OpenE2K
/
gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

SVE: Fix wrong sve predicate split (PR100048) 

The attached testcase generates the following paradoxical subregs when creating
the predicates.

(insn 22 21 23 2 (set (reg:VNx8BI 100)
	(subreg:VNx8BI (reg:VNx2BI 103) 0))
     (expr_list:REG_EQUAL (const_vector:VNx8BI [
		(const_int 1 [0x1])
		(const_int 0 [0])
		(const_int 1 [0x1])
	(const_int 0 [0]) repeated x5
	    ])
	(nil)))

and

(insn 15 14 16 2 (set (reg:VNx8BI 96)
	(subreg:VNx8BI (reg:VNx2BI 99) 0))
     (expr_list:REG_EQUAL (const_vector:VNx8BI [
		(const_int 1 [0x1])
		(const_int 0 [0]) repeated x7
	    ])
	(nil)))

This causes CSE to incorrectly think that the two predicates are equal because
some of the significant bits get ignored due to the subreg.

The attached patch instead makes it so it always looks at all 16-bits of the
predicate, but in turn means we need to generate a TRN that matches the expected
result mode.  In effect in RTL we keep the mode as VNx16BI but during codegen
re-interpret them as the mode the predicate instruction wanted:

(insn 10 9 11 2 (set (reg:VNx8BI 96)
	(subreg:VNx8BI (reg:VNx16BI 99) 0))
     (expr_list:REG_EQUAL (const_vector:VNx8BI [
		(const_int 1 [0x1])
		(const_int 0 [0]) repeated x7
	    ])
	(nil)))

Which needed correction to the TRN pattern.  A new TRN1_CONV unspec is
introduced which allows one to keep the arguments as VNx16BI but encode the
instruction as a type of the last operand.

(insn 9 8 10 2 (set (reg:VNx16BI 99)
	(unspec:VNx16BI [
		(reg:VNx16BI 97)
		(reg:VNx16BI 98)
		(reg:VNx2BI 100)
	    ] UNSPEC_TRN1_CONV))
	(nil))

This allows us remove all the paradoxical subregs and end up with

(insn 16 15 17 2 (set (reg:VNx8BI 101)
	(subreg:VNx8BI (reg:VNx16BI 104) 0))
	(expr_list:REG_EQUAL (const_vector:VNx8BI [
		(const_int 1 [0x1])
		(const_int 0 [0])
		(const_int 1 [0x1])
		(const_int 0 [0]) repeated x5
	    ])
	(nil)))

gcc/ChangeLog:

	PR target/100048
	* config/aarch64/aarch64-sve.md (@aarch64_sve_trn1_conv<mode>): New.
	* config/aarch64/aarch64.c (aarch64_expand_sve_const_pred_trn): Use new
	TRN optab.
	* config/aarch64/iterators.md (UNSPEC_TRN1_CONV): New.

gcc/testsuite/ChangeLog:

	PR target/100048
	* gcc.target/aarch64/sve/pr100048.c: New test.
This commit is contained in:
Tamar Christina 2021-04-16 16:58:50 +01:00
parent 35e8b38a91
commit 8535755af7
4 changed files with 45 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
gcc/config/aarch64/aarch64-sve.md
View File

@ -8657,6 +8657,20 @@
"<perm_insn>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
)
;; Special purpose permute used by the predicate generation instructions.
;; Unlike the normal permute patterns, these instructions operate on VNx16BI
;; regardless of the element size, so that all input and output bits are
;; well-defined. Operand 3 then indicates the size of the permute.
(define_insn "@aarch64_sve_trn1_conv<mode>"
[(set (match_operand:VNx16BI 0 "register_operand" "=Upa")
(unspec:VNx16BI [(match_operand:VNx16BI 1 "register_operand" "Upa")
(match_operand:VNx16BI 2 "register_operand" "Upa")
(match_operand:PRED_ALL 3 "aarch64_simd_imm_zero")]
UNSPEC_TRN1_CONV))]
"TARGET_SVE"
"trn1\t%0.<PRED_ALL:Vetype>, %1.<PRED_ALL:Vetype>, %2.<PRED_ALL:Vetype>"
)
;; =========================================================================
;; == Conversions
;; =========================================================================

10
gcc/config/aarch64/aarch64.c
View File

@ -5535,12 +5535,12 @@ aarch64_expand_sve_const_pred_trn (rtx target, rtx_vector_builder &builder,
}
}
/* Emit the TRN1 itself. */
/* Emit the TRN1 itself. We emit a TRN that operates on VNx16BI
operands but permutes them as though they had mode MODE. */
machine_mode mode = aarch64_sve_pred_mode (permute_size).require ();
target = aarch64_target_reg (target, mode);
emit_insn (gen_aarch64_sve (UNSPEC_TRN1, mode, target,
gen_lowpart (mode, a),
gen_lowpart (mode, b)));
target = aarch64_target_reg (target, GET_MODE (a));
rtx type_reg = CONST0_RTX (mode);
emit_insn (gen_aarch64_sve_trn1_conv (mode, target, a, b, type_reg));
return target;
}

1
gcc/config/aarch64/iterators.md
View File

@ -649,6 +649,7 @@
UNSPEC_UZP2Q ; Used in aarch64-sve.md.
UNSPEC_ZIP1Q ; Used in aarch64-sve.md.
UNSPEC_ZIP2Q ; Used in aarch64-sve.md.
UNSPEC_TRN1_CONV ; Used in aarch64-sve.md.
UNSPEC_COND_CMPEQ_WIDE ; Used in aarch64-sve.md.
UNSPEC_COND_CMPGE_WIDE ; Used in aarch64-sve.md.
UNSPEC_COND_CMPGT_WIDE ; Used in aarch64-sve.md.

25
gcc/testsuite/gcc.target/aarch64/sve/pr100048.c Normal file
View File

@ -0,0 +1,25 @@
/* { dg-additional-options "-O2 -fno-schedule-insns" } */
/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
#include "arm_sve.h"
/*
** foo:
** ptrue (p[0-7])\.d, all
** pfalse (p[0-7])\.b
** ptrue (p[0-7])\.s, all
** trn1 (p[0-7])\.d, \2\.d, \3\.d
** trn1 \2\.d, \1\.d, \3\.d
** faddv (h[0-31]), \4\, (z[0-31]).h
** faddv (h[0-31]), \2\, \6\.h
** str \5, [x0]
** str \7, [x0, 2]
** ret
*/
void foo(svfloat16_t in, float16_t *dst) {
const svbool_t pg_q0 = svdupq_n_b16(1, 0, 1, 0, 0, 0, 0, 0);
const svbool_t pg_f0 = svdupq_n_b16(1, 0, 0, 0, 0, 0, 0, 0);
dst[0] = svaddv_f16(pg_f0, in);
dst[1] = svaddv_f16(pg_q0, in);
}