[llvm] [AMDGPU] Combine build_vector patterns for i16 and f16. NFCI. (PR #91806)

[Stanislav Mekhanoshin via llvm-commits]

https://github.com/rampitec updated https://github.com/llvm/llvm-project/pull/91806

>From 8bb76838520d4da2d178de489e78f17c8bd30b02 Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Fri, 10 May 2024 13:43:03 -0700
Subject: [PATCH 1/2] [AMDGPU] Combine build_vector patterns for i16 and f16.
 NFCI.

---
 llvm/lib/Target/AMDGPU/SIInstructions.td | 48 ++++++++----------------
 1 file changed, 16 insertions(+), 32 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index cca8d96f29c0f..97ff61bde45ec 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -3162,49 +3162,34 @@ def : GCNPat <
   (v2f16 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), VGPR_32:$src1))
 >;
 
-def : GCNPat <
-  (v2i16 (UniformBinFrag<build_vector> (i16 SReg_32:$src0), (i16 undef))),
-  (COPY_TO_REGCLASS SReg_32:$src0, SReg_32)
->;
-
-def : GCNPat <
-  (v2i16 (DivergentBinFrag<build_vector> (i16 VGPR_32:$src0), (i16 undef))),
-  (COPY_TO_REGCLASS VGPR_32:$src0, VGPR_32)
->;
+foreach Ty = [i16, f16] in {
 
-def : GCNPat <
-  (v2f16 (build_vector f16:$src0, (f16 undef))),
-  (COPY $src0)
->;
+defvar vecTy = !cond(!eq(Ty, i16)  : v2i16,
+                     !eq(Ty, f16)  : v2f16);
 
 def : GCNPat <
-  (v2i16 (UniformBinFrag<build_vector> (i16 undef), (i16 SReg_32:$src1))),
-  (S_LSHL_B32 SReg_32:$src1, (i32 16))
+  (vecTy (UniformBinFrag<build_vector> (Ty SReg_32:$src0), (Ty undef))),
+  (COPY_TO_REGCLASS SReg_32:$src0, SReg_32)
 >;
 
 def : GCNPat <
-  (v2i16 (DivergentBinFrag<build_vector> (i16 undef), (i16 VGPR_32:$src1))),
-  (v2i16 (V_LSHLREV_B32_e64 (i32 16), VGPR_32:$src1))
+  (vecTy (DivergentBinFrag<build_vector> (Ty VGPR_32:$src0), (Ty undef))),
+  (COPY_TO_REGCLASS VGPR_32:$src0, VGPR_32)
 >;
 
-
 def : GCNPat <
-  (v2f16 (UniformBinFrag<build_vector> (f16 undef), (f16 SReg_32:$src1))),
+  (vecTy (UniformBinFrag<build_vector> (Ty undef), (Ty SReg_32:$src1))),
   (S_LSHL_B32 SReg_32:$src1, (i32 16))
 >;
 
 def : GCNPat <
-  (v2f16 (DivergentBinFrag<build_vector> (f16 undef), (f16 VGPR_32:$src1))),
-  (v2f16 (V_LSHLREV_B32_e64 (i32 16), VGPR_32:$src1))
+  (vecTy (DivergentBinFrag<build_vector> (Ty undef), (Ty VGPR_32:$src1))),
+  (vecTy (V_LSHLREV_B32_e64 (i32 16), VGPR_32:$src1))
 >;
+} // End foreach Ty = ...
 }
 
 let SubtargetPredicate = HasVOP3PInsts in {
-def : GCNPat <
-  (v2i16 (UniformBinFrag<build_vector> (i16 SReg_32:$src0), (i16 SReg_32:$src1))),
-  (S_PACK_LL_B32_B16 SReg_32:$src0, SReg_32:$src1)
->;
-
 def : GCNPat <
   (v2i16 (DivergentBinFrag<build_vector> (i16 VGPR_32:$src0), (i16 VGPR_32:$src1))),
   (v2i16 (V_LSHL_OR_B32_e64 $src1, (i32 16), (i32 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), $src0))))
@@ -3223,18 +3208,17 @@ def : GCNPat <
   (S_PACK_HH_B32_B16 SReg_32:$src0, SReg_32:$src1)
 >;
 
-def : GCNPat <
-  (v2f16 (UniformBinFrag<build_vector> (f16 SReg_32:$src0), (f16 SReg_32:$src1))),
-  (S_PACK_LL_B32_B16 SReg_32:$src0, SReg_32:$src1)
->;
-
-
 
 foreach Ty = [i16, f16] in {
 
 defvar vecTy = !if(!eq(Ty, i16), v2i16, v2f16);
 defvar immzeroTy = !if(!eq(Ty, i16), immzero, fpimmzero);
 
+def : GCNPat <
+  (vecTy (UniformBinFrag<build_vector> (Ty SReg_32:$src0), (Ty SReg_32:$src1))),
+  (S_PACK_LL_B32_B16 SReg_32:$src0, SReg_32:$src1)
+>;
+
 // Take the lower 16 bits from each VGPR_32 and concat them
 def : GCNPat <
   (vecTy (DivergentBinFrag<build_vector> (Ty VGPR_32:$a), (Ty VGPR_32:$b))),

>From f5e8e2e5e9dc8cc5f0cdd2d1c59460fb7cc51a6b Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Fri, 10 May 2024 14:37:29 -0700
Subject: [PATCH 2/2] Use vecTy.ElementType instead of switch

---
 llvm/lib/Target/AMDGPU/SIInstructions.td | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 97ff61bde45ec..8be3818d7d249 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -3162,10 +3162,9 @@ def : GCNPat <
   (v2f16 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), VGPR_32:$src1))
 >;
 
-foreach Ty = [i16, f16] in {
+foreach vecTy = [v2i16, v2f16] in {
 
-defvar vecTy = !cond(!eq(Ty, i16)  : v2i16,
-                     !eq(Ty, f16)  : v2f16);
+defvar Ty = vecTy.ElementType;
 
 def : GCNPat <
   (vecTy (UniformBinFrag<build_vector> (Ty SReg_32:$src0), (Ty undef))),
@@ -3209,9 +3208,9 @@ def : GCNPat <
 >;
 
 
-foreach Ty = [i16, f16] in {
+foreach vecTy = [v2i16, v2f16] in {
 
-defvar vecTy = !if(!eq(Ty, i16), v2i16, v2f16);
+defvar Ty = vecTy.ElementType;
 defvar immzeroTy = !if(!eq(Ty, i16), immzero, fpimmzero);
 
 def : GCNPat <