summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td87
-rw-r--r--llvm/test/CodeGen/Hexagon/extload-combine.ll80
2 files changed, 167 insertions, 0 deletions
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td b/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td
index 022a7f61365..fee83fb8110 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td
@@ -3188,6 +3188,93 @@ def STriw_offset_ext_V4 : STInst<(outs),
(add IntRegs:$src1, u6_2ImmPred:$src2))]>,
Requires<[HasV4T]>;
+def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))),
+ (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTLZ64_rr DoubleRegs:$src1))))>,
+ Requires<[HasV4T]>;
+
+def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))),
+ (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTTZ64_rr DoubleRegs:$src1))))>,
+ Requires<[HasV4T]>;
+
+
+// i8 -> i64 loads
+// We need a complexity of 120 here to overide preceeding handling of
+// zextloadi8.
+let Predicates = [HasV4T], AddedComplexity = 120 in {
+def: Pat <(i64 (extloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 tglobaladdr:$addr)))>;
+
+def: Pat <(i64 (zextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 tglobaladdr:$addr)))>;
+
+def: Pat <(i64 (sextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (SXTW (LDrib_abs_V4 tglobaladdr:$addr)))>;
+
+def: Pat <(i64 (extloadi8 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 FoldGlobalAddr:$addr)))>;
+
+def: Pat <(i64 (zextloadi8 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 FoldGlobalAddr:$addr)))>;
+
+def: Pat <(i64 (sextloadi8 FoldGlobalAddr:$addr)),
+ (i64 (SXTW (LDrib_abs_V4 FoldGlobalAddr:$addr)))>;
+}
+// i16 -> i64 loads
+// We need a complexity of 120 here to overide preceeding handling of
+// zextloadi16.
+let AddedComplexity = 120 in {
+def: Pat <(i64 (extloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (zextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (sextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (SXTW (LDrih_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (extloadi16 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (zextloadi16 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (sextloadi16 FoldGlobalAddr:$addr)),
+ (i64 (SXTW (LDrih_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+}
+// i32->i64 loads
+// We need a complexity of 120 here to overide preceeding handling of
+// zextloadi32.
+let AddedComplexity = 120 in {
+def: Pat <(i64 (extloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (zextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (sextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))),
+ (i64 (SXTW (LDriw_abs_V4 tglobaladdr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (extloadi32 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (zextloadi32 FoldGlobalAddr:$addr)),
+ (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+
+def: Pat <(i64 (sextloadi32 FoldGlobalAddr:$addr)),
+ (i64 (SXTW (LDriw_abs_V4 FoldGlobalAddr:$addr)))>,
+ Requires<[HasV4T]>;
+}
// Indexed store double word - global address.
// memw(Rs+#u6:2)=#S8
diff --git a/llvm/test/CodeGen/Hexagon/extload-combine.ll b/llvm/test/CodeGen/Hexagon/extload-combine.ll
new file mode 100644
index 00000000000..b3b8bf07032
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/extload-combine.ll
@@ -0,0 +1,80 @@
+; RUN: llc -march=hexagon -mcpu=hexagonv4 -O3 < %s | FileCheck %s
+; Check that the combine/stxw instructions are being generated.
+; In case of combine one of the operand should be 0 and another should be
+; the output of absolute addressing load instruction.
+
+@a = external global i16
+@b = external global i16
+@c = external global i16
+@char_a = external global i8
+@char_b = external global i8
+@char_c = external global i8
+@int_a = external global i32
+@int_b = external global i32
+@int_c = external global i32
+
+; Function Attrs: nounwind
+define i64 @short_test1() #0 {
+; CHECK: [[VAR:r[0-9]+]]{{ *}}={{ *}}memuh(##
+; CHECK: combine(#0, [[VAR]])
+entry:
+ store i16 0, i16* @a, align 2
+ %0 = load i16* @b, align 2
+ %conv2 = zext i16 %0 to i64
+ ret i64 %conv2
+}
+
+; Function Attrs: nounwind
+define i64 @short_test2() #0 {
+; CHECK: [[VAR1:r[0-9]+]]{{ *}}={{ *}}memh(##
+; CHECK: sxtw([[VAR1]])
+entry:
+ store i16 0, i16* @a, align 2
+ %0 = load i16* @c, align 2
+ %conv2 = sext i16 %0 to i64
+ ret i64 %conv2
+}
+
+; Function Attrs: nounwind
+define i64 @char_test1() #0 {
+; CHECK: [[VAR2:r[0-9]+]]{{ *}}={{ *}}memub(##
+; CHECK: combine(#0, [[VAR2]])
+entry:
+ store i8 0, i8* @char_a, align 1
+ %0 = load i8* @char_b, align 1
+ %conv2 = zext i8 %0 to i64
+ ret i64 %conv2
+}
+
+; Function Attrs: nounwind
+define i64 @char_test2() #0 {
+; CHECK: [[VAR3:r[0-9]+]]{{ *}}={{ *}}memb(##
+; CHECK: sxtw([[VAR3]])
+entry:
+ store i8 0, i8* @char_a, align 1
+ %0 = load i8* @char_c, align 1
+ %conv2 = sext i8 %0 to i64
+ ret i64 %conv2
+}
+
+; Function Attrs: nounwind
+define i64 @int_test1() #0 {
+; CHECK: [[VAR4:r[0-9]+]]{{ *}}={{ *}}memw(##
+; CHECK: combine(#0, [[VAR4]])
+entry:
+ store i32 0, i32* @int_a, align 4
+ %0 = load i32* @int_b, align 4
+ %conv = zext i32 %0 to i64
+ ret i64 %conv
+}
+
+; Function Attrs: nounwind
+define i64 @int_test2() #0 {
+; CHECK: [[VAR5:r[0-9]+]]{{ *}}={{ *}}memw(##
+; CHECK: sxtw([[VAR5]])
+entry:
+ store i32 0, i32* @int_a, align 4
+ %0 = load i32* @int_c, align 4
+ %conv = sext i32 %0 to i64
+ ret i64 %conv
+}
OpenPOWER on IntegriCloud