mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-22 18:54:02 +01:00
0c16dcd701
This provides an optimized implementation of SADDO/SSUBO/UADDO/USUBO as well as ADDCARRY/SUBCARRY on top of the new CC implementation. In particular, multi-word arithmetic now uses UADDO/ADDCARRY instead of the old ADDC/ADDE logic, which means we no longer need to use "glue" links for those instructions. This also allows making full use of the memory-based instructions like ALSI, which couldn't be recognized due to limitations in the DAG matcher previously. Also, the llvm.sadd.with.overflow et.al. intrinsincs now expand to directly using the ADD instructions and checking for a CC 3 result. llvm-svn: 331203
83 lines
2.2 KiB
LLVM
83 lines
2.2 KiB
LLVM
; Test the three-operand form of 64-bit addition.
|
|
;
|
|
; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z196 | FileCheck %s
|
|
|
|
declare i64 @foo(i64, i64, i64)
|
|
|
|
; Check SLGRK.
|
|
define i64 @f1(i64 %dummy, i64 %a, i64 %b, i64 *%flag) {
|
|
; CHECK-LABEL: f1:
|
|
; CHECK: slgrk %r2, %r3, %r4
|
|
; CHECK: ipm [[REG1:%r[0-5]]]
|
|
; CHECK: afi [[REG1]], -536870912
|
|
; CHECK: risbg [[REG2:%r[0-5]]], [[REG1]], 63, 191, 33
|
|
; CHECK: stg [[REG2]], 0(%r5)
|
|
; CHECK: br %r14
|
|
%t = call {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
|
|
%val = extractvalue {i64, i1} %t, 0
|
|
%obit = extractvalue {i64, i1} %t, 1
|
|
%ext = zext i1 %obit to i64
|
|
store i64 %ext, i64 *%flag
|
|
ret i64 %val
|
|
}
|
|
|
|
; Check using the overflow result for a branch.
|
|
define i64 @f2(i64 %dummy, i64 %a, i64 %b) {
|
|
; CHECK-LABEL: f2:
|
|
; CHECK: slgrk %r2, %r3, %r4
|
|
; CHECK-NEXT: bnler %r14
|
|
; CHECK: lghi %r2, 0
|
|
; CHECK: jg foo@PLT
|
|
%t = call {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
|
|
%val = extractvalue {i64, i1} %t, 0
|
|
%obit = extractvalue {i64, i1} %t, 1
|
|
br i1 %obit, label %call, label %exit
|
|
|
|
call:
|
|
%res = tail call i64 @foo(i64 0, i64 %a, i64 %b)
|
|
ret i64 %res
|
|
|
|
exit:
|
|
ret i64 %val
|
|
}
|
|
|
|
; ... and the same with the inverted direction.
|
|
define i64 @f3(i64 %dummy, i64 %a, i64 %b) {
|
|
; CHECK-LABEL: f3:
|
|
; CHECK: slgrk %r2, %r3, %r4
|
|
; CHECK-NEXT: bler %r14
|
|
; CHECK: lghi %r2, 0
|
|
; CHECK: jg foo@PLT
|
|
%t = call {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
|
|
%val = extractvalue {i64, i1} %t, 0
|
|
%obit = extractvalue {i64, i1} %t, 1
|
|
br i1 %obit, label %exit, label %call
|
|
|
|
call:
|
|
%res = tail call i64 @foo(i64 0, i64 %a, i64 %b)
|
|
ret i64 %res
|
|
|
|
exit:
|
|
ret i64 %val
|
|
}
|
|
|
|
; Check that we can still use SLGR in obvious cases.
|
|
define i64 @f4(i64 %a, i64 %b, i64 *%flag) {
|
|
; CHECK-LABEL: f4:
|
|
; CHECK: slgr %r2, %r3
|
|
; CHECK: ipm [[REG1:%r[0-5]]]
|
|
; CHECK: afi [[REG1]], -536870912
|
|
; CHECK: risbg [[REG2:%r[0-5]]], [[REG1]], 63, 191, 33
|
|
; CHECK: stg [[REG2]], 0(%r4)
|
|
; CHECK: br %r14
|
|
%t = call {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
|
|
%val = extractvalue {i64, i1} %t, 0
|
|
%obit = extractvalue {i64, i1} %t, 1
|
|
%ext = zext i1 %obit to i64
|
|
store i64 %ext, i64 *%flag
|
|
ret i64 %val
|
|
}
|
|
|
|
declare {i64, i1} @llvm.usub.with.overflow.i64(i64, i64) nounwind readnone
|
|
|