Commit 3ee1b8550c6e0bb48df4614ed185964dc1a036d3
1 parent
bdadc0b5
Optimize 64 bit bswap
Use rldimi instead of rldicr/or pair, saves us one instruction. Suggested by Hollis Blanchard. git-svn-id: svn://svn.savannah.nongnu.org/qemu/trunk@5404 c046a42c-6fe2-441c-8c8c-71466251a162
Showing
1 changed file
with
5 additions
and
5 deletions
tcg/ppc64/tcg-target.c
| @@ -303,6 +303,7 @@ static int tcg_target_const_match (tcg_target_long val, | @@ -303,6 +303,7 @@ static int tcg_target_const_match (tcg_target_long val, | ||
| 303 | 303 | ||
| 304 | #define RLDICL XO30( 0) | 304 | #define RLDICL XO30( 0) |
| 305 | #define RLDICR XO30( 1) | 305 | #define RLDICR XO30( 1) |
| 306 | +#define RLDIMI XO30( 3) | ||
| 306 | 307 | ||
| 307 | #define BCLR XO19( 16) | 308 | #define BCLR XO19( 16) |
| 308 | #define BCCTR XO19(528) | 309 | #define BCCTR XO19(528) |
| @@ -691,11 +692,10 @@ static void tcg_out_qemu_ld (TCGContext *s, const TCGArg *args, int opc) | @@ -691,11 +692,10 @@ static void tcg_out_qemu_ld (TCGContext *s, const TCGArg *args, int opc) | ||
| 691 | break; | 692 | break; |
| 692 | case 3: | 693 | case 3: |
| 693 | if (bswap) { | 694 | if (bswap) { |
| 694 | - tcg_out32 (s, LWBRX | RT (0) | RB (r0)); | ||
| 695 | - tcg_out32 (s, ADDI | RT (r1) | RA (r0) | 4); | ||
| 696 | - tcg_out32 (s, LWBRX | RT (data_reg) | RB (r1)); | ||
| 697 | - tcg_out_rld (s, RLDICR, data_reg, data_reg, 32, 31); | ||
| 698 | - tcg_out32 (s, OR | SAB (0, data_reg, data_reg)); | 695 | + tcg_out_movi32 (s, 0, 4); |
| 696 | + tcg_out32 (s, LWBRX | RT (data_reg) | RB (r0)); | ||
| 697 | + tcg_out32 (s, LWBRX | RT ( r1) | RA (r0)); | ||
| 698 | + tcg_out_rld (s, RLDIMI, data_reg, r1, 32, 0); | ||
| 699 | } | 699 | } |
| 700 | else tcg_out32 (s, LD | RT (data_reg) | RA (r0)); | 700 | else tcg_out32 (s, LD | RT (data_reg) | RA (r0)); |
| 701 | break; | 701 | break; |