[clang] [llvm] [Clang] Make sdot builtins available to SME (PR #77792)

via cfe-commits cfe-commits at lists.llvm.org
Thu Jan 11 08:08:08 PST 2024 

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-aarch64

Author: Sander de Smalen (sdesmalen-arm)

<details>
<summary>Changes</summary>

See the specification for more details:
* https://github.com/ARM-software/acle/blob/main/main/acle.md#udot-sdot-fdot-vectors
* https://github.com/ARM-software/acle/blob/main/main/acle.md#udot-sdot-fdot-indexed

---
Full diff: https://github.com/llvm/llvm-project/pull/77792.diff


3 Files Affected:

- (modified) clang/include/clang/Basic/arm_sve.td (+7-7) 
- (modified) clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_dot.c (+15-6) 
- (modified) llvm/test/CodeGen/AArch64/sve2p1-intrinsics-dots.ll (+1) 


``````````diff
diff --git a/clang/include/clang/Basic/arm_sve.td b/clang/include/clang/Basic/arm_sve.td
index 6f35e25617adea..5ba7598ac61633 100644
--- a/clang/include/clang/Basic/arm_sve.td
+++ b/clang/include/clang/Basic/arm_sve.td
@@ -2039,13 +2039,13 @@ let TargetGuard = "sve2p1|sme2" in {
   defm STNT1 : MultiVecStore<"stnt1">;
 }
 
-let TargetGuard = "sve2p1" in {
-def SVDOT_X2_S : SInst<"svdot[_{d}_{2}]", "ddhh", "i",  MergeNone, "aarch64_sve_sdot_x2", [], []>;
-def SVDOT_X2_U : SInst<"svdot[_{d}_{2}]", "ddhh", "Ui", MergeNone, "aarch64_sve_udot_x2", [], []>;
-def SVDOT_X2_F : SInst<"svdot[_{d}_{2}]", "ddhh", "f",  MergeNone, "aarch64_sve_fdot_x2", [], []>;
-def SVDOT_LANE_X2_S : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "i",  MergeNone, "aarch64_sve_sdot_lane_x2", [], [ImmCheck<3, ImmCheck0_3>]>;
-def SVDOT_LANE_X2_U : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "Ui", MergeNone, "aarch64_sve_udot_lane_x2", [], [ImmCheck<3, ImmCheck0_3>]>;
-def SVDOT_LANE_X2_F : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "f",  MergeNone, "aarch64_sve_fdot_lane_x2", [], [ImmCheck<3, ImmCheck0_3>]>;
+let TargetGuard = "sve2p1|sme2" in {
+def SVDOT_X2_S : SInst<"svdot[_{d}_{2}]", "ddhh", "i",  MergeNone, "aarch64_sve_sdot_x2", [IsStreamingOrSVE2p1], []>;
+def SVDOT_X2_U : SInst<"svdot[_{d}_{2}]", "ddhh", "Ui", MergeNone, "aarch64_sve_udot_x2", [IsStreamingOrSVE2p1], []>;
+def SVDOT_X2_F : SInst<"svdot[_{d}_{2}]", "ddhh", "f",  MergeNone, "aarch64_sve_fdot_x2", [IsStreamingOrSVE2p1], []>;
+def SVDOT_LANE_X2_S : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "i",  MergeNone, "aarch64_sve_sdot_lane_x2", [IsStreamingOrSVE2p1], [ImmCheck<3, ImmCheck0_3>]>;
+def SVDOT_LANE_X2_U : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "Ui", MergeNone, "aarch64_sve_udot_lane_x2", [IsStreamingOrSVE2p1], [ImmCheck<3, ImmCheck0_3>]>;
+def SVDOT_LANE_X2_F : SInst<"svdot_lane[_{d}_{2}]", "ddhhi", "f",  MergeNone, "aarch64_sve_fdot_lane_x2", [IsStreamingOrSVE2p1], [ImmCheck<3, ImmCheck0_3>]>;
 }
 
 let TargetGuard = "sve2p1|sme2" in {
diff --git a/clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_dot.c b/clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_dot.c
index 035ba244f9441e..97099c1f1061c7 100644
--- a/clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_dot.c
+++ b/clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_dot.c
@@ -2,11 +2,20 @@
 // REQUIRES: aarch64-registered-target
 // RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sve2p1 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
 // RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sve2p1 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - -x c++ %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CPP-CHECK
+// RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme2 -DTEST_SME2 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
+// RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme2 -DTEST_SME2 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - -x c++ %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CPP-CHECK
 // RUN: %clang_cc1 -fclang-abi-compat=latest -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve2p1 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
 // RUN: %clang_cc1 -fclang-abi-compat=latest -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve2p1 -S -disable-O0-optnone -Werror -Wall -emit-llvm -o - -x c++ %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CPP-CHECK
 // RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sve2p1 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
+// RUN: %clang_cc1 -fclang-abi-compat=latest -triple aarch64-none-linux-gnu -target-feature +sme2 -DTEST_SME2 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
 #include <arm_sve.h>
 
+#ifndef TEST_SME2
+#define ATTR
+#else
+#define ATTR __arm_streaming
+#endif
+
 #ifdef SVE_OVERLOADED_FORMS
 // A simple used,unused... macro, long enough to represent any SVE builtin.
 #define SVE_ACLE_FUNC(A1,A2_UNUSED,A3) A1##A3
@@ -24,7 +33,7 @@
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sve.sdot.x2.nxv4i32(<vscale x 4 x i32> [[OP1:%.*]], <vscale x 8 x i16> [[OP2:%.*]], <vscale x 8 x i16> [[OP3:%.*]])
 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 //
-svint32_t test_svdot_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3)
+svint32_t test_svdot_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot,_s32_s16,)(op1, op2, op3);
 }
@@ -39,7 +48,7 @@ svint32_t test_svdot_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3)
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sve.udot.x2.nxv4i32(<vscale x 4 x i32> [[OP1:%.*]], <vscale x 8 x i16> [[OP2:%.*]], <vscale x 8 x i16> [[OP3:%.*]])
 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 //
-svuint32_t test_svdot_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3)
+svuint32_t test_svdot_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot,_u32_u16,)(op1, op2, op3);
 }
@@ -54,7 +63,7 @@ svuint32_t test_svdot_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3)
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sve.fdot.x2.nxv4f32(<vscale x 4 x float> [[OP1:%.*]], <vscale x 8 x half> [[OP2:%.*]], <vscale x 8 x half> [[OP3:%.*]])
 // CPP-CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
 //
-svfloat32_t test_svdot_f32_x2(svfloat32_t op1, svfloat16_t op2, svfloat16_t op3)
+svfloat32_t test_svdot_f32_x2(svfloat32_t op1, svfloat16_t op2, svfloat16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot,_f32_f16,)(op1, op2, op3);
 }
@@ -71,7 +80,7 @@ svfloat32_t test_svdot_f32_x2(svfloat32_t op1, svfloat16_t op2, svfloat16_t op3)
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sve.sdot.lane.x2.nxv4i32(<vscale x 4 x i32> [[OP1:%.*]], <vscale x 8 x i16> [[OP2:%.*]], <vscale x 8 x i16> [[OP3:%.*]], i32 3)
 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 //
-svint32_t test_svdot_lane_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3)
+svint32_t test_svdot_lane_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot_lane,_s32_s16,)(op1, op2, op3, 3);
 }
@@ -86,7 +95,7 @@ svint32_t test_svdot_lane_s32_x2(svint32_t op1, svint16_t op2, svint16_t op3)
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sve.udot.lane.x2.nxv4i32(<vscale x 4 x i32> [[OP1:%.*]], <vscale x 8 x i16> [[OP2:%.*]], <vscale x 8 x i16> [[OP3:%.*]], i32 3)
 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 //
-svuint32_t test_svdot_lane_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3)
+svuint32_t test_svdot_lane_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot_lane,_u32_u16,)(op1, op2, op3, 3);
 }
@@ -101,7 +110,7 @@ svuint32_t test_svdot_lane_u32_x2(svuint32_t op1, svuint16_t op2, svuint16_t op3
 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sve.fdot.lane.x2.nxv4f32(<vscale x 4 x float> [[OP1:%.*]], <vscale x 8 x half> [[OP2:%.*]], <vscale x 8 x half> [[OP3:%.*]], i32 3)
 // CPP-CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
 //
-svfloat32_t test_svdot_lane_f32_x2(svfloat32_t op1, svfloat16_t op2, svfloat16_t op3)
+svfloat32_t test_svdot_lane_f32_x2(svfloat32_t op1, svfloat16_t op2, svfloat16_t op3) ATTR
 {
   return SVE_ACLE_FUNC(svdot_lane,_f32_f16,)(op1, op2, op3, 3);
 }
diff --git a/llvm/test/CodeGen/AArch64/sve2p1-intrinsics-dots.ll b/llvm/test/CodeGen/AArch64/sve2p1-intrinsics-dots.ll
index d8c8043df59c3f..e54a93bdaf6580 100644
--- a/llvm/test/CodeGen/AArch64/sve2p1-intrinsics-dots.ll
+++ b/llvm/test/CodeGen/AArch64/sve2p1-intrinsics-dots.ll
@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2p1 < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme2 < %s | FileCheck %s
 
 define <vscale x 4 x i32> @sdot_x2(<vscale x 4 x i32> %zda, <vscale x 8 x i16> %zn, <vscale x 8 x i16> %zm) {
 ; CHECK-LABEL: sdot_x2:

``````````

</details>


https://github.com/llvm/llvm-project/pull/77792

More information about the cfe-commits mailing list