| Message ID | 1529677967.7264.8.camel@us.ibm.com |
|---|---|
| State | New |
| Headers | show |
| Series | [rs6000] Change word selector to prefered location for vec_insert builtin | expand |
Hi Carl, On Fri, Jun 22, 2018 at 07:32:47AM -0700, Carl Love wrote: > The following patch changes the word selected when extracting the word > from the second vector to insert into the first vector by the > vec_insert() builtin. > > Specifically, the test case > > vector float > fn2 (float a, vector float b) > { > return vec_insert (a, b, 1); > } > > without the patch generates the code sequence > > xscvdpspn vs0,vs1 > xxextractuw vs0,vs0,4 > xxinsertw vs34,vs0,8 [ For -mcpu=power9 -mlittle ] Is this from something in the testsuite? I can't find it. > The xscvdpspn places the extracted word into words 0 and 1 of the > destination. The xxextractuw extracts word 1 (offset of 4 bytes)from > the source. The patch changes the offset so that the xxexractuw will > extract word 0 (offset 0 bytes) instead of word 1. The values are the > same so there is no functional change. But it was decided that using > word 0 was preferred choice. The patch looks fine. Okay for trunk. Thanks! Segher
> On Jun 22, 2018, at 7:56 PM, Segher Boessenkool <segher@kernel.crashing.org> wrote: > > Hi Carl, > > On Fri, Jun 22, 2018 at 07:32:47AM -0700, Carl Love wrote: >> The following patch changes the word selected when extracting the word >> from the second vector to insert into the first vector by the >> vec_insert() builtin. >> >> Specifically, the test case >> >> vector float >> fn2 (float a, vector float b) >> { >> return vec_insert (a, b, 1); >> } >> >> without the patch generates the code sequence >> >> xscvdpspn vs0,vs1 >> xxextractuw vs0,vs0,4 >> xxinsertw vs34,vs0,8 > > [ For -mcpu=power9 -mlittle ] > > Is this from something in the testsuite? I can't find it. It's not from the testsuite. I found the problem while working on some vector intrinsic documentation and alerted Carl. Thanks, Bill > >> The xscvdpspn places the extracted word into words 0 and 1 of the >> destination. The xxextractuw extracts word 1 (offset of 4 bytes)from >> the source. The patch changes the offset so that the xxexractuw will >> extract word 0 (offset 0 bytes) instead of word 1. The values are the >> same so there is no functional change. But it was decided that using >> word 0 was preferred choice. > > The patch looks fine. Okay for trunk. Thanks! > > > Segher >
diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md index a528ef2e8..6e7a4277f 100644 --- a/gcc/config/rs6000/vsx.md +++ b/gcc/config/rs6000/vsx.md @@ -3934,7 +3934,7 @@ operands[5] = gen_rtx_REG (V4SFmode, tmp_regno); operands[6] = gen_rtx_REG (V4SImode, tmp_regno); - operands[7] = GEN_INT (BYTES_BIG_ENDIAN ? 1 : 2); + operands[7] = GEN_INT (BYTES_BIG_ENDIAN ? 0 : 3); operands[8] = gen_rtx_REG (V4SImode, reg_or_subregno (operands[0])); } [(set_attr "type" "vecperm")
GCC Maintainers: The following patch changes the word selected when extracting the word from the second vector to insert into the first vector by the vec_insert() builtin. Specifically, the test case vector float fn2 (float a, vector float b) { return vec_insert (a, b, 1); } without the patch generates the code sequence xscvdpspn vs0,vs1 xxextractuw vs0,vs0,4 xxinsertw vs34,vs0,8 The xscvdpspn places the extracted word into words 0 and 1 of the destination. The xxextractuw extracts word 1 (offset of 4 bytes)from the source. The patch changes the offset so that the xxexractuw will extract word 0 (offset 0 bytes) instead of word 1. The values are the same so there is no functional change. But it was decided that using word 0 was preferred choice. Carl Love ----------------------------------------------------- gcc/ChangeLog: 2018-06-19 Carl Love <cel@us.ibm.com> * config/rs6000/vsx.md: Change word selector to prefered location. Signed-off-by: Carl Love <cel@us.ibm.com> --- gcc/config/rs6000/vsx.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)