Hello,
Changes from v2:
* Some formatting.
Changes from v1:
* Added more info to commit message fixed indentation.
This patch removes redundant vector compare instructions and logic
from the vec_first_mismatch_or_eos_index intrinsic.
Currently, GCC generates extra vcmpneb instructions and additional
masking logic (xxland, xxlorc) to handle EOS and mismatch comparisons.
However, a single vcmpnezb instruction already suffices, as it covers
both By eliminating the redundant comparisons (vcmpneb) and the
associated logic (xxland/xxlorc) we produce shorter,
more efficient code.
Bootstrapped and tested on powerpc64le-linux-gnu with no regressions.
Im not sure why formatting was messed up last time but it should be
fixed now also added more info about insns and builtins
requiring ISA 3.0
> Why was it ever written this way, any clue?
Im not sure actually I tried contacting Carl love but no updates from him yet
2025-10-22 Vijay Shankar <[email protected]>
gcc/ChangeLog:
PR target/116004
* config/rs6000/vsx.md (first_mismatch_or_eos_index): Remove redundant
emit_insns.
---
gcc/config/rs6000/vsx.md | 22 +++-------------------
1 file changed, 3 insertions(+), 19 deletions(-)
diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md
index dd3573b80..cac1e6080 100644
--- a/gcc/config/rs6000/vsx.md
+++ b/gcc/config/rs6000/vsx.md
@@ -5668,29 +5668,13 @@
"TARGET_P9_VECTOR"
{
int sh;
- rtx cmpz1_result = gen_reg_rtx (<MODE>mode);
- rtx cmpz2_result = gen_reg_rtx (<MODE>mode);
- rtx cmpz_result = gen_reg_rtx (<MODE>mode);
- rtx not_cmpz_result = gen_reg_rtx (<MODE>mode);
- rtx and_result = gen_reg_rtx (<MODE>mode);
rtx result = gen_reg_rtx (<MODE>mode);
- rtx vzero = gen_reg_rtx (<MODE>mode);
-
- /* Vector with zeros in elements that correspond to zeros in operands. */
- emit_move_insn (vzero, CONST0_RTX (<MODE>mode));
- emit_insn (gen_vcmpne<VSX_EXTRACT_WIDTH> (cmpz1_result, operands[1], vzero));
- emit_insn (gen_vcmpne<VSX_EXTRACT_WIDTH> (cmpz2_result, operands[2], vzero));
- emit_insn (gen_and<mode>3 (and_result, cmpz1_result, cmpz2_result));
+ /* Vector with ones in elements that match or elements corresponding to zeros
+ in operands. */
- /* Vector with ones in elments that match. */
- emit_insn (gen_vcmpnez<VSX_EXTRACT_WIDTH> (cmpz_result, operands[1],
+ emit_insn (gen_vcmpnez<VSX_EXTRACT_WIDTH> (result, operands[1],
operands[2]));
- emit_insn (gen_one_cmpl<mode>2 (not_cmpz_result, cmpz_result));
-
- /* Create vector with ones in elements where there was a zero in one of
- the source elements or the elements did not match. */
- emit_insn (gen_nand<mode>3 (result, and_result, not_cmpz_result));
sh = GET_MODE_SIZE (GET_MODE_INNER (<MODE>mode)) / 2;
if (<MODE>mode == V16QImode)
--
2.47.3
