https://gcc.gnu.org/g:2c9a83413db78aba0b788b199cbf94d69f71f6bd
commit 2c9a83413db78aba0b788b199cbf94d69f71f6bd
Author: Michael Meissner <meiss...@linux.ibm.com>
Date: Fri Apr 12 14:37:52 2024 -0400
Simplify converting between SImode and SF/DFmode.
2024-04-12 Michael Meissner <meiss...@linux.ibm.com>
gcc/
PR target/90822
* gcc/config/rs6000.md (uns code attribute): Add sign_extend and
zero_extend.
(floatsidf2): If SImode can live in the floating point registers,
directly issue a conversion to DImode and do the floating point
conversion during expansion instead of using an UNSPEC.
(floatunssisf2): Likewise.
(floatunssidf2): Likewise.
(floatsisf2): Likewise.
* gcc/config/rs6000/vsx.md (vsx_extract_si_<uns>float_df): Adjust
combiner to expect sign/zero extend of SImode to DImode in
optimizing
V4SI extracts.
Diff:
---
gcc/config/rs6000/rs6000.md | 48 ++++++++++++++++++++++++++++++++++++++++-----
gcc/config/rs6000/vsx.md | 18 +++++++++--------
2 files changed, 53 insertions(+), 13 deletions(-)
diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index f1f120199f3..4ff80c34cb1 100644
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -667,7 +667,9 @@
(define_code_attr uns [(fix "")
(unsigned_fix "uns")
(float "")
- (unsigned_float "uns")])
+ (unsigned_float "uns")
+ (sign_extend "")
+ (zero_extend "uns")])
; Various instructions that come in SI and DI forms.
; A generic w/d attribute, for things like cmpw/cmpd.
@@ -6041,6 +6043,10 @@
; with a '#' template, and a define_split (with C code). The idea is
; to allow constant folding with the template of the define_insn,
; then to have the insns split later (between sched1 and final).
+;
+; If we have direct support for SImode in floating point registers, just
+; convert the SImode value to DImode. If we are loading the value from memory,
+; we will use the LFIWAX/LXSIWAX
(define_expand "floatsidf2"
[(parallel [(set (match_operand:DF 0 "gpc_reg_operand")
@@ -6054,7 +6060,13 @@
{
if (TARGET_LFIWAX && TARGET_FCFID)
{
- emit_insn (gen_floatsidf2_lfiwax (operands[0], operands[1]));
+ if (TARGET_POWERPC64 && TARGET_P8_VECTOR)
+ {
+ rtx di_tmp = convert_to_mode (DImode, operands[1], false);
+ emit_insn (gen_floatdidf2 (operands[0], di_tmp));
+ }
+ else
+ emit_insn (gen_floatsidf2_lfiwax (operands[0], operands[1]));
DONE;
}
else if (TARGET_FCFID)
@@ -6110,6 +6122,10 @@
;; If we don't have a direct conversion to single precision, don't enable this
;; conversion for 32-bit without fast math, because we don't have the insn to
;; generate the fixup swizzle to avoid double rounding problems.
+;
+; If we have direct support for SImode in floating point registers, just
+; convert the SImode value to DImode. If we are loading the value from memory,
+; we will use the LFIWAX/LXSIWAX
(define_expand "floatunssisf2"
[(set (match_operand:SF 0 "gpc_reg_operand")
(unsigned_float:SF (match_operand:SI 1 "nonimmediate_operand")))]
@@ -6120,7 +6136,13 @@
{
if (TARGET_LFIWZX && TARGET_FCFIDUS)
{
- emit_insn (gen_floatunssisf2_lfiwzx (operands[0], operands[1]));
+ if (TARGET_POWERPC64 && TARGET_P8_VECTOR)
+ {
+ rtx di_tmp = convert_to_mode (DImode, operands[1], true);
+ emit_insn (gen_floatdisf2 (operands[0], di_tmp));
+ }
+ else
+ emit_insn (gen_floatunssisf2_lfiwzx (operands[0], operands[1]));
DONE;
}
else
@@ -6145,7 +6167,13 @@
{
if (TARGET_LFIWZX && TARGET_FCFID)
{
- emit_insn (gen_floatunssidf2_lfiwzx (operands[0], operands[1]));
+ if (TARGET_POWERPC64 && TARGET_P8_VECTOR)
+ {
+ rtx di_tmp = convert_to_mode (DImode, operands[1], true);
+ emit_insn (gen_floatdidf2 (operands[0], di_tmp));
+ }
+ else
+ emit_insn (gen_floatunssidf2_lfiwzx (operands[0], operands[1]));
DONE;
}
else if (TARGET_FCFID)
@@ -6905,6 +6933,10 @@
;; If we don't have a direct conversion to single precision, don't enable this
;; conversion for 32-bit without fast math, because we don't have the insn to
;; generate the fixup swizzle to avoid double rounding problems.
+;
+; If we have direct support for SImode in floating point registers, just
+; convert the SImode value to DImode. If we are loading the value from memory,
+; we will use the LFIWAX/LXSIWAX
(define_expand "floatsisf2"
[(set (match_operand:SF 0 "gpc_reg_operand")
(float:SF (match_operand:SI 1 "nonimmediate_operand")))]
@@ -6915,7 +6947,13 @@
{
if (TARGET_FCFIDS && TARGET_LFIWAX)
{
- emit_insn (gen_floatsisf2_lfiwax (operands[0], operands[1]));
+ if (TARGET_POWERPC64 && TARGET_P8_VECTOR)
+ {
+ rtx di_tmp = convert_to_mode (DImode, operands[1], false);
+ emit_insn (gen_floatdisf2 (operands[0], di_tmp));
+ }
+ else
+ emit_insn (gen_floatsisf2_lfiwax (operands[0], operands[1]));
DONE;
}
else if (TARGET_FCFID && TARGET_LFIWAX)
diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md
index f135fa079bd..2be5303d572 100644
--- a/gcc/config/rs6000/vsx.md
+++ b/gcc/config/rs6000/vsx.md
@@ -4319,10 +4319,11 @@
;; Get the element into the top position and use XVCVSWDP/XVCVUWDP
(define_insn_and_split "*vsx_extract_si_<uns>float_df"
[(set (match_operand:DF 0 "gpc_reg_operand" "=wa")
- (any_float:DF
- (vec_select:SI
- (match_operand:V4SI 1 "gpc_reg_operand" "v")
- (parallel [(match_operand:QI 2 "const_0_to_3_operand" "n")]))))
+ (float:DF
+ (any_extend:DI
+ (vec_select:SI
+ (match_operand:V4SI 1 "gpc_reg_operand" "v")
+ (parallel [(match_operand:QI 2 "const_0_to_3_operand" "n")])))))
(clobber (match_scratch:V4SI 3 "=v"))]
"VECTOR_MEM_VSX_P (V4SImode) && TARGET_DIRECT_MOVE_64BIT"
"#"
@@ -4362,10 +4363,11 @@
;; type.
(define_insn_and_split "*vsx_extract_si_<uns>float_<mode>"
[(set (match_operand:VSX_EXTRACT_FL 0 "gpc_reg_operand" "=wa")
- (any_float:VSX_EXTRACT_FL
- (vec_select:SI
- (match_operand:V4SI 1 "gpc_reg_operand" "v")
- (parallel [(match_operand:QI 2 "const_0_to_3_operand" "n")]))))
+ (float:VSX_EXTRACT_FL
+ (any_extend:DI
+ (vec_select:SI
+ (match_operand:V4SI 1 "gpc_reg_operand" "v")
+ (parallel [(match_operand:QI 2 "const_0_to_3_operand" "n")])))))
(clobber (match_scratch:V4SI 3 "=v"))
(clobber (match_scratch:DF 4 "=wa"))]
"VECTOR_MEM_VSX_P (V4SImode) && TARGET_DIRECT_MOVE_64BIT"
