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date: Thu Oct 29 11:20:43 2015 +0000
[ARM][cleanup] Remove uses of CONST_DOUBLE_HIGH/LOW
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 5c97786..36958af 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -12386,32 +12386,15 @@ neon_valid_immediate (rtx op, machine_mode mode, int inverse,
{
rtx el = vector ? CONST_VECTOR_ELT (op, i) : op;
unsigned HOST_WIDE_INT elpart;
- unsigned int part, parts;
- if (CONST_INT_P (el))
- {
- elpart = INTVAL (el);
- parts = 1;
- }
- else if (CONST_DOUBLE_P (el))
- {
- elpart = CONST_DOUBLE_LOW (el);
- parts = 2;
- }
- else
- gcc_unreachable ();
+ gcc_assert (CONST_INT_P (el));
+ elpart = INTVAL (el);
- for (part = 0; part < parts; part++)
- {
- unsigned int byte;
- for (byte = 0; byte < innersize; byte++)
- {
- bytes[idx++] = (elpart & 0xff) ^ invmask;
- elpart >>= BITS_PER_UNIT;
- }
- if (CONST_DOUBLE_P (el))
- elpart = CONST_DOUBLE_HIGH (el);
- }
+ for (unsigned int byte = 0; byte < innersize; byte++)
+ {
+ bytes[idx++] = (elpart & 0xff) ^ invmask;
+ elpart >>= BITS_PER_UNIT;
+ }
}
/* Sanity check. */