diff --git a/.travis.yml b/.travis.yml index b0ddbe226cd..fb192ae82a2 100644 --- a/.travis.yml +++ b/.travis.yml @@ -125,7 +125,7 @@ matrix: - BUILD=make - MAKEFLAGS="-j4" - MAKE_CHECK_COMMAND="true" - - LLVM_VERSION=5.0 + - LLVM_VERSION=6.0 - LLVM_CONFIG="llvm-config-${LLVM_VERSION}" - DRI_LOADERS="--disable-glx --disable-gbm --disable-egl" - DRI_DRIVERS="" @@ -136,12 +136,12 @@ matrix: addons: apt: sources: - - llvm-toolchain-trusty-5.0 + - llvm-toolchain-trusty-6.0 packages: # LLVM packaging is broken and misses these dependencies - libedit-dev # From sources above - - llvm-5.0-dev + - llvm-6.0-dev # Common - xz-utils - x11proto-xf86vidmode-dev @@ -269,7 +269,7 @@ matrix: - DRI_LOADERS="--disable-glx --disable-gbm --disable-egl" - DRI_DRIVERS="" - GALLIUM_ST="--disable-dri --enable-opencl --enable-opencl-icd --enable-llvm --disable-xa --disable-nine --disable-xvmc --disable-vdpau --disable-va --disable-omx-bellagio --disable-gallium-osmesa" - - GALLIUM_DRIVERS="r600,radeonsi" + - GALLIUM_DRIVERS="r600" - VULKAN_DRIVERS="" - LIBUNWIND_FLAGS="--enable-libunwind" addons: @@ -366,7 +366,7 @@ matrix: - BUILD=make - MAKEFLAGS="-j4" - MAKE_CHECK_COMMAND="make -C src/gtest check && make -C src/intel check" - - LLVM_VERSION=5.0 + - LLVM_VERSION=6.0 - LLVM_CONFIG="llvm-config-${LLVM_VERSION}" - DRI_LOADERS="--disable-glx --disable-gbm --disable-egl --with-platforms=x11,wayland" - DRI_DRIVERS="" @@ -377,12 +377,12 @@ matrix: addons: apt: sources: - - llvm-toolchain-trusty-5.0 + - llvm-toolchain-trusty-6.0 packages: # LLVM packaging is broken and misses these dependencies - libedit-dev # From sources above - - llvm-5.0-dev + - llvm-6.0-dev # Common - xz-utils - x11proto-xf86vidmode-dev diff --git a/configure.ac b/configure.ac index e7e057c8f56..7d898aeda9e 100644 --- a/configure.ac +++ b/configure.ac @@ -108,8 +108,8 @@ dnl LLVM versions LLVM_REQUIRED_GALLIUM=3.3.0 LLVM_REQUIRED_OPENCL=3.9.0 LLVM_REQUIRED_R600=3.9.0 -LLVM_REQUIRED_RADEONSI=5.0.0 -LLVM_REQUIRED_RADV=5.0.0 +LLVM_REQUIRED_RADEONSI=6.0.0 +LLVM_REQUIRED_RADV=6.0.0 LLVM_REQUIRED_SWR=5.0.0 dnl Check for progs diff --git a/meson.build b/meson.build index 9a15c8ae3f2..cbaee5ceb22 100644 --- a/meson.build +++ b/meson.build @@ -1151,7 +1151,9 @@ if with_gallium_opencl llvm_optional_modules += ['coroutines', 'opencl'] endif -if with_amd_vk or with_gallium_radeonsi or with_gallium_swr +if with_amd_vk or with_gallium_radeonsi + _llvm_version = '>= 6.0.0' +elif with_gallium_swr _llvm_version = '>= 5.0.0' elif with_gallium_opencl or with_gallium_r600 _llvm_version = '>= 3.9.0' diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c index 54b7e987015..87e36df6431 100644 --- a/src/amd/common/ac_llvm_build.c +++ b/src/amd/common/ac_llvm_build.c @@ -1854,83 +1854,24 @@ LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx, args, 2, AC_FUNC_ATTR_READNONE); } -/* Upper 16 bits must be zero. */ -static LLVMValueRef ac_llvm_pack_two_int16(struct ac_llvm_context *ctx, - LLVMValueRef val[2]) -{ - return LLVMBuildOr(ctx->builder, val[0], - LLVMBuildShl(ctx->builder, val[1], - LLVMConstInt(ctx->i32, 16, 0), - ""), ""); -} - -/* Upper 16 bits are ignored and will be dropped. */ -static LLVMValueRef ac_llvm_pack_two_int32_as_int16(struct ac_llvm_context *ctx, - LLVMValueRef val[2]) -{ - LLVMValueRef v[2] = { - LLVMBuildAnd(ctx->builder, val[0], - LLVMConstInt(ctx->i32, 0xffff, 0), ""), - val[1], - }; - return ac_llvm_pack_two_int16(ctx, v); -} - LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx, LLVMValueRef args[2]) { - if (HAVE_LLVM >= 0x0600) { - LLVMValueRef res = - ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16", - ctx->v2i16, args, 2, - AC_FUNC_ATTR_READNONE); - return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); - } - - LLVMValueRef val[2]; - - for (int chan = 0; chan < 2; chan++) { - /* Clamp between [-1, 1]. */ - val[chan] = ac_build_fmin(ctx, args[chan], ctx->f32_1); - val[chan] = ac_build_fmax(ctx, val[chan], LLVMConstReal(ctx->f32, -1)); - /* Convert to a signed integer in [-32767, 32767]. */ - val[chan] = LLVMBuildFMul(ctx->builder, val[chan], - LLVMConstReal(ctx->f32, 32767), ""); - /* If positive, add 0.5, else add -0.5. */ - val[chan] = LLVMBuildFAdd(ctx->builder, val[chan], - LLVMBuildSelect(ctx->builder, - LLVMBuildFCmp(ctx->builder, LLVMRealOGE, - val[chan], ctx->f32_0, ""), - LLVMConstReal(ctx->f32, 0.5), - LLVMConstReal(ctx->f32, -0.5), ""), ""); - val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->i32, ""); - } - return ac_llvm_pack_two_int32_as_int16(ctx, val); + LLVMValueRef res = + ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16", + ctx->v2i16, args, 2, + AC_FUNC_ATTR_READNONE); + return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); } LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx, LLVMValueRef args[2]) { - if (HAVE_LLVM >= 0x0600) { - LLVMValueRef res = - ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16", - ctx->v2i16, args, 2, - AC_FUNC_ATTR_READNONE); - return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); - } - - LLVMValueRef val[2]; - - for (int chan = 0; chan < 2; chan++) { - val[chan] = ac_build_clamp(ctx, args[chan]); - val[chan] = LLVMBuildFMul(ctx->builder, val[chan], - LLVMConstReal(ctx->f32, 65535), ""); - val[chan] = LLVMBuildFAdd(ctx->builder, val[chan], - LLVMConstReal(ctx->f32, 0.5), ""); - val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan], - ctx->i32, ""); - } - return ac_llvm_pack_two_int32_as_int16(ctx, val); + LLVMValueRef res = + ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16", + ctx->v2i16, args, 2, + AC_FUNC_ATTR_READNONE); + return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); } /* The 8-bit and 10-bit clamping is for HW workarounds. */ @@ -1947,10 +1888,9 @@ LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx, bits != 10 ? max_rgb : ctx->i32_1; LLVMValueRef min_alpha = bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0); - bool has_intrinsic = HAVE_LLVM >= 0x0600; /* Clamp. */ - if (!has_intrinsic || bits != 16) { + if (bits != 16) { for (int i = 0; i < 2; i++) { bool alpha = hi && i == 1; args[i] = ac_build_imin(ctx, args[i], @@ -1960,15 +1900,11 @@ LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx, } } - if (has_intrinsic) { - LLVMValueRef res = - ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16", - ctx->v2i16, args, 2, - AC_FUNC_ATTR_READNONE); - return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); - } - - return ac_llvm_pack_two_int32_as_int16(ctx, args); + LLVMValueRef res = + ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16", + ctx->v2i16, args, 2, + AC_FUNC_ATTR_READNONE); + return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); } /* The 8-bit and 10-bit clamping is for HW workarounds. */ @@ -1981,10 +1917,9 @@ LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx, bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0); LLVMValueRef max_alpha = bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0); - bool has_intrinsic = HAVE_LLVM >= 0x0600; /* Clamp. */ - if (!has_intrinsic || bits != 16) { + if (bits != 16) { for (int i = 0; i < 2; i++) { bool alpha = hi && i == 1; args[i] = ac_build_umin(ctx, args[i], @@ -1992,37 +1927,23 @@ LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx, } } - if (has_intrinsic) { - LLVMValueRef res = - ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16", - ctx->v2i16, args, 2, - AC_FUNC_ATTR_READNONE); - return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); - } - - return ac_llvm_pack_two_int16(ctx, args); + LLVMValueRef res = + ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16", + ctx->v2i16, args, 2, + AC_FUNC_ATTR_READNONE); + return LLVMBuildBitCast(ctx->builder, res, ctx->i32, ""); } LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1) { - assert(HAVE_LLVM >= 0x0600); return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1, &i1, 1, AC_FUNC_ATTR_READNONE); } void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1) { - if (HAVE_LLVM >= 0x0600) { - ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt, - &i1, 1, 0); - return; - } - - LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1, - LLVMConstReal(ctx->f32, 1), - LLVMConstReal(ctx->f32, -1), ""); - ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt, - &value, 1, AC_FUNC_ATTR_LEGACY); + ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt, + &i1, 1, 0); } LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input, @@ -3231,7 +3152,7 @@ ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3) { unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3); - if (ctx->chip_class >= VI && HAVE_LLVM >= 0x0600) { + if (ctx->chip_class >= VI) { return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false); } else { return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask); diff --git a/src/amd/common/ac_llvm_helper.cpp b/src/amd/common/ac_llvm_helper.cpp index a4b2fde786a..e022e12c7f3 100644 --- a/src/amd/common/ac_llvm_helper.cpp +++ b/src/amd/common/ac_llvm_helper.cpp @@ -92,11 +92,7 @@ LLVMBuilderRef ac_create_builder(LLVMContextRef ctx, llvm::unwrap(builder)->setFastMathFlags(flags); break; case AC_FLOAT_MODE_UNSAFE_FP_MATH: -#if HAVE_LLVM >= 0x0600 flags.setFast(); -#else - flags.setUnsafeAlgebra(); -#endif llvm::unwrap(builder)->setFastMathFlags(flags); break; } diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 405833d3ba7..3afdca52ea6 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -2450,30 +2450,6 @@ static LLVMValueRef fetch_constant( /* Fast path when user data SGPRs point to constant buffer 0 directly. */ if (sel->info.const_buffers_declared == 1 && sel->info.shader_buffers_declared == 0) { - - /* This enables use of s_load_dword and flat_load_dword for const buffer 0 - * loads, and up to x4 load opcode merging. However, it leads to horrible - * code reducing SIMD wave occupancy from 8 to 2 in many cases. - * - * Using s_buffer_load_dword (x1) seems to be the best option right now. - * - * LLVM 5.0 on SI doesn't insert a required s_nop between SALU setting - * a descriptor and s_buffer_load_dword using it, so we can't expand - * the pointer into a full descriptor like below. We have to use - * s_load_dword instead. The only case when LLVM 5.0 would select - * s_buffer_load_dword (that we have to prevent) is when we use use - * a literal offset where we don't need bounds checking. - */ - if (ctx->screen->info.chip_class == SI && HAVE_LLVM < 0x0600 && - !reg->Register.Indirect) { - LLVMValueRef ptr = - LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers); - - addr = LLVMBuildLShr(ctx->ac.builder, addr, LLVMConstInt(ctx->i32, 2, 0), ""); - LLVMValueRef result = ac_build_load_invariant(&ctx->ac, ptr, addr); - return bitcast(bld_base, type, result); - } - LLVMValueRef desc = load_const_buffer_desc_fast_path(ctx); LLVMValueRef result = buffer_load_const(ctx, desc, addr); return bitcast(bld_base, type, result); diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c index 4ae5b006593..0aefca22385 100644 --- a/src/gallium/drivers/radeonsi/si_shader_nir.c +++ b/src/gallium/drivers/radeonsi/si_shader_nir.c @@ -748,13 +748,6 @@ void si_nir_scan_shader(const struct nir_shader *nir, void si_lower_nir(struct si_shader_selector* sel) { - /* Disable const buffer fast path for old LLVM versions */ - if (sel->screen->info.chip_class == SI && HAVE_LLVM < 0x0600 && - sel->info.const_buffers_declared == 1 && - sel->info.shader_buffers_declared == 0) { - sel->info.const_buffers_declared |= 0x2; - } - /* Adjust the driver location of inputs and outputs. The state tracker * interprets them as slots, while the ac/nir backend interprets them * as individual components. diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c index c51d057967c..4c908745313 100644 --- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c +++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c @@ -56,11 +56,9 @@ void si_llvm_emit_kill(struct ac_shader_abi *abi, LLVMValueRef visible) LLVMBuilderRef builder = ctx->ac.builder; if (ctx->shader->selector->force_correct_derivs_after_kill) { - /* LLVM 6.0 can kill immediately while maintaining WQM. */ - if (HAVE_LLVM >= 0x0600) { - ac_build_kill_if_false(&ctx->ac, - ac_build_wqm_vote(&ctx->ac, visible)); - } + /* Kill immediately while maintaining WQM. */ + ac_build_kill_if_false(&ctx->ac, + ac_build_wqm_vote(&ctx->ac, visible)); LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, ""); mask = LLVMBuildAnd(builder, mask, visible, "");