Various instructions are supported as integer operations as well as SIMD on AArch64. When register pressure is high, lra-constraints inserts spill code without taking the allocation class into account, and basically chooses the first available pattern that matches. Since this instruction has the SIMD version first it is usually chosen eventhough some of the operands are eventually allocated to integer registers. The result is inefficient code not only due to the higher latency of SIMD instructions but also due to the extra int<->FP moves. Placing the integer variant first in the shl pattern generates far more optimal spill code. A few more patterns are the wrong way around, which I'll address in a separate patch. I'm also looking into fixing lra-constraints to generate the expected code by taking the allocno class into account in the cost calculations during spilling.
2015-04-27 Wilco Dijkstra <wdijk...@arm.com> * gcc/config/aarch64/aarch64.md (aarch64_ashl_sisd_or_int_<mode>3): Place integer variant first. --- gcc/config/aarch64/aarch64.md | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md index 7163025..baef56a 100644 --- a/gcc/config/aarch64/aarch64.md +++ b/gcc/config/aarch64/aarch64.md @@ -3334,17 +3334,17 @@ ;; Logical left shift using SISD or Integer instruction (define_insn "*aarch64_ashl_sisd_or_int_<mode>3" - [(set (match_operand:GPI 0 "register_operand" "=w,w,r") + [(set (match_operand:GPI 0 "register_operand" "=r,w,w") (ashift:GPI - (match_operand:GPI 1 "register_operand" "w,w,r") - (match_operand:QI 2 "aarch64_reg_or_shift_imm_<mode>" "Us<cmode>,w,rUs<cmode>")))] + (match_operand:GPI 1 "register_operand" "r,w,w") + (match_operand:QI 2 "aarch64_reg_or_shift_imm_<mode>" "rUs<cmode>,Us<cmode>,w")))] "" "@ + lsl\t%<w>0, %<w>1, %<w>2 shl\t%<rtn>0<vas>, %<rtn>1<vas>, %2 - ushl\t%<rtn>0<vas>, %<rtn>1<vas>, %<rtn>2<vas> - lsl\t%<w>0, %<w>1, %<w>2" - [(set_attr "simd" "yes,yes,no") - (set_attr "type" "neon_shift_imm<q>, neon_shift_reg<q>,shift_reg")] + ushl\t%<rtn>0<vas>, %<rtn>1<vas>, %<rtn>2<vas>" + [(set_attr "simd" "no,yes,yes") + (set_attr "type" "shift_reg,neon_shift_imm<q>, neon_shift_reg<q>")] ) ;; Logical right shift using SISD or Integer instruction -- 1.9.1