[PATCH] S/390: Implement SD<->TD conversions
Andreas Krebbel
krebbel@linux.vnet.ibm.com
Wed Sep 10 08:42:00 GMT 2014
Hi,
on S/390 we have DFP conversions for SD<->DD and DD<->TD but miss
SD<->TD. The attached patch expands the missing patterns to SD->DD->TD
and TD->DD->SD to avoid library calls being emitted.
No regressions on s390 and s390x.
Bye,
-Andreas-
2014-09-10 Andreas Krebbel <Andreas.Krebbel@de.ibm.com>
* config/s390/s390.md ("trunctdsd2", "extendsdtd2"): New
expanders.
2014-09-10 Andreas Krebbel <Andreas.Krebbel@de.ibm.com>
* gcc.target/s390/dfp-conv1.c: New testcase.
diff --git a/gcc/config/s390/s390.md b/gcc/config/s390/s390.md
index bcfe8f9..73ac0dc 100644
--- a/gcc/config/s390/s390.md
+++ b/gcc/config/s390/s390.md
@@ -4406,6 +4406,18 @@
[(set_attr "op_type" "RRF")
(set_attr "type" "ftruncsd")])
+(define_expand "trunctdsd2"
+ [(parallel
+ [(set (match_dup 3)
+ (float_truncate:DD (match_operand:TD 1 "register_operand" "")))
+ (clobber (match_scratch:TD 2 ""))])
+ (set (match_operand:SD 0 "register_operand" "")
+ (float_truncate:SD (match_dup 3)))]
+ "TARGET_HARD_DFP"
+{
+ operands[3] = gen_reg_rtx (DDmode);
+})
+
;
; extend(sf|df)(df|tf)2 instruction pattern(s).
;
@@ -4442,6 +4454,16 @@
[(set_attr "op_type" "RRF")
(set_attr "type" "fsimptf")])
+(define_expand "extendsdtd2"
+ [(set (match_dup 2)
+ (float_extend:DD (match_operand:SD 1 "register_operand" "")))
+ (set (match_operand:TD 0 "register_operand" "")
+ (float_extend:TD (match_dup 2)))]
+ "TARGET_HARD_DFP"
+{
+ operands[2] = gen_reg_rtx (DDmode);
+})
+
; Binary Floating Point - load fp integer
; Expanders for: floor, btrunc, round, ceil, and nearbyint
diff --git a/gcc/testsuite/gcc.target/s390/dfp-conv1.c b/gcc/testsuite/gcc.target/s390/dfp-conv1.c
new file mode 100644
index 0000000..3c553a3
--- /dev/null
+++ b/gcc/testsuite/gcc.target/s390/dfp-conv1.c
@@ -0,0 +1,22 @@
+/* We do not have hardware instructions which do a direct conversion
+ between the 32 and 128 bit DFP types. But we can easily do it in
+ two steps. Older libdfp implementations require this not to call
+ into a lib in order to prevent an endless loop. */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -march=z10 -mzarch" } */
+
+_Decimal32
+foo (_Decimal128 a)
+{
+ return (_Decimal32)a;
+}
+
+_Decimal128
+bar (_Decimal32 a)
+{
+ return (_Decimal128)a;
+}
+
+/* Make sure no library call is emitted. */
+/* { dg-final { scan-assembler-not "brasl" } } */
--
1.7.11.4
More information about the Gcc-patches
mailing list