| Message ID | 1385080047.30314.13.camel@gnopaine |
|---|---|
| State | New |
| Headers | show |
On Thu, Nov 21, 2013 at 7:27 PM, Bill Schmidt <wschmidt@linux.vnet.ibm.com> wrote: > Hi, > > This patch fixes two issues to allow correct compilation of > gcc.dg/torture/vec-cvt-1.c in little endian mode. The first reverts a > change in three patterns in vector.md. This is from an early patch that > preceded the general fix for vector permutes. As a consequence we ended > up swapping the input arguments twice. So we can simplify the code here > and have it operate the same for big and little endian. > > The other issue corrects a scenario where I managed to check for > endianness twice, with the effect that the code acts the same for both > big and little endian when it shouldn't. > > Bootstrapped and tested on powerpc64{,le}-unknown-linux-gnu with no > regressions. Is this ok for trunk? > > Thanks, > Bill > > > 2013-11-21 Bill Schmidt <wschmidt@vnet.ibm.com> > > * config/rs6000/vector.md (vec_pack_trunc_v2df): Revert previous > little endian change. > (vec_pack_sfix_trunc_v2df): Likewise. > (vec_pack_ufix_trunc_v2df): Likewise. > * config/rs6000/rs6000.c (rs6000_expand_interleave): Correct > double checking of endianness. Okay. Thanks, David
Index: gcc/config/rs6000/vector.md =================================================================== --- gcc/config/rs6000/vector.md (revision 205145) +++ gcc/config/rs6000/vector.md (working copy) @@ -831,12 +831,7 @@ emit_insn (gen_vsx_xvcvdpsp (r1, operands[1])); emit_insn (gen_vsx_xvcvdpsp (r2, operands[2])); - - if (BYTES_BIG_ENDIAN) - rs6000_expand_extract_even (operands[0], r1, r2); - else - rs6000_expand_extract_even (operands[0], r2, r1); - + rs6000_expand_extract_even (operands[0], r1, r2); DONE; }) @@ -851,12 +846,7 @@ emit_insn (gen_vsx_xvcvdpsxws (r1, operands[1])); emit_insn (gen_vsx_xvcvdpsxws (r2, operands[2])); - - if (BYTES_BIG_ENDIAN) - rs6000_expand_extract_even (operands[0], r1, r2); - else - rs6000_expand_extract_even (operands[0], r2, r1); - + rs6000_expand_extract_even (operands[0], r1, r2); DONE; }) @@ -871,12 +861,7 @@ emit_insn (gen_vsx_xvcvdpuxws (r1, operands[1])); emit_insn (gen_vsx_xvcvdpuxws (r2, operands[2])); - - if (BYTES_BIG_ENDIAN) - rs6000_expand_extract_even (operands[0], r1, r2); - else - rs6000_expand_extract_even (operands[0], r2, r1); - + rs6000_expand_extract_even (operands[0], r1, r2); DONE; }) Index: gcc/config/rs6000/rs6000.c =================================================================== --- gcc/config/rs6000/rs6000.c (revision 205145) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -30116,7 +30116,7 @@ rs6000_expand_interleave (rtx target, rtx op0, rtx unsigned i, high, nelt = GET_MODE_NUNITS (vmode); rtx perm[16]; - high = (highp == BYTES_BIG_ENDIAN ? 0 : nelt / 2); + high = (highp ? 0 : nelt / 2); for (i = 0; i < nelt / 2; i++) { perm[i * 2] = GEN_INT (i + high);
Hi, This patch fixes two issues to allow correct compilation of gcc.dg/torture/vec-cvt-1.c in little endian mode. The first reverts a change in three patterns in vector.md. This is from an early patch that preceded the general fix for vector permutes. As a consequence we ended up swapping the input arguments twice. So we can simplify the code here and have it operate the same for big and little endian. The other issue corrects a scenario where I managed to check for endianness twice, with the effect that the code acts the same for both big and little endian when it shouldn't. Bootstrapped and tested on powerpc64{,le}-unknown-linux-gnu with no regressions. Is this ok for trunk? Thanks, Bill 2013-11-21 Bill Schmidt <wschmidt@vnet.ibm.com> * config/rs6000/vector.md (vec_pack_trunc_v2df): Revert previous little endian change. (vec_pack_sfix_trunc_v2df): Likewise. (vec_pack_ufix_trunc_v2df): Likewise. * config/rs6000/rs6000.c (rs6000_expand_interleave): Correct double checking of endianness.