haraldwolff/qemu-patch-raspberry4
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

tcg-arm: Improve constant generation

Browse source 
Try fully rotated arguments to mov and mvn before trying movt
or full decomposition.  Begin decomposition with mvn when it
looks like it'll help.  Examples include

-:        mov   r9, #0x00000fa0
-:        orr   r9, r9, #0x000ee000
-:        orr   r9, r9, #0x0ff00000
-:        orr   r9, r9, #0xf0000000
+:        mvn   r9, #0x0000005f
+:        eor   r9, r9, #0x00011000

Reviewed-by: Aurelien Jarno <aurelien@aurel32.net>
Signed-off-by: Richard Henderson <rth@twiddle.net>
This commit is contained in:
Richard Henderson 2013-03-04 23:16:24 -08:00 committed by Aurelien Jarno
parent 2df3f1ee68
commit e86e0f2807
1 changed files with 44 additions and 23 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

67
tcg/arm/tcg-target.c
View file

@ -427,15 +427,31 @@ static inline void tcg_out_dat_imm(TCGContext *s,
(rn << 16) | (rd << 12) | im);
}
static inline void tcg_out_movi32(TCGContext *s,
int cond, int rd, uint32_t arg)
static void tcg_out_movi32(TCGContext *s, int cond, int rd, uint32_t arg)
{
/* TODO: This is very suboptimal, we can easily have a constant
* pool somewhere after all the instructions. */
if ((int)arg < 0 && (int)arg >= -0x100) {
tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, (~arg) & 0xff);
} else if (use_armv7_instructions) {
/* use movw/movt */
int rot, opc, rn;
/* For armv7, make sure not to use movw+movt when mov/mvn would do.
Speed things up by only checking when movt would be required.
Prior to armv7, have one go at fully rotated immediates before
doing the decomposition thing below. */
if (!use_armv7_instructions || (arg & 0xffff0000)) {
rot = encode_imm(arg);
if (rot >= 0) {
tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0,
rotl(arg, rot) | (rot << 7));
return;
}
rot = encode_imm(~arg);
if (rot >= 0) {
tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0,
rotl(~arg, rot) | (rot << 7));
return;
}
}
/* Use movw + movt. */
if (use_armv7_instructions) {
/* movw */
tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
| ((arg << 4) & 0x000f0000) | (arg & 0xfff));
@ -444,22 +460,27 @@ static inline void tcg_out_movi32(TCGContext *s,
tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
| ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
}
} else {
int opc = ARITH_MOV;
int rn = 0;
do {
int i, rot;
i = ctz32(arg) & ~1;
rot = ((32 - i) << 7) & 0xf00;
tcg_out_dat_imm(s, cond, opc, rd, rn, ((arg >> i) & 0xff) | rot);
arg &= ~(0xff << i);
opc = ARITH_ORR;
rn = rd;
} while (arg);
return;
}
/* TODO: This is very suboptimal, we can easily have a constant
pool somewhere after all the instructions. */
opc = ARITH_MOV;
rn = 0;
/* If we have lots of leading 1's, we can shorten the sequence by
beginning with mvn and then clearing higher bits with eor. */
if (clz32(~arg) > clz32(arg)) {
opc = ARITH_MVN, arg = ~arg;
}
do {
int i = ctz32(arg) & ~1;
rot = ((32 - i) << 7) & 0xf00;
tcg_out_dat_imm(s, cond, opc, rd, rn, ((arg >> i) & 0xff) | rot);
arg &= ~(0xff << i);
opc = ARITH_EOR;
rn = rd;
} while (arg);
}
static inline void tcg_out_dat_rI(TCGContext *s, int cond, int opc, TCGArg dst,