[AMDGPU][SDAG] Handle ISD::PTRADD in VOP3 patterns

This patch mirrors similar patterns for ISD::ADD. The main difference is
that ISD::ADD is commutative, so that a pattern definition for, e.g.,
(add (mul x, y), z), automatically also handles (add z, (mul x, y)).
ISD::PTRADD is not commutative, so we would need to handle these cases
explicitly. This patch only implements (ptradd z, (op x, y)) patterns,
where the nested operation (shift or multiply) is the offset of the
ptradd (i.e., the right operand), since base pointers that are the
result of a shift or multiply seem less likely.

For SWDEV-516125.

Author: ritter-x2a

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -554,12 +554,13 @@ let OtherPredicates = [isGFX10Plus, Has16BitInsts], True16Predicate = NotHasTrue
   defm: Ternary_i16_Pats_gfx9<mul, add, V_MAD_U16_gfx9_e64>;
 } // End OtherPredicates = [isGFX10Plus, Has16BitInsts], True16Predicate = NotHasTrue16BitInsts
 
-class ThreeOpFragSDAG<SDPatternOperator op1, SDPatternOperator op2> : PatFrag<
+class ThreeOpFragSDAG<SDPatternOperator op1, SDPatternOperator op2, bit op1IsRight = 0> : PatFrag<
   (ops node:$x, node:$y, node:$z),
   // When the inner operation is used multiple times, selecting 3-op
   // instructions may still be beneficial -- if the other users can be
   // combined similarly. Let's be conservative for now.
-  (op2 (HasOneUseBinOp<op1> node:$x, node:$y), node:$z),
+  !if(op1IsRight, (op2 node:$z, (HasOneUseBinOp<op1> node:$x, node:$y)),
+                  (op2 (HasOneUseBinOp<op1> node:$x, node:$y), node:$z)),
   [{
     // Only use VALU ops when the result is divergent.
     if (!N->isDivergent())
@@ -586,7 +587,10 @@ class ThreeOpFragSDAG<SDPatternOperator op1, SDPatternOperator op2> : PatFrag<
   let PredicateCodeUsesOperands = 1;
 }
 
-class ThreeOpFrag<SDPatternOperator op1, SDPatternOperator op2> : ThreeOpFragSDAG<op1, op2> {
+// Matches (op2 (op1 x, y), z) if op1IsRight = 0 and
+// matches (op2 z, (op1, x, y)) if op1IsRight = 1.
+class ThreeOpFrag<SDPatternOperator op1, SDPatternOperator op2,
+                  bit op1IsRight = 0> : ThreeOpFragSDAG<op1, op2, op1IsRight> {
   // The divergence predicate is irrelevant in GlobalISel, as we have
   // proper register bank checks. We just need to verify the constant
   // bus restriction when all the sources are considered.
@@ -938,12 +942,19 @@ def : GCNPat<
  (DivergentBinFrag<mul> i32:$src0, IsPow2Plus1:$src1),
  (V_LSHL_ADD_U32_e64 i32:$src0, (i32 (Log2_32 imm:$src1)), i32:$src0)>;
 
-let SubtargetPredicate = HasLshlAddU64Inst in
+let SubtargetPredicate = HasLshlAddU64Inst in {
 def : GCNPat<
   (ThreeOpFrag<shl_0_to_4, add> i64:$src0, i32:$src1, i64:$src2),
   (V_LSHL_ADD_U64_e64 VSrc_b64:$src0, VSrc_b32:$src1, VSrc_b64:$src2)
 >;
 
+def : GCNPat <
+  // (ptradd z, (shl x, y)) -> ((x << y) + z)
+  (ThreeOpFrag<shl_0_to_4, ptradd, /*op1IsRight=*/1> i64:$src0, i32:$src1, i64:$src2),
+  (V_LSHL_ADD_U64_e64 VSrc_b64:$src0, VSrc_b32:$src1, VSrc_b64:$src2)
+>;
+} // End SubtargetPredicate = HasLshlAddU64Inst
+
 let SubtargetPredicate = HasAddMinMaxInsts in {
 def : ThreeOp_i32_Pats<add, smax, V_ADD_MAX_I32_e64>;
 def : ThreeOp_i32_Pats<add, umax, V_ADD_MAX_U32_e64>;
@@ -1019,19 +1030,24 @@ multiclass IMAD32_Pats <VOP3_Pseudo inst> {
 
 // Handle cases where amdgpu-codegenprepare-mul24 made a mul24 instead of a normal mul.
 // We need to separate this because otherwise OtherPredicates would be overriden.
-class IMAD32_Mul24_Pat<VOP3_Pseudo inst>: GCNPat <
-    (i64 (add (i64 (AMDGPUmul_u24 i32:$src0, i32:$src1)), i64:$src2)),
-    (inst $src0, $src1, $src2, 0 /* clamp */)
-    >;
+class IMAD32_Mul24_Pats_Impl<VOP3_Pseudo inst, SDPatternOperator AddOp, bit mulIsRight = 0> : GCNPat <
+    !if(mulIsRight, (i64 (AddOp i64:$src2, (i64 (AMDGPUmul_u24 i32:$src0, i32:$src1)))),
+                    (i64 (AddOp (i64 (AMDGPUmul_u24 i32:$src0, i32:$src1)), i64:$src2))),
+    (inst $src0, $src1, $src2, 0 /* clamp */)>;
+
+multiclass IMAD32_Mul24_Pats<VOP3_Pseudo inst> {
+  def : IMAD32_Mul24_Pats_Impl<inst, add>;
+  def : IMAD32_Mul24_Pats_Impl<inst, ptradd, /*mulIsRight=*/1>;
+}
 
 // exclude pre-GFX9 where it was slow
 let OtherPredicates = [HasNotMADIntraFwdBug], SubtargetPredicate = isGFX9Plus in {
   defm : IMAD32_Pats<V_MAD_U64_U32_e64>;
-  def : IMAD32_Mul24_Pat<V_MAD_U64_U32_e64>;
+  defm : IMAD32_Mul24_Pats<V_MAD_U64_U32_e64>;
 }
 let OtherPredicates = [HasMADIntraFwdBug], SubtargetPredicate = isGFX11Only in {
   defm : IMAD32_Pats<V_MAD_U64_U32_gfx11_e64>;
-  def : IMAD32_Mul24_Pat<V_MAD_U64_U32_gfx11_e64>;
+  defm : IMAD32_Mul24_Pats<V_MAD_U64_U32_gfx11_e64>;
 }
 
 def VOP3_PERMLANE_Profile : VOP3_Profile<VOPProfile <[i32, i32, i32, i32]>, VOP3_OPSEL> {

llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll

@@ -266,42 +266,25 @@ define amdgpu_kernel void @fold_mad64(ptr addrspace(1) %p) {
 
 ; Use non-zero shift amounts in v_lshl_add_u64.
 define ptr @select_v_lshl_add_u64(ptr %base, i64 %voffset) {
-; GFX942_PTRADD-LABEL: select_v_lshl_add_u64:
-; GFX942_PTRADD:       ; %bb.0:
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshlrev_b64 v[2:3], 3, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: select_v_lshl_add_u64:
-; GFX942_LEGACY:       ; %bb.0:
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 3, v[0:1]
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: select_v_lshl_add_u64:
+; GFX942:       ; %bb.0:
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 3, v[0:1]
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr inbounds i64, ptr %base, i64 %voffset
   ret ptr %gep
 }
 
 ; Fold mul and add into v_mad, even if amdgpu-codegenprepare-mul24 turned the
 ; mul into a mul24.
 define ptr @fold_mul24_into_mad(ptr %base, i64 %a, i64 %b) {
-; GFX942_PTRADD-LABEL: fold_mul24_into_mad:
-; GFX942_PTRADD:       ; %bb.0:
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_and_b32_e32 v2, 0xfffff, v2
-; GFX942_PTRADD-NEXT:    v_and_b32_e32 v4, 0xfffff, v4
-; GFX942_PTRADD-NEXT:    v_mul_hi_u32_u24_e32 v3, v2, v4
-; GFX942_PTRADD-NEXT:    v_mul_u32_u24_e32 v2, v2, v4
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: fold_mul24_into_mad:
-; GFX942_LEGACY:       ; %bb.0:
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_and_b32_e32 v2, 0xfffff, v2
-; GFX942_LEGACY-NEXT:    v_and_b32_e32 v3, 0xfffff, v4
-; GFX942_LEGACY-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, v3, v[0:1]
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: fold_mul24_into_mad:
+; GFX942:       ; %bb.0:
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_and_b32_e32 v2, 0xfffff, v2
+; GFX942-NEXT:    v_and_b32_e32 v3, 0xfffff, v4
+; GFX942-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, v3, v[0:1]
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
   %a_masked = and i64 %a, u0xfffff
   %b_masked = and i64 %b, u0xfffff
   %mul = mul i64 %a_masked, %b_masked

llvm/test/CodeGen/AMDGPU/ptradd-sdag.ll

@@ -25,20 +25,12 @@ define ptr @gep_as0(ptr %p, i64 %offset) {
 ; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX942_PTRADD-LABEL: gep_as0:
-; GFX942_PTRADD:       ; %bb.0: ; %entry
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: gep_as0:
-; GFX942_LEGACY:       ; %bb.0: ; %entry
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 2, v[0:1]
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: gep_as0:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 2, v[0:1]
+; GFX942-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: gep_as0:
 ; GFX10:       ; %bb.0: ; %entry
@@ -188,20 +180,12 @@ define ptr @multi_gep_as0(ptr %p, i64 %offset) {
 ; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX942_PTRADD-LABEL: multi_gep_as0:
-; GFX942_PTRADD:       ; %bb.0: ; %entry
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: multi_gep_as0:
-; GFX942_LEGACY:       ; %bb.0: ; %entry
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 2, v[0:1]
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: multi_gep_as0:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 2, v[0:1]
+; GFX942-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 5
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: multi_gep_as0:
 ; GFX10:       ; %bb.0: ; %entry
@@ -537,3 +521,5 @@ entry:
 ; GFX12_PTRADD: {{.*}}
 ; GFX8_LEGACY: {{.*}}
 ; GFX8_PTRADD: {{.*}}
+; GFX942_LEGACY: {{.*}}
+; GFX942_PTRADD: {{.*}}