From: Andrew Cooper <andrew.cooper3@citrix.com>
Date: Fri, 15 Mar 2024 11:31:33 +0000 (+0000)
Subject: xen/vpci: Improve code generation in mask_write()
X-Git-Url: http://xenbits.xensource.com/gitweb?a=commitdiff_plain;h=188fa82305e72b725473db9146e20cc9abf7bff3;p=people%2Froyger%2Fxen.git

xen/vpci: Improve code generation in mask_write()

The use of __clear_bit() forces dmask to be spilled to the stack, and
interferes with the compiler heuristcs for some upcoming improvements to the
ffs() code generation.

First, shrink dmask to just the active vectors by making out the upper bits.
This replaces the "i < msi->vectors" part of the loop condition.

Next, use a simple while() loop with "clear bottom bit" expressed in plane C,
which affords the optimiser a far better understanding of what the loop is
doing.

Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
---

diff --git a/xen/drivers/vpci/msi.c b/xen/drivers/vpci/msi.c
index d3aa5df089..30adcf7df0 100644
--- a/xen/drivers/vpci/msi.c
+++ b/xen/drivers/vpci/msi.c
@@ -169,13 +169,15 @@ static void cf_check mask_write(
 
     if ( msi->enabled )
     {
-        unsigned int i;
+        /* Skip changes to vectors which aren't enabled. */
+        dmask &= (~0U >> (32 - msi->vectors));
 
-        for ( i = ffs(dmask) - 1; dmask && i < msi->vectors;
-              i = ffs(dmask) - 1 )
+        while ( dmask )
         {
+            unsigned int i = ffs(dmask) - 1;
+
             vpci_msi_arch_mask(msi, pdev, i, (val >> i) & 1);
-            __clear_bit(i, &dmask);
+            dmask &= (dmask - 1);
         }
     }