ASM optimization for fiq_playback(). Saves about 0.4MHz of CPU while playback on PP502x/PP5002.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@17097 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 61 insertions, 23 deletions

diff --git a/firmware/target/arm/pcm-pp.c b/firmware/target/arm/pcm-pp.c
index 64c6d0cdc8..489862e5de 100644
--- a/firmware/target/arm/pcm-pp.c
+++ b/firmware/target/arm/pcm-pp.c
@@ -107,11 +107,17 @@ void pcm_apply_settings(void)
     pcm_curr_sampr = pcm_freq;
 }
 
-/* ASM optimised FIQ handler. GCC fails to make use of the fact that FIQ mode
-   has registers r8-r14 banked, and so does not need to be saved. This routine
-   uses only these registers, and so will never touch the stack unless it
-   actually needs to do so when calling pcm_callback_for_more. C version is
-   still included below for reference and testing.
+/* ASM optimised FIQ handler. Checks for the minimum allowed loop cycles by evalutation of
+ * free IISFIFO-slots against available source buffer words. Through this it is possible to 
+ * move the check for IIS_TX_FREE_COUNT outside the loop and do some further optimization.
+ * Right after the loops (source buffer -> IISFIFO) are done we need to check whether we
+ * have to exit FIQ handler (this must be done, if all free FIFO slots were filled) or
+ * we will have to get some new source data.
+ * Important information kept from former ASM implementation (not used anymore): GCC fails 
+ * to make use of the fact that FIQ mode has registers r8-r14 banked, and so does not need 
+ * to be saved. This routine uses only these registers, and so will never touch the stack 
+ * unless it actually needs to do so when calling pcm_callback_for_more. C version is still 
+ * included below for reference and testing.
  */
 #if 1
 void fiq_playback(void) ICODE_ATTR __attribute__((naked));
@@ -122,9 +128,11 @@ void fiq_playback(void)
      * addresses we need are generated by using offsets with these two.
      * r10 + 0x40 is IISFIFO_WR, and r10 + 0x0c is IISFIFO_CFG.
      * r8 and r9 contains local copies of p and size respectively.
-     * r12 is a working register.
+     * r0-r3 and r12 is a working register.
      */
     asm volatile (
+        "stmfd   sp!, { r0-r3, lr }  \n" /* stack scratch regs and lr */
+        
 #if CONFIG_CPU == PP5002
         "ldr     r12, =0xcf001040    \n" /* Some magic from iPodLinux */
         "ldr     r12, [r12]          \n"
@@ -132,24 +140,54 @@ void fiq_playback(void)
         "ldmia   r11, { r8-r9 }      \n" /* r8 = p, r9 = size */
         "cmp     r9, #0              \n" /* is size 0? */
         "beq     .more_data          \n" /* if so, ask pcmbuf for more data */
-    ".fifo_loop:                     \n"
-        "ldr     r12, [r10, %[cfg]]  \n" /* read IISFIFO_CFG to check FIFO status */
-        "ands    r12, r12, %[mask]   \n"
-        "beq     .exit               \n" /* FIFO full, exit */
+
 #if SAMPLE_SIZE == 16
-        "ldr     r12, [r8], #4       \n" /* load two samples */
-        "str     r12, [r10, %[wr]]   \n" /* write them */
+    ".check_fifo:                    \n"   
+        "ldr     r0, [r10, %[cfg]]   \n" /* read IISFIFO_CFG to check FIFO status */
+        "and     r0, r0, %[mask]     \n" /* r0 = IIS_TX_FREE_COUNT << 16 (PP502x) */
+        
+        "mov     r1, r0, lsr #16     \n" /* number of free FIFO slots */     
+        "cmp     r1, r9, lsr #2      \n" /* number of words from source */
+        "movgt   r1, r9, lsr #2      \n" /* r1 = amount of allowed loops */
+        "sub     r9, r9, r1, lsl #2  \n" /* r1 words will be written in following loop */
+        
+        "subs    r1, r1, #2          \n"
+    ".fifo_loop_2:                   \n"
+        "ldmgeia r8!, {r2, r12}      \n" /* load four samples */
+        "strge   r2 , [r10, %[wr]]   \n" /* write sample 0-1 to IISFIFO_WR */
+        "strge   r12, [r10, %[wr]]   \n" /* write sample 2-3 to IISFIFO_WR */
+        "subges  r1, r1, #2          \n" /* one more loop? */
+        "bge     .fifo_loop_2        \n" /* yes, continue */
+        
+        "tst     r1, #1              \n" /* two samples (one word) left? */
+        "ldrne   r12, [r8], #4       \n" /* load two samples */
+        "strne   r12, [r10, %[wr]]   \n" /* write sample 0-1 to IISFIFO_WR */
+        
+        "cmp     r9, #0              \n" /* either FIFO is full or source buffer is empty */
+        "bgt     .exit               \n" /* if source buffer is not empty, FIFO must be full */
 #elif SAMPLE_SIZE == 32
+    ".check_fifo:                    \n"   
+        "ldr     r0, [r10, %[cfg]]   \n" /* read IISFIFO_CFG to check FIFO status */
+        "and     r0, r0, %[mask]     \n" /* r0 = IIS_TX_FREE_COUNT << 23 (PP5002) */
+        
+        "mov     r1, r0, lsr #24     \n" /* number of free pairs of FIFO slots */     
+        "cmp     r1, r9, lsr #2      \n" /* number of words from source */
+        "movgt   r1, r9, lsr #2      \n" /* r1 = amount of allowed loops */
+        "sub     r9, r9, r1, lsl #2  \n" /* r1 words will be written in following loop */
+        
+    ".fifo_loop:                     \n"
         "ldr     r12, [r8], #4       \n" /* load two samples */
-        "mov     r12, r12, ror #16   \n" /* put left sample at the top bits */
-        "str     r12, [r10, %[wr]]   \n" /* write top sample, lower sample ignored */
-        "mov     r12, r12, lsl #16   \n" /* shift lower sample up */
-        "str     r12, [r10, %[wr]]   \n" /* then write it */
+        "mov     r2 , r12, lsl #16   \n" /* put left sample at the top bits */
+        "str     r2 , [r10, %[wr]]   \n" /* write top sample to IISFIFO_WR */
+        "str     r12, [r10, %[wr]]   \n" /* write low sample to IISFIFO_WR*/
+        "subs    r1, r1, #1          \n" /* one more loop? */
+        "bgt     .fifo_loop          \n" /* yes, continue */
+        
+        "cmp     r9, #0              \n" /* either FIFO is full or source buffer is empty */
+        "bgt     .exit               \n" /* if source buffer is not empty, FIFO must be full */
 #endif
-        "subs    r9, r9, #4          \n" /* check if we have more samples */
-        "bne     .fifo_loop          \n" /* yes, continue */
+        
     ".more_data:                     \n"
-        "stmfd   sp!, { r0-r3, lr }  \n" /* stack scratch regs and lr */
         "ldr     r2, =pcm_callback_for_more \n"
         "ldr     r2, [r2]            \n" /* get callback address */
         "cmp     r2, #0              \n" /* check for null pointer */
@@ -160,21 +198,21 @@ void fiq_playback(void)
         "bxne    r2                  \n"
         "ldmia   r11, { r8-r9 }      \n" /* reload p and size */
         "cmp     r9, #0              \n" /* did we actually get more data? */
-        "ldmnefd sp!, { r0-r3, lr }  \n"
-        "bne     .fifo_loop          \n" /* yes, continue to try feeding FIFO */
+        "bne     .check_fifo         \n"
         "ldr     r12, =pcm_play_dma_stop \n"
         "mov     lr, pc              \n"
         "bx      r12                 \n"
         "ldr     r12, =pcm_play_dma_stopped_callback \n"
         "mov     lr, pc              \n"
         "bx      r12                 \n"
-        "ldmfd   sp!, { r0-r3, lr }  \n"
+
     ".exit:                          \n" /* (r8=0 if stopping, look above) */
         "stmia   r11, { r8-r9 }      \n" /* save p and size */
+        "ldmfd   sp!, { r0-r3, lr }  \n"
         "subs    pc, lr, #4          \n" /* FIQ specific return sequence */
         ".ltorg                      \n"
         : /* These must only be integers! No regs */
-        : [mask]"i"(IIS_TX_FREE_MASK & (IIS_TX_FREE_MASK-1)),
+        : [mask]"i"(IIS_TX_FREE_MASK),
           [cfg]"i"((int)&IISFIFO_CFG - (int)&IISCONFIG),
           [wr]"i"((int)&IISFIFO_WR - (int)&IISCONFIG)
     );