https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109052
Bug ID: 109052
Summary: Unnecessary reload with -mfpmath=both
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ubizjak at gmail dot com
Target Milestone: ---
Following testcase:
--cut here--
double foo (double a)
{
double tmp = a;
asm ("" : "+t" (tmp));
return a * tmp;
}
--cut here--
compiles with -O2 -mfpmath=both -msse2 to:
foo:
movsd %xmm0, -8(%rsp)
fldl -8(%rsp)
fstpl -8(%rsp)
--> movsd -8(%rsp), %xmm1
mulsd %xmm1, %xmm0
ret
There is no need to reload from -8(%rsp) to %xmm1, mulsd accepts memory
operands. Optimal code would be:
foo:
movsd %xmm0, -8(%rsp)
fldl -8(%rsp)
fstpl -8(%rsp)
mulsd -8(%rsp), %xmm0
ret
Please note that with -mfpmath=sse, expected code is generated.
The insns pattern is defined in i386.md:
(define_insn "*fop_<mode>_comm"
[(set (match_operand:MODEF 0 "register_operand" "=f,x,v")
(match_operator:MODEF 3 "binary_fp_operator"
[(match_operand:MODEF 1 "nonimmediate_operand" "%0,0,v")
(match_operand:MODEF 2 "nonimmediate_operand" "fm,xm,vm")]))]
"((SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
|| (TARGET_80387 && X87_ENABLE_ARITH (<MODE>mode)))
&& COMMUTATIVE_ARITH_P (operands[3])
&& !(MEM_P (operands[1]) && MEM_P (operands[2]))"
"* return output_387_binary_op (insn, operands);"
[(set (attr "type")
(if_then_else (eq_attr "alternative" "1,2")
(if_then_else (match_operand:MODEF 3 "mult_operator")
(const_string "ssemul")
(const_string "sseadd"))
(if_then_else (match_operand:MODEF 3 "mult_operator")
(const_string "fmul")
(const_string "fop"))))
(set_attr "isa" "*,noavx,avx")
(set_attr "prefix" "orig,orig,vex")
(set_attr "mode" "<MODE>")
(set (attr "enabled")
(if_then_else
(match_test ("SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH"))
(if_then_else
(eq_attr "alternative" "0")
(symbol_ref "TARGET_MIX_SSE_I387
&& X87_ENABLE_ARITH (<MODE>mode)")
(const_string "*"))
(if_then_else
(eq_attr "alternative" "0")
(symbol_ref "true")
(symbol_ref "false"))))])
The difference between -mfpmath=both and -mfpmath=sse is in alternative 0,
which is enabled with -mfpmath=both and disabled with -mpmath=sse. Using -mavx
instead of -msse2 with -mfpmath=both does not solve the problem.