Hello!
The problem here was with x86 FP move patterns that allowed FP
constants for CM_MEDIUM and CM_LARGE memory models, but not others.
However, this interfered with ifcvt cmove detection, since FP constant
was rematerialized in, not copied to the output register.
According to ix86_in_large_data_p, automatic variables are never large
data, so there is no justification to treat CM_MEDIUM memory models
any different than CM_SMALL model. In CM_SMALL models, constants are
always expanded as loads from memory, so CSE passes can do their job,
and after registers are allocated, we simplify relevant loads from
memory to a simple constant move insn anyway.
In contrast, CM_LARGE{,_PIC} models have costly memory access, so we
should avoid loads from memory as much as possible. These models
should be treated like -Os case, where we already avoid memory loads,
*unless* we are sure that the constant will result in a simple
constant move insn. Unfortunately, above mentioned ifcvt cmove
detection interference will trigger for -Os and CM_LARGE models, but
hopefully, STRICT_MIN/MAX_EXPR patches [1] (not yet fully reviewed and
committed) will resolve this problem.
2016-05-14 Uros Bizjak <ubiz...@gmail.com>
PR target/71097
* config/i386/i386.md (*movtf_internal): Before register allocation,
do not allow FP constants for CM_MEDIUM memory model, allow only
standard FP constants for CM_LARGE and CM_LARGE_PIC models.
(*movxf_internal): Ditto.
(*movdf_internal): Ditto.
(*movsf_internal): Ditto.
Patch was bootstrapped and regression tested on
x86_64-linux-gnu/-mcmodel=medium/{,-fpic}.
Committed to mainline SVN.
[1] https://gcc.gnu.org/ml/gcc-patches/2015-08/msg00330.html
Uros.
Index: i386.md
===================================================================
--- i386.md (revision 236239)
+++ i386.md (working copy)
@@ -3114,9 +3114,9 @@
"(TARGET_64BIT || TARGET_SSE)
&& !(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (!can_create_pseudo_p ()
- || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| !CONST_DOUBLE_P (operands[1])
- || (optimize_function_for_size_p (cfun)
+ || ((optimize_function_for_size_p (cfun)
+ || (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC))
&& standard_sse_constant_p (operands[1], TFmode) == 1
&& !memory_operand (operands[0], TFmode))
|| (!TARGET_MEMORY_MISMATCH_STALL
@@ -3200,9 +3200,9 @@
"fm,f,G,roF,r , *roF,*r,F ,C,roF,rF"))]
"!(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (!can_create_pseudo_p ()
- || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| !CONST_DOUBLE_P (operands[1])
- || (optimize_function_for_size_p (cfun)
+ || ((optimize_function_for_size_p (cfun)
+ || (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC))
&& standard_80387_constant_p (operands[1]) > 0
&& !memory_operand (operands[0], XFmode))
|| (!TARGET_MEMORY_MISMATCH_STALL
@@ -3273,9 +3273,9 @@
"Yf*fm,Yf*f,G ,roF,r ,*roF,*r,F ,rm,rC,C ,F ,C,v,m,v,C ,*x,m ,*x,Yj,r
,roF,rF,rmF,rC"))]
"!(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (!can_create_pseudo_p ()
- || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| !CONST_DOUBLE_P (operands[1])
- || (optimize_function_for_size_p (cfun)
+ || ((optimize_function_for_size_p (cfun)
+ || (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC))
&& ((!(TARGET_SSE2 && TARGET_SSE_MATH)
&& standard_80387_constant_p (operands[1]) > 0)
|| (TARGET_SSE2 && TARGET_SSE_MATH
@@ -3475,9 +3475,9 @@
"Yf*fm,Yf*f,G ,rmF,rF,C,v,m,v,Yj,r ,*y ,m ,*y,*Yn,r ,rmF,rF"))]
"!(MEM_P (operands[0]) && MEM_P (operands[1]))
&& (!can_create_pseudo_p ()
- || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
|| !CONST_DOUBLE_P (operands[1])
- || (optimize_function_for_size_p (cfun)
+ || ((optimize_function_for_size_p (cfun)
+ || (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC))
&& ((!TARGET_SSE_MATH
&& standard_80387_constant_p (operands[1]) > 0)
|| (TARGET_SSE_MATH
