Hi all,
Sphinx3 benchmark segmented when built and ran with -Ofast and
-fprefecth-loop-arrays.
When I debugged the benchmark executable, found that it is caused due to
register swapping happening
at reload for the smaxdf3 pattern.
197r.ira
---------
(insn 457 456 458 28 (set (reg:DF 687)
(smax:DF (reg/v:DF 367 [ dval1 ])
(reg/v:DF 154 [ dval1 ]))) eval.c:58 914 {smaxdf3}
(expr_list:REG_DEAD (reg/v:DF 367 [ dval1 ])
(expr_list:REG_DEAD (reg/v:DF 154 [ dval1 ])
(nil))))
198.reload
-----------
(insn 457 456 458 28 (set (reg:DF 21 xmm0 [687])
(smax:DF (reg/v:DF 21 xmm0 [orig:154 dval1 ] [154])
(reg/v:DF 51 xmm14 [orig:367 dval1 ] [367]))) eval.c:58 914
{smaxdf3}
(nil))
In my case the register 367 is sNAN.
MAXSD instruction is sensitive to sNAN and is not commutative.
DEST[63-0] IF (DEST[63-0] == SNaN) THEN SRC[63-0];
ELSE IF SRC[63-0] == SNaN) THEN SRC[63-0];
ELSE IF (DEST[63-0] > SRC[63-0])
THEN DEST[63-0]
ELSE SRC[63-0];
FI;
So I think the commutative % should be removed from the pattern when -Ofast
(fastmath) is set.
Index: gcc/config/i386/i386.md
===================================================================
--- gcc/config/i386/i386.md (revision 184372)
+++ gcc/config/i386/i386.md (working copy)
@@ -16507,7 +16507,7 @@
(define_insn "<code><mode>3"
[(set (match_operand:MODEF 0 "register_operand" "=x,x")
(smaxmin:MODEF
- (match_operand:MODEF 1 "nonimmediate_operand" "%0,x")
+ (match_operand:MODEF 1 "nonimmediate_operand" "0,x")
(match_operand:MODEF 2 "nonimmediate_operand" "xm,xm")))]
"SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH"
"@
This solved the sphinx segmentation. But the comments on i386.md says the
commutative % mark is correct.
Please provide your opinion.
Regards,
Venkat
