From 2b3ebe307c2250f3cc8ffea203e3a2196a70d496 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Marek=20Ol=C5=A1=C3=A1k?= Date: Thu, 23 Feb 2017 02:15:54 +0100 Subject: [PATCH] ac: replace SI.export with amdgcn.exp.* Reviewed-by: Dave Airlie --- src/amd/common/ac_llvm_build.c | 31 ++++++++++++++++++++++++ src/gallium/drivers/radeonsi/si_shader.c | 8 +++--- 2 files changed, 36 insertions(+), 3 deletions(-) diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c index fae551055ed..bd1b63d40fa 100644 --- a/src/amd/common/ac_llvm_build.c +++ b/src/amd/common/ac_llvm_build.c @@ -891,6 +891,37 @@ void ac_emit_export(struct ac_llvm_context *ctx, struct ac_export_args *a) { LLVMValueRef args[9]; + if (HAVE_LLVM >= 0x0500) { + args[0] = LLVMConstInt(ctx->i32, a->target, 0); + args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0); + + if (a->compr) { + LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context); + LLVMTypeRef v2i16 = LLVMVectorType(i16, 2); + + args[2] = LLVMBuildBitCast(ctx->builder, a->out[0], + v2i16, ""); + args[3] = LLVMBuildBitCast(ctx->builder, a->out[1], + v2i16, ""); + args[4] = LLVMConstInt(ctx->i1, a->done, 0); + args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0); + + ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16", + ctx->voidt, args, 6, 0); + } else { + args[2] = a->out[0]; + args[3] = a->out[1]; + args[4] = a->out[2]; + args[5] = a->out[3]; + args[6] = LLVMConstInt(ctx->i1, a->done, 0); + args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0); + + ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.exp.f32", + ctx->voidt, args, 8, 0); + } + return; + } + args[0] = LLVMConstInt(ctx->i32, a->enabled_channels, 0); args[1] = LLVMConstInt(ctx->i32, a->valid_mask, 0); args[2] = LLVMConstInt(ctx->i32, a->done, 0); diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 055c8f7daed..7bc977b6ed7 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -6490,7 +6490,7 @@ static bool si_eliminate_const_output(struct si_shader_context *ctx, for (i = 0; i < 4; i++) { LLVMBool loses_info; - LLVMValueRef p = LLVMGetOperand(inst, 5 + i); + LLVMValueRef p = LLVMGetOperand(inst, (HAVE_LLVM >= 0x0500 ? 2 : 5) + i); /* It's a constant expression. Undef outputs are eliminated too. */ if (LLVMIsUndef(p)) { @@ -6574,10 +6574,12 @@ static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx) unsigned num_args = LLVMCountParams(callee); /* Check if this is an export instruction. */ - if (num_args != 9 || strcmp(name, "llvm.SI.export")) + if ((num_args != 9 && num_args != 8) || + (strcmp(name, "llvm.SI.export") && + strcmp(name, "llvm.amdgcn.exp."))) continue; - LLVMValueRef arg = LLVMGetOperand(cur, 3); + LLVMValueRef arg = LLVMGetOperand(cur, HAVE_LLVM >= 0x0500 ? 0 : 3); unsigned target = LLVMConstIntGetZExtValue(arg); if (target < V_008DFC_SQ_EXP_PARAM) -- 2.30.2