[AMDGPU] (x or y) xor -1 -> x nor y (#130264)

Added pattern so s_nor is selected for ((i1 x or i1 y) xor -1) instead
of s_or and s_xor . This patch is for i1 divergent. The ballot in the
test is added for the retrieval of lanemask. The control flow is needed
because the combiner can't pass through phi instructions.

Author: mihajlovicana

llvm/lib/Target/AMDGPU/SOPInstructions.td

@@ -1925,6 +1925,20 @@ def : ScalarNot2Pat<S_ORN2_B32, or, v2i16>;
 def : ScalarNot2Pat<S_ORN2_B64, or, v4i16>;
 def : ScalarNot2Pat<S_ORN2_B64, or, v2i32>;
 
+let WaveSizePredicate = isWave32 in {
+def : GCNPat<
+  (i1 (not (or_oneuse i1:$src0, i1:$src1))),
+  (S_NOR_B32 i1:$src0, i1:$src1)
+>;
+}
+
+let WaveSizePredicate = isWave64 in {
+def : GCNPat<
+  (i1 (not (or_oneuse i1:$src0, i1:$src1))),
+  (S_NOR_B64 i1:$src0, i1:$src1)
+>;
+}
+
 //===----------------------------------------------------------------------===//
 // Target-specific instruction encodings.
 //===----------------------------------------------------------------------===//

llvm/test/CodeGen/AMDGPU/nor-divergent-lanemask.ll

@@ -0,0 +1,107 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -mattr="-wavefrontsize32,+wavefrontsize64" -o - < %s | FileCheck -check-prefixes=SDAG-W64 %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -mattr="-wavefrontsize32,+wavefrontsize64" -o - < %s | FileCheck -check-prefixes=GISEL-W64 %s
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -mattr="+wavefrontsize32,-wavefrontsize64" -o - < %s | FileCheck -check-prefixes=SDAG-W32 %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -mattr="+wavefrontsize32,-wavefrontsize64" -o - < %s | FileCheck -check-prefixes=GISEL-W32 %s
+
+; Use ballot for easy access to lanemask
+
+define amdgpu_ps i64 @test_nor(i64 inreg %a, i64 inreg %b) {
+; SDAG-W64-LABEL: test_nor:
+; SDAG-W64:       ; %bb.0:
+; SDAG-W64-NEXT:    s_nor_b64 s[0:1], s[0:1], s[2:3]
+; SDAG-W64-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; SDAG-W64-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; SDAG-W64-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v0
+; SDAG-W64-NEXT:    ; return to shader part epilog
+;
+; GISEL-W64-LABEL: test_nor:
+; GISEL-W64:       ; %bb.0:
+; GISEL-W64-NEXT:    s_nor_b64 s[0:1], s[0:1], s[2:3]
+; GISEL-W64-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GISEL-W64-NEXT:    s_and_b64 s[0:1], s[0:1], exec
+; GISEL-W64-NEXT:    ; return to shader part epilog
+;
+; SDAG-W32-LABEL: test_nor:
+; SDAG-W32:       ; %bb.0:
+; SDAG-W32-NEXT:    s_nor_b32 s0, s0, s2
+; SDAG-W32-NEXT:    s_mov_b32 s1, 0
+; SDAG-W32-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; SDAG-W32-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; SDAG-W32-NEXT:    v_cmp_ne_u32_e64 s0, 0, v0
+; SDAG-W32-NEXT:    ; return to shader part epilog
+;
+; GISEL-W32-LABEL: test_nor:
+; GISEL-W32:       ; %bb.0:
+; GISEL-W32-NEXT:    s_nor_b32 s0, s0, s2
+; GISEL-W32-NEXT:    s_mov_b32 s1, 0
+; GISEL-W32-NEXT:    s_and_b32 s0, s0, exec_lo
+; GISEL-W32-NEXT:    ; return to shader part epilog
+  %a.lanemask = call i1 @llvm.amdgcn.inverse.ballot.i64(i64 %a)
+  %b.lanemask = call i1 @llvm.amdgcn.inverse.ballot.i64(i64 %b)
+  %or = or i1 %a.lanemask, %b.lanemask
+  %xor = xor i1 %or, true
+  %r = call i64 @llvm.amdgcn.ballot.i64(i1 %xor)
+  ret i64 %r
+}
+
+define amdgpu_ps i64 @test_or_two_uses(i64 inreg %a, i64 inreg %b) {
+; SDAG-W64-LABEL: test_or_two_uses:
+; SDAG-W64:       ; %bb.0:
+; SDAG-W64-NEXT:    s_or_b64 s[0:1], s[0:1], s[2:3]
+; SDAG-W64-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; SDAG-W64-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; SDAG-W64-NEXT:    s_xor_b64 s[0:1], s[0:1], -1
+; SDAG-W64-NEXT:    s_waitcnt_depctr 0xfffe
+; SDAG-W64-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s[0:1]
+; SDAG-W64-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SDAG-W64-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; SDAG-W64-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v1
+; SDAG-W64-NEXT:    s_and_b64 s[0:1], s[0:1], vcc
+; SDAG-W64-NEXT:    s_waitcnt_depctr 0xfffe
+; SDAG-W64-NEXT:    ; return to shader part epilog
+;
+; GISEL-W64-LABEL: test_or_two_uses:
+; GISEL-W64:       ; %bb.0:
+; GISEL-W64-NEXT:    s_or_b64 s[0:1], s[0:1], s[2:3]
+; GISEL-W64-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GISEL-W64-NEXT:    s_xor_b64 s[2:3], s[0:1], -1
+; GISEL-W64-NEXT:    s_and_b64 s[0:1], s[0:1], exec
+; GISEL-W64-NEXT:    s_and_b64 s[2:3], s[2:3], exec
+; GISEL-W64-NEXT:    s_and_b64 s[0:1], s[2:3], s[0:1]
+; GISEL-W64-NEXT:    ; return to shader part epilog
+;
+; SDAG-W32-LABEL: test_or_two_uses:
+; SDAG-W32:       ; %bb.0:
+; SDAG-W32-NEXT:    s_or_b32 s0, s0, s2
+; SDAG-W32-NEXT:    s_mov_b32 s3, 0
+; SDAG-W32-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; SDAG-W32-NEXT:    s_xor_b32 s0, s0, -1
+; SDAG-W32-NEXT:    s_mov_b32 s1, s3
+; SDAG-W32-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; SDAG-W32-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; SDAG-W32-NEXT:    v_cmp_ne_u32_e64 s0, 0, v0
+; SDAG-W32-NEXT:    v_cmp_ne_u32_e64 s2, 0, v1
+; SDAG-W32-NEXT:    s_and_b64 s[0:1], s[2:3], s[0:1]
+; SDAG-W32-NEXT:    ; return to shader part epilog
+;
+; GISEL-W32-LABEL: test_or_two_uses:
+; GISEL-W32:       ; %bb.0:
+; GISEL-W32-NEXT:    s_or_b32 s0, s0, s2
+; GISEL-W32-NEXT:    s_mov_b32 s1, 0
+; GISEL-W32-NEXT:    s_xor_b32 s4, s0, -1
+; GISEL-W32-NEXT:    s_and_b32 s2, s0, exec_lo
+; GISEL-W32-NEXT:    s_mov_b32 s3, s1
+; GISEL-W32-NEXT:    s_and_b32 s0, s4, exec_lo
+; GISEL-W32-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GISEL-W32-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
+; GISEL-W32-NEXT:    ; return to shader part epilog
+  %a.lanemask = call i1 @llvm.amdgcn.inverse.ballot.i64(i64 %a)
+  %b.lanemask = call i1 @llvm.amdgcn.inverse.ballot.i64(i64 %b)
+  %or = or i1 %a.lanemask, %b.lanemask
+  %xor = xor i1 %or, true
+  %r0 = call i64 @llvm.amdgcn.ballot.i64(i1 %xor)
+  %r1 = call i64 @llvm.amdgcn.ballot.i64(i1 %or)
+  %r = and i64 %r0, %r1
+  ret i64 %r
+}