Fix PR8365 by adding a more specialized Pat that checks if an 'and' with
8 bit constants can be used.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@116403 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 6bfd605..5f13885 100644
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -20,6 +20,11 @@
return getI32Imm((unsigned)N->getZExtValue());
}]>;
+def GetLo8XForm : SDNodeXForm<imm, [{
+ // Transformation function: get the low 8 bits.
+ return getI8Imm((uint8_t)N->getZExtValue());
+}]>;
+
//===----------------------------------------------------------------------===//
// Random Pseudo Instructions.
@@ -1106,9 +1111,19 @@
def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
(SUB64mi32 addr:$dst, 0xffffffff80000000)>;
-// Use a 32-bit and with implicit zero-extension instead of a 64-bit and if it
-// has an immediate with at least 32 bits of leading zeros, to avoid needing to
-// materialize that immediate in a register first.
+// To avoid needing to materialize an immediate in a register, use a 32-bit and
+// with implicit zero-extension instead of a 64-bit and if the immediate has at
+// least 32 bits of leading zeros. If in addition the last 32 bits can be
+// represented with a sign extension of a 8 bit constant, use that.
+
+def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
+ (SUBREG_TO_REG
+ (i64 0),
+ (AND32ri8
+ (EXTRACT_SUBREG GR64:$src, sub_32bit),
+ (i32 (GetLo8XForm imm:$imm))),
+ sub_32bit)>;
+
def : Pat<(and GR64:$src, i64immZExt32:$imm),
(SUBREG_TO_REG
(i64 0),
diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index ed5e407..14d673a 100644
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -473,6 +473,11 @@
return (uint64_t)N->getZExtValue() == (uint32_t)N->getZExtValue();
}]>;
+def i64immZExt32SExt8 : PatLeaf<(i64 imm), [{
+ uint64_t v = N->getZExtValue();
+ return v == (uint32_t)v && (int32_t)v == (int8_t)v;
+}]>;
+
// Helper fragments for loads.
// It's always safe to treat a anyext i16 load as a i32 load if the i16 is
// known to be 32-bit aligned or better. Ditto for i8 to i16.