This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH, ARM] Allow high regs when optimising for size on Thumb-2
- From: Julian Brown <julian at codesourcery dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: paul at codesourcery dot com
- Date: Thu, 1 Apr 2010 09:32:36 +0100
- Subject: [PATCH, ARM] Allow high regs when optimising for size on Thumb-2
Hi,
This is a size optimisation for Thumb-2. With the patch applied, high
registers are permitted when optimising for size for Thumb-2 (but
still not for Thumb-1). Also SImode loads and stores have their
constraints split into high/low register halves with the high parts not
taken into account when choosing register preferences, to reflect the
greater expense of the high-register alternatives (we don't want to
influence the compiler to use them unless necessary).
Tested on mainline (with flags to generate Thumb-2 code). (We've
also been carrying this in our internal trees for quite a while). OK to
apply?
Julian
ChangeLog
2007-07-05 Mark Shinwell <shinwell@codesourcery.com>
gcc/
* config/arm/thumb2.md (thumb2_movsi_insn): Split ldr and
str alternatives according to use of high and low regs.
* config/arm/vfp.md (thumb2_movsi_vfp): Likewise.
* config/arm/arm.h (CONDITIONAL_REGISTER_USAGE): Use high
regs when optimizing for size on Thumb-2.
Index: gcc/config/arm/thumb2.md
===================================================================
--- gcc/config/arm/thumb2.md (revision 157802)
+++ gcc/config/arm/thumb2.md (working copy)
@@ -223,9 +223,14 @@
(set_attr "neg_pool_range" "*,*,*,0,*")]
)
+;; We have two alternatives here for memory loads (and similarly for stores)
+;; to reflect the fact that the permissible constant pool ranges differ
+;; between ldr instructions taking low regs and ldr instructions taking high
+;; regs. The high register alternatives are not taken into account when
+;; choosing register preferences in order to reflect their expense.
(define_insn "*thumb2_movsi_insn"
- [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,rk,m")
- (match_operand:SI 1 "general_operand" "rk ,I,K,j,mi,rk"))]
+ [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,l ,*hk,m,*m")
+ (match_operand:SI 1 "general_operand" "rk ,I,K,j,mi,*mi,l,*hk"))]
"TARGET_THUMB2 && ! TARGET_IWMMXT
&& !(TARGET_HARD_FLOAT && TARGET_VFP)
&& ( register_operand (operands[0], SImode)
@@ -236,11 +241,13 @@
mvn%?\\t%0, #%B1
movw%?\\t%0, %1
ldr%?\\t%0, %1
+ ldr%?\\t%0, %1
+ str%?\\t%1, %0
str%?\\t%1, %0"
- [(set_attr "type" "*,*,*,*,load1,store1")
+ [(set_attr "type" "*,*,*,*,load1,load1,store1,store1")
(set_attr "predicable" "yes")
- (set_attr "pool_range" "*,*,*,*,4096,*")
- (set_attr "neg_pool_range" "*,*,*,*,0,*")]
+ (set_attr "pool_range" "*,*,*,*,1020,4096,*,*")
+ (set_attr "neg_pool_range" "*,*,*,*,0,0,*,*")]
)
(define_insn "tls_load_dot_plus_four"
Index: gcc/config/arm/arm.h
===================================================================
--- gcc/config/arm/arm.h (revision 157802)
+++ gcc/config/arm/arm.h (working copy)
@@ -771,12 +771,11 @@ extern int arm_structure_size_boundary;
fixed_regs[regno] = call_used_regs[regno] = 1; \
} \
\
- if (TARGET_THUMB && optimize_size) \
- { \
- /* When optimizing for size, it's better not to use \
- the HI regs, because of the overhead of stacking \
- them. */ \
- /* ??? Is this still true for thumb2? */ \
+ if (TARGET_THUMB1 && optimize_size) \
+ { \
+ /* When optimizing for size on Thumb-1, it's better not \
+ to use the HI regs, because of the overhead of \
+ stacking them. */ \
for (regno = FIRST_HI_REGNUM; \
regno <= LAST_HI_REGNUM; ++regno) \
fixed_regs[regno] = call_used_regs[regno] = 1; \
Index: gcc/config/arm/vfp.md
===================================================================
--- gcc/config/arm/vfp.md (revision 157802)
+++ gcc/config/arm/vfp.md (working copy)
@@ -86,9 +86,11 @@
(set_attr "neg_pool_range" "*,*,*,*,4084,*,*,*,*,1008,*")]
)
+;; See thumb2.md:thumb2_movsi_insn for an explanation of the split
+;; high/low register alternatives for loads and stores here.
(define_insn "*thumb2_movsi_vfp"
- [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,rk,m,*t,r, *t,*t, *Uv")
- (match_operand:SI 1 "general_operand" "rk, I,K,j,mi,rk,r,*t,*t,*Uvi,*t"))]
+ [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,l,*hk,m,*m,*t,r, *t,*t, *Uv")
+ (match_operand:SI 1 "general_operand" "rk, I,K,j,mi,*mi,l,*hk,r,*t,*t,*Uvi,*t"))]
"TARGET_THUMB2 && TARGET_VFP && TARGET_HARD_FLOAT
&& ( s_register_operand (operands[0], SImode)
|| s_register_operand (operands[1], SImode))"
@@ -102,25 +104,27 @@
case 3:
return \"movw%?\\t%0, %1\";
case 4:
- return \"ldr%?\\t%0, %1\";
case 5:
- return \"str%?\\t%1, %0\";
+ return \"ldr%?\\t%0, %1\";
case 6:
- return \"fmsr%?\\t%0, %1\\t%@ int\";
case 7:
- return \"fmrs%?\\t%0, %1\\t%@ int\";
+ return \"str%?\\t%1, %0\";
case 8:
+ return \"fmsr%?\\t%0, %1\\t%@ int\";
+ case 9:
+ return \"fmrs%?\\t%0, %1\\t%@ int\";
+ case 10:
return \"fcpys%?\\t%0, %1\\t%@ int\";
- case 9: case 10:
+ case 11: case 12:
return output_move_vfp (operands);
default:
gcc_unreachable ();
}
"
[(set_attr "predicable" "yes")
- (set_attr "type" "*,*,*,*,load1,store1,r_2_f,f_2_r,fcpys,f_load,f_store")
- (set_attr "pool_range" "*,*,*,*,4096,*,*,*,*,1020,*")
- (set_attr "neg_pool_range" "*,*,*,*, 0,*,*,*,*,1008,*")]
+ (set_attr "type" "*,*,*,*,load1,load1,store1,store1,r_2_f,f_2_r,fcpys,f_load,f_store")
+ (set_attr "pool_range" "*,*,*,*,1020,4096,*,*,*,*,*,1020,*")
+ (set_attr "neg_pool_range" "*,*,*,*, 0, 0,*,*,*,*,*,1008,*")]
)