[committed] [RISC-V] Fix detection of store pair fusion cases

[Jeff Law <jlaw@ventanamicro.com>]

[-- Attachment #1: Type: text/plain, Size: 837 bytes --]


We've got the ability to count the number of store pair fusions 
happening in the front-end of the pipeline.  When comparing some code 
from last year vs the current trunk we saw a fairly dramatic drop.

The problem is the store pair fusion detection code was actively harmful 
due to a minor bug in checking offsets.   So instead of pairing up 8 
byte stores such as sp+0 with sp+8, it tried to pair up sp+8 and sp+16.

Given uarch sensitivity I didn't try to pull together a testcase.  But 
we could certainly see the undesirable behavior in benchmarks as 
simplistic as dhrystone up through spec2017.

Anyway, bootstrapped a while back.  Also verified through our 
performance counters that store pair fusion rates are back up. 
Regression tested with crosses a few minutes ago.



Pushing to the trunk and coordination branch.

jeff

[-- Attachment #2: P --]
[-- Type: text/plain, Size: 4127 bytes --]

commit fad93e7617ce1aafb006983a71b6edc9ae1eb2d1
Author: Jeff Law <jlaw@ventanamicro.com>
Date:   Wed May 1 11:28:41 2024 -0600

    [committed] [RISC-V] Fix detection of store pair fusion cases
    
    We've got the ability to count the number of store pair fusions happening in
    the front-end of the pipeline.  When comparing some code from last year vs the
    current trunk we saw a fairly dramatic drop.
    
    The problem is the store pair fusion detection code was actively harmful due to
    a minor bug in checking offsets.   So instead of pairing up 8 byte stores such
    as sp+0 with sp+8, it tried to pair up sp+8 and sp+16.
    
    Given uarch sensitivity I didn't try to pull together a testcase.  But we could
    certainly see the undesirable behavior in benchmarks as simplistic as dhrystone
    up through spec2017.
    
    Anyway, bootstrapped a while back.  Also verified through our performance
    counters that store pair fusion rates are back up.  Regression tested with
    crosses a few minutes ago.
    
    gcc/
            * config/riscv/riscv.cc (riscv_macro_fusion_pair_p): Break out
            tests for easier debugging in store pair fusion case.  Fix offset
            check in same.

diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 0f62b295b96..24d1ead3902 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -8874,26 +8874,43 @@ riscv_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr)
 	  extract_base_offset_in_addr (SET_DEST (prev_set), &base_prev, &offset_prev);
 	  extract_base_offset_in_addr (SET_DEST (curr_set), &base_curr, &offset_curr);
 
-	  /* The two stores must be contained within opposite halves of the same
-	     16 byte aligned block of memory.  We know that the stack pointer and
-	     the frame pointer have suitable alignment.  So we just need to check
-	     the offsets of the two stores for suitable alignment.
-
-	     Originally the thought was to check MEM_ALIGN, but that was reporting
-	     incorrect alignments, even for SP/FP accesses, so we gave up on that
-	     approach.  */
-	  if (base_prev != NULL_RTX
-	      && base_curr != NULL_RTX
-	      && REG_P (base_prev)
-	      && REG_P (base_curr)
-	      && REGNO (base_prev) == REGNO (base_curr)
-	      && (REGNO (base_prev) == STACK_POINTER_REGNUM
-		  || REGNO (base_prev) == HARD_FRAME_POINTER_REGNUM)
-	      && ((INTVAL (offset_prev) == INTVAL (offset_curr) + 8
-		   && (INTVAL (offset_prev) % 16) == 0)
-		  || ((INTVAL (offset_curr) == INTVAL (offset_prev) + 8)
-		      && (INTVAL (offset_curr) % 16) == 0)))
-	    return true;
+	  /* Fail if we did not find both bases.  */
+	  if (base_prev == NULL_RTX || base_curr == NULL_RTX)
+	    return false;
+
+	  /* Fail if either base is not a register.  */
+	  if (!REG_P (base_prev) || !REG_P (base_curr))
+	    return false;
+
+	  /* Fail if the bases are not the same register.  */
+	  if (REGNO (base_prev) != REGNO (base_curr))
+	    return false;
+
+	  /* Originally the thought was to check MEM_ALIGN, but that was
+	     reporting incorrect alignments, even for SP/FP accesses, so we
+	     gave up on that approach.  Instead just check for stack/hfp
+	     which we know are aligned.  */
+	  if (REGNO (base_prev) != STACK_POINTER_REGNUM
+	      && REGNO (base_prev) != HARD_FRAME_POINTER_REGNUM)
+	    return false;
+
+	  /* The two stores must be contained within opposite halves of the
+	     same 16 byte aligned block of memory.  We know that the stack
+	     pointer and the frame pointer have suitable alignment.  So we
+	     just need to check the offsets of the two stores for suitable
+	     alignment.  */
+	  /* Get the smaller offset into OFFSET_PREV.  */
+	  if (INTVAL (offset_prev) > INTVAL (offset_curr))
+	    std::swap (offset_prev, offset_curr);
+
+	  /* If the smaller offset (OFFSET_PREV) is not 16 byte aligned,
+	     then fail.  */
+	  if ((INTVAL (offset_prev) % 16) != 0)
+	    return false;
+
+	  /* The higher offset must be 8 bytes more than the lower
+	     offset.  */
+	  return (INTVAL (offset_prev) + 8 == INTVAL (offset_curr));
 	}
     }