[PATCH] D31265: Fix computeKnownBits for ARMISD::CMOV
Pirama Arumuga Nainar via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 22 15:40:32 PDT 2017
pirama created this revision.
Herald added subscribers: rengolin, aemerson.
The true and false operands for the CMOV are operands 0 and 1.
ARMISelLowering.cpp::computeKnownBits was looking at operands 1 and 2
instead. This can cause CMOV instructions to be incorrectly folded into
BFI if value set by the CMOV is another CMOV, whose known bits are
computed incorrectly.
This patch fixes the issue and adds a test case.
https://reviews.llvm.org/D31265
Files:
lib/Target/ARM/ARMISelLowering.cpp
test/CodeGen/ARM/no-cmov2bfi.ll
Index: test/CodeGen/ARM/no-cmov2bfi.ll
===================================================================
--- /dev/null
+++ test/CodeGen/ARM/no-cmov2bfi.ll
@@ -0,0 +1,19 @@
+; RUN: llc < %s -mtriple=thumbv7 | FileCheck --check-prefix=CHECK-NOBFI %s
+
+declare zeroext i1 @dummy()
+
+define i8 @test(i8 %a1, i1 %c) {
+; CHECK-NOBFI-NOT: bfi
+; CHECK-NOBFI: bl dummy
+; CHECK-NOBFI: cmp r0, #0
+; CHECK-NOBFI: it ne
+; CHECK-NOBFI: orrne [[REG:r[0-9]+]], [[REG]], #8
+; CHECK-NOBFI: mov r0, [[REG]]
+
+ %1 = and i8 %a1, -9
+ %2 = select i1 %c, i8 %1, i8 %a1
+ %3 = tail call zeroext i1 @dummy()
+ %4 = or i8 %2, 8
+ %ret = select i1 %3, i8 %4, i8 %2
+ ret i8 %ret
+}
Index: lib/Target/ARM/ARMISelLowering.cpp
===================================================================
--- lib/Target/ARM/ARMISelLowering.cpp
+++ lib/Target/ARM/ARMISelLowering.cpp
@@ -11628,8 +11628,8 @@
if (Op.getOpcode() == ARMISD::CMOV) {
APInt KZ2(KnownZero.getBitWidth(), 0);
APInt KO2(KnownOne.getBitWidth(), 0);
- computeKnownBits(DAG, Op.getOperand(1), KnownZero, KnownOne);
- computeKnownBits(DAG, Op.getOperand(2), KZ2, KO2);
+ computeKnownBits(DAG, Op.getOperand(0), KnownZero, KnownOne);
+ computeKnownBits(DAG, Op.getOperand(1), KZ2, KO2);
KnownZero &= KZ2;
KnownOne &= KO2;
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D31265.92725.patch
Type: text/x-patch
Size: 1325 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170322/31093e0f/attachment.bin>
More information about the llvm-commits
mailing list