ac/llvm,radeonsi: lower nir_fpow for aco and llvm

aco does not implement fpow, need nir to lower it
first. llvm will do by itself in the same way, so
we always lower fpow in nir now.

Remove the llvm fpow implementation that has special
handling for the muliplication. It's not used any
more and does not match GLSL spec as fpow(0,0)=NaN
but here we get 0.

There's some pixel changes for gl-radeonsi-stoney:
  ror-default 2 (no tolerance), 0 (1% tol.)

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22573>

src/amd/ci/traces-amd.yml

@@ -104,7 +104,7 @@ traces:
       checksum: 60f74020451e9beaf586b4551541b763
   ror/ror-default.trace:
     gl-radeonsi-stoney:
-      checksum: ea53f93df31703bf7a07a4efb922608f
+      checksum: a37b58424c4289a6de77e61d599b6fab
   nheko/nheko-colors.trace:
     gl-radeonsi-stoney:
       checksum: d3234cd6ccc2ab5d3ceab7db79300c69

src/amd/llvm/ac_nir_to_llvm.c

@@ -844,30 +844,6 @@ static bool visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
       src[0] = ac_to_float(&ctx->ac, src[0]);
       result = ac_build_frexp_mant(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
       break;
-   case nir_op_fpow:
-      if (instr->dest.dest.ssa.bit_size != 32) {
-         /* 16 and 64 bits */
-         result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
-                                       ac_to_float_type(&ctx->ac, def_type), src[0]);
-         result = LLVMBuildFMul(ctx->ac.builder, result, ac_to_float(&ctx->ac, src[1]), "");
-         result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
-                                       ac_to_float_type(&ctx->ac, def_type), result);
-         break;
-      }
-      if (LLVM_VERSION_MAJOR >= 12) {
-         result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
-                                       ac_to_float_type(&ctx->ac, def_type), src[0]);
-         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fmul.legacy", ctx->ac.f32,
-                                     (LLVMValueRef[]){result, ac_to_float(&ctx->ac, src[1])},
-                                     2, 0);
-         result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
-                                       ac_to_float_type(&ctx->ac, def_type), result);
-         break;
-      }
-      /* Older LLVM doesn't have fmul.legacy. */
-      result = emit_intrin_2f_param(&ctx->ac, "llvm.pow", ac_to_float_type(&ctx->ac, def_type),
-                                    src[0], src[1]);
-      break;
    case nir_op_fmax:
       result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum", ac_to_float_type(&ctx->ac, def_type),
                                     src[0], src[1]);

src/gallium/drivers/radeonsi/si_get.c

@@ -1280,6 +1280,7 @@ void si_init_screen_get_functions(struct si_screen *sscreen)
       .fuse_ffma32 = use_fma32,
       .fuse_ffma64 = true,
       .lower_fmod = true,
+      .lower_fpow = true,
       .lower_ineg = true,
       .lower_pack_snorm_4x8 = true,
       .lower_pack_unorm_4x8 = true,