[Bug rtl-optimization/102178] [12 Regression] SPECFP 2006 470.lbm regressions on AMD Zen CPUs after r12-897-gde56f95afaaa22

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102178

--- Comment #32 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the bad "head" can be fixed via

diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index c74edd1aaef..8f9f26e0a82 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -3580,9 +3580,9 @@
 ;; Possible store forwarding (partial memory) stall in alternatives 4, 6 and
7.
 (define_insn "*movdf_internal"
   [(set (match_operand:DF 0 "nonimmediate_operand"
-    "=Yf*f,m   ,Yf*f,?r ,!o,?*r ,!o,!o,?r,?m,?r,?r,v,v,v,m,*x,*x,*x,m ,r ,v,r 
,o ,r  ,m")
+    "=Yf*f,m   ,Yf*f,?r ,!o,?*r ,!o,!o,?r,?m,?r,?r,v,v,v,m,*x,*x,*x,m ,!r,!v,r
 ,o ,r  ,m")
        (match_operand:DF 1 "general_operand"
-    "Yf*fm,Yf*f,G   ,roF,r ,*roF,*r,F ,rm,rC,C ,F ,C,v,m,v,C ,*x,m ,*x,v,r
,roF,rF,rmF,rC"))]
+    "Yf*fm,Yf*f,G   ,roF,r ,*roF,*r,F ,rm,rC,C ,F ,C,v,m,v,C ,*x,m
,*x,!v,!r,roF,rF,rmF,rC"))]
   "!(MEM_P (operands[0]) && MEM_P (operands[1]))
    && (lra_in_progress || reload_completed
        || !CONST_DOUBLE_P (operands[1])

which is adding ! to r<->v alternatives.  That should eventually be done
by duplicating the alternatives and enabling one set via some enable
attribute based on some tunable.  I see those alternatives are already

   (set (attr "preferred_for_speed")
     (cond [(eq_attr "alternative" "3,4")
              (symbol_ref "TARGET_INTEGER_DFMODE_MOVES")
            (eq_attr "alternative" "20")
              (symbol_ref "TARGET_INTER_UNIT_MOVES_FROM_VEC")
            (eq_attr "alternative" "21")
              (symbol_ref "TARGET_INTER_UNIT_MOVES_TO_VEC")
           ]     
           (symbol_ref "true")))

not sure why it's preferred_for_speed here though - shouldn't that be
enabled for size if !TARGET_INTER_UNIT_MOVES_{TO,FROM}_VEC and otherwise
disabled?  Not sure if combining enabled and preferred_for_speed is
reasonably possible, but we have a preferred_for_size attribute here.

The diff with ! added is quite short, I've yet have to measure any
effect on LBM:

--- streamcollide.s.orig        2022-04-25 11:37:01.638733951 +0200
+++ streamcollide.s2    2022-04-25 11:35:54.885849296 +0200
@@ -33,28 +33,24 @@
        .p2align 4
        .p2align 3
 .L12:
-       movq    .LC0(%rip), %rax
-       vmovsd  .LC4(%rip), %xmm6
+       vmovsd  .LC0(%rip), %xmm2
+       vmovsd  .LC1(%rip), %xmm13
+       movabsq $0x3ff01878b7a1c25d, %rax
        movabsq $0x3fef85af6c69b5a6, %rdi
+       vmovsd  .LC2(%rip), %xmm12
+       vmovsd  .LC3(%rip), %xmm14
        movabsq $0x3ff03db8fde2ef4e, %r8
+       movabsq $0x3fefcea39c51dabe, %r9
+       vmovsd  .LC4(%rip), %xmm6
        vmovsd  .LC5(%rip), %xmm7
        movq    .LC8(%rip), %r11
-       movabsq $0x3fefcea39c51dabe, %r9
        movq    .LC6(%rip), %rdx
        movq    .LC7(%rip), %rcx
-       vmovq   %rax, %xmm2
-       vmovq   %rax, %xmm4
-       movq    .LC1(%rip), %rax
        movq    %r11, %rsi
        movq    %r11, %r12
-       vmovq   %rax, %xmm13
-       vmovq   %rax, %xmm8
-       movq    .LC2(%rip), %rax
-       vmovq   %rax, %xmm12
-       vmovq   %rax, %xmm5
-       movq    .LC3(%rip), %rax
-       vmovq   %rax, %xmm14
-       movabsq $0x3ff01878b7a1c25d, %rax
+       vmovsd  %xmm2, %xmm2, %xmm4
+       vmovsd  %xmm13, %xmm13, %xmm8
+       vmovsd  %xmm12, %xmm12, %xmm5
        vmovsd  %xmm14, -16(%rsp)
 .L5:
        vmulsd  .LC9(%rip), %xmm0, %xmm3