Hi!
As Peter remarked, the input_operand's in the disassemble patterns are
better as something more specific (input_operand has everything a "mov"
pattern can handle, but pretty much nothing else can on a load/store
architecture like Power -- it will likely still work, but only after
reloads, so n
From: Aaron Sawdey
Segher & Bergner -
Thanks for the reviews, here's the updated patch after fixing those things.
We now have an UNSPEC for xxsetaccz, and an accompanying change to
rs6000_rtx_costs to make it be cost 0 so that CSE doesn't try to replace it
with a bunch of register moves.
If bo
For some reason this patch never showed up on gcc-patches.
Aaron Sawdey, Ph.D. saw...@linux.ibm.com
IBM Linux on POWER Toolchain
> Begin forwarded message:
>
> From: acsaw...@linux.ibm.com
> Subject: [PATCH,rs6000] Make MMA builtins use opaque modes [v2]
> Date: November 19, 2
On 11/19/20 12:58 PM, acsaw...@linux.ibm.com wrote:
> +(define_expand "mma_disassemble_pair"
> + [(match_operand:V16QI 0 "mma_disassemble_output_operand")
> + (match_operand:OO 1 "input_operand")
> + (match_operand 2 "const_0_to_1_operand")]
Maybe we should use vsx_register_operand instead of
On 11/17/20 5:01 PM, Segher Boessenkool wrote:
> On Tue, Nov 17, 2020 at 12:41:58PM -0600, Peter Bergner wrote:
>>> +;; Return 1 if this operand is valid for an MMA disassemble insn.
>>> +(define_predicate "mma_disassemble_output_operand"
>>> + (match_code "reg,subreg,mem")
>>> +{
>>> + if (REG_P
On Tue, Nov 17, 2020 at 12:41:58PM -0600, Peter Bergner wrote:
> > +;; Return 1 if this operand is valid for an MMA disassemble insn.
> > +(define_predicate "mma_disassemble_output_operand"
> > + (match_code "reg,subreg,mem")
> > +{
> > + if (REG_P (op) && !vsx_register_operand (op, mode))
> > +
Hi!
On Tue, Nov 17, 2020 at 11:48:04AM -0600, acsaw...@linux.ibm.com wrote:
> This patch changes powerpc MMA builtins to use the new opaque
> mode class and use modes OO (32 bytes) and XO (64 bytes)
> instead of POI/PXI. Using the opaque modes prevents
> optimization from trying to do anything wit
On 11/17/20 11:48 AM, acsaw...@linux.ibm.com wrote:
> -;; The MMA patterns use the multi-register PXImode and POImode partial
> +;; The MMA patterns use the multi-register XOmode and OOmode partial
> ;; integer modes to implement the target specific __vector_quad and
XOmode and OOmode are not par
From: Aaron Sawdey
This patch changes powerpc MMA builtins to use the new opaque
mode class and use modes OO (32 bytes) and XO (64 bytes)
instead of POI/PXI. Using the opaque modes prevents
optimization from trying to do anything with vector
pair/quad, which was the problem we were seeing with th
