Patchwork [4/N,rs6000,cft] -mfused-add cleanup

login
register
mail settings
Submitter IainS
Date Nov. 13, 2010, 6 p.m.
Message ID <8BBDEE74-1398-44A3-91F9-74BA5F1DDD39@sandoe-acoustics.co.uk>
Download mbox | patch
Permalink /patch/71065/
State New
Headers show

Comments

IainS - Nov. 13, 2010, 6 p.m.
Hi Richard,

On 11 Nov 2010, at 20:06, Michael Meissner wrote:

> On Thu, Nov 11, 2010 at 10:09:29AM -0800, Richard Henderson wrote:
>> According to Meissner via IRC, there are a number of regressions on  
>> the
>> powerpc port.  One example is cc.target/powerpc/ppc-fmadd-3.c, a  
>> testcase
>> that wants to see FMA generated even when the exact operation isn't
>> available.  I think there is some sense to this, as indicated in the
>> comment I added to the code.
>>
>> Ok?

This (I think it became r166611) caused:

http://gcc.gnu.org/ml/gcc-regression/2010-11/msg00148.html

with the diff between stage 2 & stage 3  libcpp/symtab.o:

$ diff -u s2s.txt s3s.txt
  00000864       lfd     f0,0xa30(r26)
  00000868       lwz     r3,0x0(r31)
-0000086c       fneg    f30,f28
-00000870       fmadd   f31,f30,f28,f31
+0000086c       fmul    f13,f28,f28
+00000870       fsub    f31,f31,f13
  00000874       fcmpu   cr7,f31,f0
  00000878       blt     cr7,0x940
  0000087c       beq     cr7,0x938


any more detail/analysis needed ... let me know.
cheers
Iain

Patch

--- s2s.txt     2010-11-13 17:56:37.000000000 +0000
+++ s3s.txt     2010-11-13 17:56:43.000000000 +0000
@@ -1,4 +1,4 @@ 
-s2st.o.stripped:
+s3st.o.stripped:

@@ -547,8 +547,8 @@ 
  00000860       fdiv    f31,f31,f30