I noticed that the switch code on PowerPC little endian systems (with medium
code mode) did not follow the ABI in terms of page 69:
Table 2.36. Position-Independent Switch Code for Small/Medium Models
(preferred, with TOC-relative addressing)
The code we currently generate is:
.section ".toc","aw"
.align 3
.LC0:
.quad .L4
.section ".text"
# ...
addis 10,2,.LC0@toc@ha
ld 10,.LC0@toc@l(10)
sldi 3,3,2
add 9,10,3
lwa 9,0(9)
add 9,9,10
mtctr 9
bctr
.L4:
.long .L2-.L4
.long .L12-.L4
.long .L11-.L4
.long .L10-.L4
.long .L9-.L4
.long .L8-.L4
.long .L7-.L4
.long .L6-.L4
.long .L5-.L4
.long .L3-.L4
While the suggested code would be something like:
addis 10,2,.L4@toc@ha
addi 10,10,.L4@toc@l
sldi 3,3,2
lwax 9,10,3
add 9,9,10
mtctr 9
bctr
.p2align 2
.align 2
.L4:
.long .L2-.L4
.long .L12-.L4
.long .L11-.L4
.long .L10-.L4
.long .L9-.L4
.long .L8-.L4
.long .L7-.L4
.long .L6-.L4
.long .L5-.L4
.long .L3-.L4
This patch adds an insn to load a LABEL_REF into a GPR. This is needed so the
FWPROP1 pass can convert the load the of the label address from the TOC to a
direct load to a GPR.
While working on the patch, I discovered that the LWA instruction did not
support indexed loads. This was due to it using the 'Y' constraint, which
accepts DS-form offsettable addresses, but not X-form indexed addresses. I
added the Z constraint so that the indexed form is accepted.
I am in the middle of doing spec 2006 runs on both power8 and power9 systems
with this change. So far after 2 runs out 3, I'm seeing several minor wins on
power9 (1-2%, perlbench, gcc, sjeng, sphinx3) and no regressions. On power8 I
see 3 minor wins (1-3%, perlbench, sjeng, omnetpp) and 1 minor regression (1%,
povray).
I have done bootstrap builds with/without the change and there were no
regressions in the test suite. Can I check this change into the trunk? It is
a simple enough change for back ports, if desired.
Note, I will be on vacation for 11 days starting this Saturday. I will not be
actively checking my mail in that time period. If I get the approval early
enough, I can check it in. Otherwise, somebody else can check it in if they
monitor for failure, or we can wait until I get around August 14th to check it
in.
2018-07-31 Michael Meissner <meiss...@linux.ibm.com>
* config/rs6000/predicates.md (label_ref_operand): New predicate
to recognize LABEL_REF.
* config/rs6000/rs6000.c (rs6000_output_addr_const_extra): Allow
LABEL_REF's inside of UNSPEC_TOCREL's.
* config/rs6000/rs6000.md (extendsi<mode>2): Allow reg+reg indexed
addressing.
(labelref): New insn to optimize loading a label address into
registers on a medium code system.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meiss...@linux.ibm.com, phone: +1 (978) 899-4797
Index: gcc/config/rs6000/predicates.md
===================================================================
--- gcc/config/rs6000/predicates.md (revision 263040)
+++ gcc/config/rs6000/predicates.md (working copy)
@@ -1662,6 +1662,10 @@ (define_predicate "small_toc_ref"
return GET_CODE (op) == UNSPEC && XINT (op, 1) == UNSPEC_TOCREL;
})
+;; Match a LABEL_REF operand
+(define_predicate "label_ref_operand"
+ (match_code "label_ref"))
+
;; Match the first insn (addis) in fusing the combination of addis and loads to
;; GPR registers on power8.
(define_predicate "fusion_gpr_addis"
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 263040)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -20807,7 +20807,8 @@ rs6000_output_addr_const_extra (FILE *fi
switch (XINT (x, 1))
{
case UNSPEC_TOCREL:
- gcc_checking_assert (GET_CODE (XVECEXP (x, 0, 0)) == SYMBOL_REF
+ gcc_checking_assert ((GET_CODE (XVECEXP (x, 0, 0)) == SYMBOL_REF
+ || GET_CODE (XVECEXP (x, 0, 0)) == LABEL_REF)
&& REG_P (XVECEXP (x, 0, 1))
&& REGNO (XVECEXP (x, 0, 1)) == TOC_REGISTER);
output_addr_const (file, XVECEXP (x, 0, 0));
Index: gcc/config/rs6000/rs6000.md
===================================================================
--- gcc/config/rs6000/rs6000.md (revision 263040)
+++ gcc/config/rs6000/rs6000.md (working copy)
@@ -998,7 +998,7 @@ (define_insn "extendsi<mode>2"
"=r, r, wl, wu, wj, wK, wH, wr")
(sign_extend:EXTSI (match_operand:SI 1 "lwa_operand"
- "Y, r, Z, Z, r, wK, wH, ?wIwH")))]
+ "YZ, r, Z, Z, r, wK, wH, ?wIwH")))]
""
"@
lwa%U1%X1 %0,%1
@@ -10246,6 +10246,21 @@ (define_insn_and_split "*tocref<mode>"
[(set (match_dup 0) (high:P (match_dup 1)))
(set (match_dup 0) (lo_sum:P (match_dup 0) (match_dup 1)))])
+;; This pattern exists to allow fwprop to convert the load of the LABEL_REF
+;; address for switches into using ADDIS/ADDI if we can use the TOC
+;; to get the address.
+(define_insn_and_split "*labelref"
+ [(set (match_operand:DI 0 "gpc_reg_operand" "=b*r")
+ (match_operand:DI 1 "label_ref_operand"))]
+ "TARGET_POWERPC64 && TARGET_ELF && TARGET_CMODEL == CMODEL_MEDIUM"
+ "#"
+ "&& 1"
+ [(set (match_dup 0)
+ (match_dup 2))]
+{
+ operands[2] = create_TOC_reference (operands[1], NULL_RTX);
+})
+
;; Elf specific ways of loading addresses for non-PIC code.
;; The output of this could be r0, but we make a very strong
;; preference for a base register because it will usually
